Subscribe now

Teams seem to be spending more time talking to Claude, ChatGPT, or some other AI-powered agent, and less time thinking deeply or talking to one another. But we’re all getting more done, right?

Meanwhile, your finance team is scrambling to keep up. So your chart of accounts looks more like an over-decorated tree than a system for reporting performance.

Take a breath, I’ve got you. The following should help you regain control over your ledger. First, let’s define “AI spend.”

Separate pure-play AI from AI-flavored SaaS

Let’s talk about inference, a word that tends to get thrown around liberally. But it has a real definition.

Inference is the doing part. It is when a trained model takes what it has learned from training to turn data into findings, predictions, classifications, or decisions in real time.

Training comes first. You can think of it as showing your kid how to tie their shoes, or giving an agent a flux analysis skill. Inference is the application of that skill. Your kid sees the laces are undone, you remind them, and they remember how to tie their laces. Getting them to tie their laces on their own is another story.

Once trained, chatbots are probably the cleanest example of inference in action. Like when you ask your month-end close agent to run the /me-flux-analysis skill and it jumps to action, writing out a summary of the accounts that exceeded your thresholds with links to the associated transactions.

Alternatively, when you log into your accounting automation tool and click the “start close” button and it generates an in-app flux analysis report, that report could be generated based on a series of rules or it could be powered by a trained agent. Depending on how the flux gets generated, there could be some inference involved. But is it worth pulling out and tracking separately? Nope, that’s just a software tool powered by AI on the back-end.

Every account needs a job

Generally speaking, when it comes to managing a chart of accounts, every account should clear one of four hurdles:

• Volume: Ten or more transactions in a period

• Size: More than 2% of its statement section

• Decision: It informs a strategic or budgetary decision

• Key Metric: It feeds a number or input to a KPI or other tracked figure

Inference spend clears three of those with room to spare. And even when it might still be small today, it is a topic of conversation deserving of a presence in your chart of accounts.

But how many accounts should you be using for your AI spend? At least four, and up to six depending on the size of your organization.

• Embedded in your product — COGS

• Improving operations — Opex

• Prepaid commitments — Asset

• Usage billed in arrears — Liability

Two more worth considering. If you include software spend in your opex family for GTM, give sales and marketing a dedicated account there. And larger teams will want to consider one for their engineering team, but only if you already have a family of opex accounts for R&D.

Let your departments do the splitting

Speaking of departments, I’d like to use this as an opportunity to remind you to let your departments split the bill, not your accounts.

Some companies will likely have a dedicated family of opex for the sales and marketing team, others will go so far as to have one for the product and engineering team. That’s great! But every team shouldn’t have their own set of accounts. You eventually need to draw the line.

When you do, please just bear in mind that headcount allocation works fine for seat-based software, but then quickly falls apart for consumption pricing. In some cases, engineering can burn ten times what finance does. Allocate all of your costs by heads and you get a materially wrong P&L and a gross margin that is effectively useless.

Pro tip: Track your balance sheet by department.

This is one I learned trying to cobble together an answer for a CMO. He asked me how much he had sitting in prepaid commitments that had not yet hit his budget, because the event had not happened yet. Great question, I said. As I quickly ran off to pull some magic out of a hat.

If only I’d been tracking our prepaids and accrued expenses by department. Something I have never had to wonder about since. Now and forever, my balance sheet will use departments.

Watch the ratio, not just the total

Now that you have the what, time for the why.

At some point your board is going to ask what you are spending on the frontier models and you want that number before they ask. This is the fastest-growing cost line at most companies right now, and anything growing that fast that you cannot see clearly is exactly where things go off the rails.

But a total only tells you the size. It does not tell you whether the money is working. That’s where being able to compare it relative to traditional software and people costs becomes a strategic unlock.

That relationship is what tells you whether you are seeing real impact in productivity and revenue growth. Or whether you are just paying to play in a company sandbox, handing the kids more sand to push around.

Make the tens mean something

What does that even mean?

This is something I picked up—might have invented, not sure—but one thing I have learned remixing too many CoAs over the years. Step back and take the time to truly think about and look at your numbers. Creating a numbering convention that means something can quickly become a strategic unlock for both your finance and accounting teams (potentially others).

Pick a tens slot, give it a meaning, and hold that meaning everywhere it shows up. I’m referring to the tens digit. This is the “10” in a 4010 revenue account. And make sure its deferred revenue twin matches at 2410.

Now do the same with your product costs, whether they’re sitting in COGS or Opex. Your in-product inference and your engineering inference should share a ten. And again with people costs sitting above or below the line.

Final Thoughts

You do not need more accounts to see your AI spend. You need four to six, carefully designed. Each with a purpose.

Do that and your chart of accounts stops looking like a tree somebody kept hanging ornaments on, and starts looking like a system you can actually report from.

If you got this far and want more, I wrote a much longer (some would say long-winded) piece on the art of the chart of accounts.

Subscribe now

For my first article, I wrote my playbook for leveraging AI in financial modeling.

How to build financial models with Claude

This past summer I interviewed for a role that was right in my wheelhouse and quickly found myself at the take-home stage. My homework assignment: build a revenue forecast with an accompanying outline for how I would pitch the business to a potential acquirer.

Easy enough. I have done this hundreds of times. And of course I was going to leverage AI to build a robust model.

The problem? I did steps 3 and 4 in reverse order because I was in a hurry, and then skipped step 5 entirely. This created a really bad model and I basically had to start over.

Below is the process that I should have followed (and that I do now)…

How I Built the Model:

1. “Model Builder” Project

I created a Project within Claude to hold the model context, files, and project instructions. You can think of a project as a self-contained workspace for related chat histories and knowledge bases. In this case, I packed 20 years of financial modeling best practices into a tidy, new Claude project.

Not sure where to begin with a project. That’s ok. Tell Claude what you are looking to do and it can help you workshop the instructions for your new project and even suggest potential pieces of context.

2. Building the Skills

Skills if you’re still new to the AI game are reusable instructions, scripts, and resources that dictate how to perform specialized, repeatable tasks. You can either build the skills from scratch or you can download pre-built ones. You can even customize pre-built skills to more squarely suit your needs.

From Scratch: Work with Claude/ChatGPT chat to build the skills you need to produce a good model.

Pre-Built Skills: One of the great things about AI being so popular is you don’t need to start from scratch. There are lots of smart people sharing skills they’ve built. For instance, I downloaded some skills that I saw an FP&A vendor promoting on a recent webinar and combined them with some of OnlyCFO’s model Skills from goClose (see Napoleon Dynamite characters below 😂)

3. Start with a sketch

This is my cardinal rule of model building.

Build a schema for the model in a spreadsheet and then you can give that to Claude so it knows what you’re generally looking for. It doesn’t need all the formulas, just the rows and columns to get an idea of what you’re thinking. This will save a ton of back and forth.

If you already have a working model, you can do the reverse. Ask Claude to audit your model and build a flowchart, showing you how it’s connected and test that it is fully wired.

4. Write the prompt

The prompt is where the assignment gets handed over: the target (quote it word for word), the timeframe, and the questions the model needs to answer. If you did step 3, this is mostly transcription…the prompt is easy because it is based on the sketch.

However, this was where I broke my own order. In my rush I skipped the sketch and dove right into hammering out a prompt where I effectively converted (or tried to convert) the assignment into written instructions for what I wanted Claude to build. Only to lose everything after about two hours of hyperactive typing and start from scratch to do it a second time.

5. Humans are for judgment calls, agents need details

I like to equate AI with a new puppy. It wants to please you… mostly. It wants to give you an answer more than anything. Which is great. Except for when it’s not. Our job as humans is to make judgment calls and deliver specificity to the agent.

Tell it exactly what you want. Tell it the knobs you want, how you want them designed, where to start for the assumptions, and how you want it stitched together.

You can even feed it benchmarks, industry norms, seasonality, ratios, and other trends you know about the business and then test its alignment with that data. AI agents can handle a lot of context. The more specific and structured, the better.

You’d be surprised about how quickly AI can organize those scattered thoughts into a series of quantitative checks to test the reasonableness of a model.

And then back-test the model to make sure it is working right. Have it recalculate a month you already lived through and confirm it lands on the number you actually reported.

6. One change per turn, and keep a change log

Compounding edits compound bugs.

When iterating with AI, limit feedback to one change per turn, with a confirmation of exactly which cells changed and one downstream output that responded. It can be tempting to load it up, but when you start trying to change too many things at once, the AI can get lost. Especially if you start making changes to the file and don’t detail every change.

Your colleagues will thank you as well. We all know what it’s like to receive a model from someone, only to try and diff it yourself to figure out what changed.

Regardless of whether you are having AI build, validate, or update a model for you, changelogs are the CliffsNotes of a model.

7. Make the model prove it’s alive

A model that looks right and a model that is right are different things.

My favorite test: nudge one input and confirm the output moves in the expected direction and by the expected amount. And then check it yourself. All models can be validated with some old-fashioned napkin math.

If you change the sales capacity ramp from three months to six, does that bring down your revenue forecast accordingly? If not, something is hardcoded where a formula should be.

This is one great place where Claude has a superpower to help test your model. You can give it a laundry list of scenarios to run, and it will synthesize the results. You don’t have to be building a new model to get value from this tip.

Another test you can run to prove it’s alive is to ask it to build an independent replica of the math in Python and compare it to the spreadsheet. When they disagree, one of the two is wrong, and finding out which teaches you something every time. Just beware that Python and Excel behave differently when it comes to rounding.

Final Thoughts

If you’ve made it this far, I hope you’ve learned something from this piece. I’d hate for you to make the same mistake I made this summer.

AI can be super helpful both building and validating financial models. But…only if you do it the right way. AI can also just make the model say whatever you want.

Hope you learned something and I’ll see you next week!

Footnotes:

• Checkout goClose for a month-end close workspace, free AI Skills, and other templates.

• Subscribe and share with your friends

Subscribe now

One project that started as a fun vibecoding experiment became a full month-end close production app (close checklist, streamlined recons, flux analysis, etc). I worked with some friends who are way smarter than me to move from “vibecoded” to production.

Check it out and let me know what you think…

There are a lot of things I learned as I built stuff with AI and also as I watched my teams “vibecode” stuff to help them in their functions. After about one year of building with AI in the finance function, here are my 1) five reasons not to vibecode internal apps and 2) two reasons why it may still be a good idea.

5 Reasons Not to Vibecode Internal Apps

1. It Always Takes Longer Than You Think

Everyone I have talked to on my teams has experienced the same thing. They get really excited after Claude farts out a pretty artifact that feels like it’s 90% of the way there. “We can cancel our forecasting tool! I just vibecoded something so we will save lots of money.”

The last 10% though takes 100,000% longer than the seemingly easy first 90% for any somewhat complicated app. This will get easier as AI and the tools continue to improve, but…it’s still really hard.

2. Cost Savings Aren’t What You Think

A lot of folks talk about cost savings as the reason they decided to build something in-house. And it sounds nice on a slide deck to the executive team or your board about how AI-pilled the company is becoming (have seen several of those).

But…it’s always more expensive than you think. It takes a lot longer to build (as mentioned above), but also sucks up a lot of time for everyone involved.

Also, any assumed cost savings will continue to shrink. We are seeing an explosion of AI-native, bootstrapped companies today thanks to AI and they are putting a ton of pricing pressure on the incumbents. Just go with a cheap tool from a startup…will probably be less expensive than you building it and it will very likely be better.

3. Vibecoded Apps Die With Vibecoder

Internal vibecoded apps usually have one primary owner. They designed it, QA’d it, they are behind the scenes trying to fix bugs, add features, etc when other people complain about it. It’s their baby.

So what happens when that person leaves the company?

Let me tell you from experience….nobody wants to adopt that person’s baby. Why would they? They don’t get the glory of building something that saves the company a bunch of money, but they now have to be customer support, customer success, and product development for it…

4. No One to Throw Under the Bus

Nobody ever got fired for buying IBM

IBM is obviously an outdated example and the vibecoding comparison is more extreme, but the idea is the same. By raising your hand and saying we can just vibecode this internally, you are taking on A LOT of responsibility.

It’s really nice to have someone external to throw under the bus when things go wrong. Compliance, security, user access, bugs, errors, etc. Otherwise that all falls on you…

5. How’s Your IT / Engineering Team Holding Up?

Here is another hidden cost…whoever is managing your teams’ hosted apps will get exponentially busier. When you have non-technical users vibecoding, eventually they will need IT to help “finish” it and then they need them every time something breaks (which will be a lot)

You are either going to have to hire more technical folks to maintain your internal apps or you are going to break your current folks and they rage quit.

When Vibecoding Internal Apps Makes Sense

Ok, I was pretty negative on vibecoding apps above but there are some good reasons to build internally.

1. Customization Matters

Sometimes you have a unique process or tech stack that isn’t fully supported by any existing vendors. And the lack of customization slows you down or makes things much more inefficient.

I believe this is the best reason for building internally.

If you can move faster and be more efficient because of the ability for endless customization (because you built it) then maybe it’s worth it.

2. The App Is Low Impact

Like I said above, saving money is rarely a good reason. But if the app is low impact and you don’t need a full enterprise app then building may make sense:

• Your company is small

• Doesn’t touch sensitive data

• Limited number of users

• Nothing blows up if the app is gone tomorrow

I think about the below a lot when I hear people on social media talking about the death of SaaS and all the cool stuff they are vibecoding:

Final Thoughts

There are a lot of internal AI slop apps. I have been asked to use many…

They can seriously slow down the business. Potentially saving a few thousand dollars per year means nothing for almost every VC-backed company. If you can’t grow, then you are dead anyway.

If building internally actually helps you move faster or more efficiently, then go for it. The point of this post is not to discourage it, but rather to think carefully about it. There will be lots of pride in these projects from the person leading the build, but sometimes/often the projects will need to be killed before they waste too much time.

Look up “sunk cost fallacy” and make sure your team understands it.

Subscribe now

Want all of the AI Skills that OnlyCFO Builds?

They will all be on goClose so go sign up! We just uploaded the Skills OnlyCFO built to manage the cap table (create the grant list for board approval, reconcile equity grants, monitor activity, and 409A reviewer). They are in the the Resources section

goClose offers these resources and a month-end close tool for free.

Get OnlyCFO's AI Skills

AI can do incredible things that will benefit humanity (cure diseases, solve science’s hardest problems, create abundance, etc). But for businesses, it boils down to one question…And it’s on every CFO’s mind:

What’s the return on our AI spend?

A popular quote about marketing spend (that most marketers have grown to hate) is that half of their spend is wasted, but you don’t know which half. I think a similar story exists for AI spend today…

CFOs don’t want to suffocate innovation with AI but they also need to make sure the return on AI spend is worth the cost. Measuring that return can be hard tho…

There are two ways to increase the return on tokens:

1. Lower the cost of tokens

2. Create more value from the tokens

Lower Your AI Costs!

There is currently a lot of focus on #1. Understandably so…It’s much easier to measure and there is a lot of opportunity for “tokenmining”. Just look at Claude or ChatGPT’s revenue growth rate to see the massive opportunity here. I think an uncomfortable amount of their current revenue would be eliminated if all companies had good model routers and were utilizing open-weight models where it made sense.

*this doesn’t mean Anthropic and OpenAI are cooked (as the kids say) because the TAM is massive, but there is a lot of cost-cutting opportunity right now…

Two of the biggest opportunities for “tokenmining” (ie cutting AI model costs while not impacting quality) are:

1. Shifting some usage to open-weight models

2. Using a model router to use the least expensive model required for the task

Model routing has become a very hot product. Many companies are adding a model routing product (Ramp announced theirs this week). Stripe just announced that it is acquiring one of the leaders in the space (OpenRouter) for ~$8B, which is ~6x their Series B valuation of $1.3B from just 3 months earlier. And it represents a ~54x ARR multiple (~$150M ARR at ~$8B)!

TLDR: Cutting AI token spend is the most measurable thing to do to show improvement in return on tokens. There is plenty of opportunity there so everyone should go do it.

Are We Creating Value?

This is the question that not enough people are asking yet. But good CFOs will be pushing on this and making sure we aren’t just cutting costs, but we are also driving real value.

It’s a lot squishier to measure though…

Each department needs to set goals of what AI spend is enabling them to do and why it matters (how does it align to company objectives). Examples might include:

• Incremental revenue

• Realized labor savings. Not just hours theoretically saved

• Cycle-time or speed improvement

• Error/risk reduction

Are people creating actual value or…are they building AI slop apps just because they can? I have seen too many of these apps forced upon me - they add no value and suck up everyone’s time.

And tracking spend appropriately on the P&L is an important starting point for driving higher return on tokens…

AI Costs on the P&L

Cost of Revenue & Gross Margins

When AI is embedded in your product, that portion of the costs is also included in COGS.

AI costs that may get included in COGS:

• API spend with AI models (Claude, ChatGPT, etc)

• Model hosting when you self-host (not calling an API but using AWS, Azure, etc GPU compute)

• Vector database costs

• Orchestration and observability

While AI gross margins are up a lot in the last couple of years (from 45% in 2025 to 53% expected in 2026 and 59% in 2027), they are still significantly lower than typical software gross margins. But with the growing popularity of open-weight models and model routing, we may see these accelerate even faster for the application layer.

But it’s important to track gross margins of all products (especially AI products) separately because they can have vastly different costs and gross margin profiles. The best companies at managing this can track individual features and watch AI costs trend over time. Then they know where to focus their cost-cutting efforts or where their pricing model is broken.

Create a new line just for the AI costs running through COGS. You, the broader exec team, and the board will want to pay close attention to COGS from AI so tell your accounting team to create a separate line for it and track it by product.

How to Allocate Your AI Model Bill

• Don’t treat G&A as an expense dumping ground with your AI spend. Yes, dumping everything in G&A makes all your other metrics look better but it’s incorrect.

• AI spend for unpaid customer POCs/trials should go to sales & marketing. Make sure your accounting team isn’t putting it in COGS or R&D

• Can I allocate AI model spend based on headcount?

  Yes…but only if you are paying a flat fee per seat (ie not paying based on consumption). Then a simple headcount allocation like all your seat-based tools makes sense.

• How should I allocate usage-based AI spend?

  Text within this block will maintain its original spacing when published

  Based on the usage…duh.
  
  The only good approach now is to do direct allocations based on specific department usage. Finance teams need to segment the different types of AI costs by department. Anything else will result in materially incorrect financials and unit economics.
  
  Once you have true cost allocations then you can better determine the “return on tokens”.

Footnotes:

• Subscribe and forward this newsletter to your team.

• Check out goClose to get OnlyCFO’s downloadable Skills

Subscribe now

What does every CFO love? An accurate and fast accounting month-end close. Numbers provided in the middle of the next month are much less useful…

Ditch Excel, don’t pay crazy prices for a glorified Excel document, and don’t vibecode a solution that kinda works.

Check out goClose - get a free month-end checklist that includes full workflows, approvals, automations, etc. Also check out its reconciliation platform, flux analysis, and free accounting templates.

Free Close Checklist

In the last OnlyCFO newsletter, I discussed the most expensive cap table mistakes I have seen. The short version is that if you don’t properly track and follow the rules then it can be very expensive and time-consuming to fix. And the longer you wait the more expensive and harder the issues are to fix.

If you have experienced any of these cap table issues, then you will do everything you can to avoid them for the rest of your career.

I am in that boat…

Below are four actually useful ways I have found to leverage AI in cap table management. Lots of time saved and significantly more confidence that everything is accurate.

1. Pending Grant List AI Preparer

Prior to every board meeting, someone on my team (HR/legal/Controller) drafts the list of grants that need to be approved by the board. The below format makes it easy to review and spot-check for concerns (for both me and the board) - “are we hiring too many VP+ folks?”, “is the fully diluted % of grants too generous?”, etc

This list isn’t something you want to mess up because if the board approves the wrong thing then it will become difficult (and embarrassing to fix with the board).

How To Use AI To Prepare The Grant List

• Create a project “Pending Grant List Creation”

• Connect your folder with all new hire offer letters, promo letters, and any other one-off grant letters

• Create a skill/instructions on how to create and/or review the pending grant list

If you are having it create the list then it can look at all letters signed since the last board meeting and also do cut-off issue checks to make sure there is no timing issue (ie no pending grants fall through the cracks).

I have seen several fat finger mistakes (eg granting 20,000 options instead of 2,000), entering someone’s name wrong, or putting the wrong year on the vest start date. If you leverage AI to create and/or review the list then these are MUCH less likely.

Want to Make it More Automated?

• Crawl: Create a project (like I said above) and manually tell ChatGPT/Claude to create the pending grant list

• Walk: Create a “scheduled task” (eg. every Monday at 9am) that pulls from your connected systems and automatically updates/reviews the pending grant list.

• Run: Create a trigger (using Zapier, Workato, etc.) so every time a new offer letter is added to the folder, it automatically updates the pending grant list in real time.

The “run” option might be overkill for many things. I use the “walk” approach for almost all equity-related items so I don’t have to remember to prompt Claude. Then it will just slack me either weekly/monthly/quarterly on things I should look at.

2. Equity Reconciliation Agent: HRIS <> Board Minutes <> Carta

This reconciliation can save a lot of equity issue headaches by catching issues early. It should be done at least quarterly.

• Control: Confirming all equity grants are approved by the board and that it matches the HRIS and your equity tool (eg Carta)

• Without AI: Every quarter/month someone manually compares board minutes of approved grants to what was added to Carta, making sure all start dates, grant size, etc match Carta. And that should match the HRIS (if you put equity data there)

How To Use AI To Perform the Reconciliation:

• Create a project “Employee Equity Reconciliation”

• Connect your folder/s with 1) board minutes, 2) equity letters (new hire offer letters, promo letters, etc), and 3) HRIS reports (if not connected via MCP)

• Create a skill/instructions on what to review, any analysis you want performed, what the output should look like.

The goal is to have AI normalize the three sources and give you an exception report instead of making someone compare every row. For example: grant approved for 50,000 options but Carta shows 500,000; vesting start date doesn't match the offer letter; employee exists in Carta but not HRIS; approved grant is missing from Carta, etc.

Below are all the exceptions that AI found doing this reconciliation.

And it’s important to continue to iterate on the skill/instructions. Every time you find a new edge case, add another check. For example, my team found the last issue above (Maya Patel), so we added this check to the Pending Grant List AI Preparer: “Review all pending grants and confirm none were already approved in prior board minutes.”

3. 409A Reviewer Agent

A 409A Valuation Report is performed at least annually, but could be more frequently (fundraising, material event, tender offer, close to an IPO, etc). It’s important to get right because it determines the strike price for all equity grants during the period, impacts employee taxes when they exercise, and your auditors will review it in detail.

TLDR: You don’t want to mess up your 409A report.

The experts preparing the 409A aren’t reviewing most of the inputs. They are just trusting management on historical financials, forecasts, and other relevant data points.

Reviewing this manually is time-consuming and error-prone.

How To Use AI To Review the 409A:

• Create a project - “409A Report Reviewer”

• Point to the folders with the relevant support for the review (historical financials, forecasted financials, cap table report, secondary transactions, etc)

• Connect Carta (or other equity tool MCP) - for the 409A review if you connect to Carta MCP then it can validate the numbers as well and check for secondaries (a somewhat common thing to miss in 409A reports).

• Create a skill/instructions for what to review. There are some standard things you always want to review (confirm historical and forecasted numbers tie to your latest reports, all secondaries are included, cap table is accurate, etc).

  Text within this block will maintain its original spacing when published

  But maybe there is also an analysis you want performed.
  
  When I review the 409A I always want to compare the forecast from the prior 409A with the current one to see how the assumptions changed. We only build a model for 3 years out, but the 409A goes 10+ years so are outer-year assumptions changing a lot?
  
  I used to do this manually comparing the two 409A Valuation PDFs, but now I have AI build a comparison in Excel in less than 1 minute...

Below is an example of what that analysis looks like:

4. Equity Monitoring Agent

There is a lot to potentially remember when managing the cap table and all equity-related stuff so I created a scheduled agent that updates me every two weeks on things like the following:

• “409A expires in [X] days”

• Reviews emails and slacks for things that may impact 409A. “You just signed a term sheet for a Series D. Consider pausing additional grants until you have a new 409A completed after fundraising completes”

• “Your remaining option pool is 3.8% and if all pending grants are approved your remaining option pool will drop to 2.1%”

• “Carta identified these cap-table red flags:”

It can check all of these things because Claude has access to my folders with 409As, Carta MCP, HRIS MCP, etc

Footnotes:

• Check out goClose for month-end close (and get a free full month-end close tool)

• Reply to this email and tell me what AI stuff you want to see

Subscribe now

Want a free month-end close checklist tool?

Sign up and ditch your Excel checklist (no credit card and no sales call). Create checklists, set approvals, track timelines, leave comments, etc. Smart month-end close checklists for the low low price of $0 (limited availability).

Free Close Checklist

Open-Weight vs Closed AI Models

I am by no means an expert on this topic but I do think there are a few things that every executive (especially CFOs) should understand about open-weight AI models.

This is not about the politics, the distillation/theft topic, the risk to humanity, etc. It is about when, how, and why open-weight models make sense.

Let’s start with some definitions because I don’t think a lot of non-technical folks actually understand the model types or how they can be deployed:

• Closed AI Model: These are the AI models that most people think of (ChatGPT, Claude, etc). These closed models have been the “frontier” models because they are so far ahead of everyone else. You can’t download the weights, run the model on your machines, and customization is limited (if available at all).

• Open-Weight AI Model: The “weights” are downloadable by anyone to host or sell (subject to the license). It can be customized and trained to produce better outputs for a company.

Weights: the billions of settings inside the AI model that generates the prompt outputs (kinda like software code). “Fine-tuning the weights” means you adjust those settings until it gives better answers.

Closed models are like renting a fully furnished office (WeWork style). Open weights are getting the building plans and keys and then told you have to figure everything else out. You can knock down walls, and change things, but the HVAC bill is now your problem.

Why Use Open-Weight AI Models?

There are two primary reasons you would use open-weight models over the closed frontier models:

1. Cost - These models can be ~80%+ cheaper than the closed frontier models. Note - often these savings can be misleading (more on that below)

2. Model Control - below are a few reasons why a company would want to control the model:

   1. Fine-tuning weights - ability to fine-tune may help deliver stronger output for your specific use case.

   2. Data/security control- you host so you control all the data, who has access, privacy, etc.

   3. No vendor lock-in - closed model providers make you sign contracts, can raise prices and you build around their model.

Deployment Type Matters

To evaluate the benefits and potential cons, you need to understand the two ways these open-weight models can be deployed.

• API / hosted: This is just like how you use Claude/ChatGPT, but anyone can download open-weight models and sell it to you via API.

• Self-hosting: You download the model and run it on your servers (or a rented one).

Model control (benefit #2) is highest when you self-host an open-weight model, but you can fine-tune weights via API if you buy a dedicated environment (if offered by the provider).

There are also possible cost savings in self-hosting the open-weight models, but it is often more expensive than most think…Similar to buying software, there are always hidden costs (hardware, training, people, etc). Self-hosting is often not the right answer unless the fine-tuning or the data/security benefit is a strategic advantage for the company.

Intelligence vs Cost Trade-off

Similar to tokenmaxxing, every company has been on a race to adopt AI and beat the competition. Using the more expensive closed models was a fairly easy answer for a long time because they were sooo much better.

But a couple of things are happening:

• The gap is closing between open-weight and closed frontier models (only maybe a few months behind now versus 12+)

• Incremental intelligence increases may not matter for most tasks. In many areas, companies don’t need frontier-level intelligence anyway.

Kimi K3, open-weight model released earlier this month, ranks as one of the top models across most benchmarks (apparently it is the best at Excel-based tasks).

While Kimi appears to be just slightly behind the latest models from claude/chatgpt in terms of intelligence, it is 50-60% the token price. And other slightly less intelligent open-weight models are significantly cheaper (like DeepSeek).

Lower Token Pricing Doesn’t Mean Cheaper

The key financial metric is cost per task. Price per token is one variable of the equation.

A lower price per token is only relevant if it can complete the same task with the same or less tokens than similar models. Some models are much more token-hungry…

Kimi K3’s token prices are 40% lower than Opus 5 (released a few days ago) for just a slightly less intelligent model. ROI seems really high for Kimi, right?

But…on longer-running tasks (like the benchmark below) Kimi K3 was actually more expensive than Opus 5 High (and Opus 5 High scored higher). So this frontier closed model is actually cheaper than the open-weight one!

The cheapest tokens do not always produce the cheapest completed task when the model is token-hungry (ie inefficient).

While open-weight models are usually less expensive on a per token basis, make sure you understand the completed task cost difference as well. Open-weight often wins there too, but the difference may be smaller than you think.

Cheap tokens can be like flying Ryanair: Seems cheap until you realize there are three connections and an overnight layover to get to the next state over.

Ryanair’s social media game is fire though…haha

What About Security Issues With Using Chinese AI Models?

I have talked to many non-engineering folks who think this…

The model may have been created in China, but you (or model providers) can download it and host it in the U.S.

This is what a lot of DeepSeek or Kimi usage is. Other companies download, host, and sell these models via API. China doesn’t have to ever see it. Several U.S. companies provide access to these models via API.

What CFOs Should Do

Most companies probably shouldn’t self-host the open-weight models (particularly the large ones). It’s usually too expensive to be worth it unless you are huge (then maybe you get the economies of scales) and/or there are specific reasons why you need to have greater control of the model.

But everyone should try different models via API. You don’t need the latest frontier model for most tasks and the security/privacy risk isn’t what most people think when they hear about open-weight models created in China…

Importantly, CFOs should understand the terms of service and actual cost per task. If companies can use the frontier models for ~10% of the hardest tasks and cheaper models for everything else, then they will save a lot of money. Most should be able to do this.

Footnotes:

• Subscribe and forward this newsletter to your team.

• Check out goClose for month-end close (get their free close checklist tool)

• Reply to this email and tell me what AI stuff you want to see

Subscribe now

Subscribe now

ICONIQ just dropped their 2026 State of AI Report. There are a few slides finance folks should take a closer look at:

Token Spend

Every CFO is trying to figure out the “right amount” of token spend. We don’t want companies wasting money but we also don’t want to slow down the business at such a pivotal time.

It’s a tricky place to be in. Right now it is likely leading to overspending at most companies, but it has been an acceptable tradeoff for most companies trying to maintain high growth.

The chart below shows token spend for internal-facing tools.

In general I think it’s stupid to spend a relatively large amount on internal tools when your headcount is small. Your team is probably too focused on internal-facing stuff and not product.

Building Is More Expensive Than You Think

While everyone can build now, not everyone fully appreciates the cost/time that goes into building a fully working internal app (particularly your non-engineering teams).

And then if you are embedding AI into the tool you are building then those costs can also rack up quickly. And then the maintenance, bug fixes, new use cases, etc….so few companies have really built something internally before that they are facing a lot of this for the first time now. Cost is usually higher than you expect.

AI Not Taking Finance Jobs Yet?

Somewhat surprised to see FP&A automation at the bottom of this list of productivity impact from AI. Also, interestingly, it is one of only two functions where higher growth companies see lower productivity gains from AI. Kinda makes sense… there is a lot more art than science (that AI can handle) when a company is growing fast.

Why will we grow 100% next year? “Um…because the CEO said that’s what he wants.”

AI Use Case: Close Faster

FYI - every board member I talk to thinks their portcos should be closing the books faster… Both accounting and FP&A. Everyone wants their numbers sooner!

💡 Plug: I am working with goClose to build cool accounting tools (will either be free or very affordable) for companies who don’t want to spend tens of thousands of dollars yet and don’t have the time to vibe code something good.

What Products Are Being Built?

More folks are going after the verticals as their differentiator in a sea of sameness. But also, AI can be a lot more powerful when vertically focused.

Cheaper Models Are Gaining

We want the best but we are starting to see now that much cheaper models get the job done just as well. Also, some of the anxiety around open-source models is disappearing as companies find solutions to reduce concerns.

Most Popular Models

No big surprises with the most popular models, but it is interesting that companies on average use 3.3 model providers (slight increase from 6 months earlier).

This is likely at least partially from a push to be more efficient in AI spend. You still get the best model (Claude/ChatGPT) but we want to use low-cost models for other things.

Pricing AI Products

A lot more variability in pricing strategies today…

Companies are spending more time on their pricing strategy in 2026 (rightfully so), but I think most aren’t experimenting and updating their pricing often enough

Gross Margins Are Rising

Can’t use AI as an excuse for low gross margins forever… Make sure you have a plan for higher gross margins with your AI products.

Final Thoughts

A lot of the data from this report just proves that benchmarks go stale REALLY quick in the age of AI. Things are moving too fast to rely on benchmarks from a report 6 months ago.

You kind of have to just ignore what everyone else is doing and use your brain…

Check out the full ICONIQ report here.

Footnotes:

• Subscribe and forward this newsletter to your team.

• Reply to this email and tell me what AI stuff you want to see

Subscribe now

Subscribe now

We have all shipped something important only to realize a few minutes after we hit the send button that there was an embarrassing mistake.

• Board deck with an internal consistency error

• Planning model with a formula error

• Financial statement rounding issue

I made a mistake in my OnlyCFO newsletter last week where I accidentally showed Shopify’s stock price falling by over 90% over the past 5 years (in a table with lots of other numbers). It didn’t….Shopify had a 10:1 stock split and my starting stock price was capturing pre-split for some reason. 39,000 of my closest friends saw the mistake….

After being upset for missing a silly error, I realized that I didn’t have the same AI quality checkers for my newsletter as I do in my actual job.

One of my favorite “AI agents" I use all the time is the Finance QA (quality assurance) Analyst.

How To Create a Finance QA Analyst

The Finance QA Analyst has already caught many potentially embarrassing mistakes on board packages and investor updates. I use it every time I put together something important.

If I was using something like this for my newsletter, I would have avoided the embarrassing Shopify price error. I corrected my mistakes and have now implemented a Newsletter QA Analyst to check my work.

*I built my QA Analyst in ChatGPT, but you can do the same things in Claude.

Step 1: Create a Project

Scope projects tightly for the best results. The more you try to combine, the longer the instructions and more chance of AI confusion/errors.

A Board Deck Review project should likely be separate from a Forecast Model project.

Step 2: Feed AI Context

When people complain about the output of AI it is often a context problem. The more context and direction (see next section), the better the output.

Project Sources:

In ChatGPT projects you can upload files or link to Google Drive folders/documents to provide context. There is a lot of context you can provide but below are a few things that I uploaded for my board deck review:

• Metrics Dashboard: confirms all board deck numbers agree to our dashboard

• Company Design Guidelines: logo, colors, formatting rules, and examples so it can keep materials on-brand.

• Board Decks: I provide all prior board decks and have it look at the most recent ones. With this context, the AI agent can compare the current material to other board material. If ARR in Q2 was $52M and now we are showing $51M for Q2, the QA Analyst should flag it and check why previous numbers changed from what was already presented.

MCP Connections:

MCP connections / custom apps are separate from project sources. In ChatGPT, these show up as Apps. Project sources are the files, docs, dashboards, and folders you want ChatGPT to reference inside the project. Apps/MCPs are the external systems ChatGPT can access when they are connected and available in the tools/app menu.

In your Skill/instructions, tell the Finance QA Analyst exactly when to use those connections. Example below:

When reviewing board deck revenue, check the financial report in Snowflake before signing off.

You want to provide fairly tight instructions in the Skill so it doesn’t just go looking everywhere via MCP.

Step 3: Create Your Skill/Instructions

💡 Pro tip: In ChatGPT I put the “Skill” (which ChatGPT just refers to as instructions) in a Google Doc so I don’t have to manually update the Skill/instruction each time. So then as I want to make changes to the instructions I just prompt ChatGPT to update for [X] in the instructions.

In the project settings I just tell it to reference the “Finance QA Analyst Skill”, which is a Google Doc that is linked in the project sources.

Skill/Instructions: Describe what you want the QA analyst to do in your Skill file. Below are a few things I have listed in my Finance QA Analyst:

• Internal consistency in numbers - very embarrassing to have the same metric with two different numbers in the board deck

• Review grammar and spelling

• Review formatting consistency - review prior board material for formatting

• Apply company design - reference our marketing company designs that is linked as a source

• Confirm numbers from source data - call relevant MCPs to confirm data from ERP, Snowflake, etc

For every important metric, tell the QA Analyst which source wins. ARR might come from Snowflake, cash from ERP, headcount from HRIS, bookings from Salesforce, and budget variance from the planning model. The QA Analyst shouldn’t guess. It should flag mismatches and tell you which source disagrees.

Finance QA Analyst Output Instructions:

It is important to define what and how you want the AI agent to provide feedback. Otherwise you will never read it…

In my instruction file I tell it to be concise and order it in the below categories:

• Major issues

• Must-fix before sending

• Data items not verified

• Grammar/spelling

• Open questions

Stop Making Stupid Mistakes

I don’t use AI to write anything for my newsletters. I don’t like how it writes and I enjoy the writing process myself. I do use it for research though. And I am now using it for quality review. Let’s see if I make any obvious errors like the Shopify one in the future…

Anyone putting together important materials (board materials, investor updates, exec presentations, fundraising, M&A stuff, etc) should create a smart AI agent to double-check your work.

Use AI to prevent embarrassing mistakes.

Footnotes:

• Subscribe and forward this newsletter to your team.

• Reply to this email and tell me what AI stuff you want to see

Subscribe now

Subscribe now

Just a few weeks ago Anthropic announced Mythos, an AI model so dangerously proficient at cybersecurity hacking that they paused releasing it to the public until large companies and security firms could use it to find their security vulnerabilities first.

TLDR: Anthropic was worried (at least for marketing purposes) that Mythos could disrupt the global economy…

Yesterday they released a safety-modified version of Mythos called Fable 5.

Fable 5 is VERY Powerful

Everyone who has used Fable 5 (me included) reports a major jump in capabilities. And the benchmarks reflect that jump…

Fable 5 is EXPENSIVE

With great power comes great costs…

Fable 5 is 2x more expensive than the previous most expensive model (Opus 4.8).

Currently Fable 5 is included in the subscription plans (subject to existing usage limits). But in a couple of weeks that changes. You will only be able to use Fable 5 with usage credits (pay-as-you-go). Anthropic did say they hope to bring it back to subscriptions soon. Anthropic is compute-constrained and Fable 5 is expensive, so they are limiting usage.

Fable 5’s Privacy Concern (breaking NDAs?)

Using Fable 5 means your prompts and the model's outputs sit on Anthropic's systems for 30 days and can be human-reviewed for safety.

It overrides the zero-retention deals enterprises negotiated. This applies even if your enterprise agreement says otherwise…So uploading anything under an NDA or other confidentiality obligations (M&A, earnings announcements, etc.) would likely put you in breach.

Don’t automatically turn on Fable 5 for all users and make sure users know not to upload confidential information to Fable 5. You could be breaking NDAs…

Three Finance Use Cases With Fable 5

I haven’t used Fable 5 with real data for finance use cases yet given the privacy concern, but below are a few things I am excited about (that I have tested with some dummy data).

1. Drafting S-1s

Gabe, President of Harvey, demonstrated that Fable 5 is a huge improvement for drafting an S-1 with better formatting, structure, and writing.

Similarly, Fable 5 is much better for drafting board memos, board decks, and other long, complex, analytical work.

2. BvA Commentary and Accounting Flux Explanations

A lot of time in both accounting and FP&A is spent researching and documenting financial trends, budget vs actuals, etc. To date, most finance folks have been disappointed by AI’s ability to draft this commentary.

But Opus 4.8 was already saving my team a bunch of time so given Fable’s jump in benchmarks, it should be really good at this.

3. ASC 606 Revenue Recognition Review

Fable 5 is particularly strong with document reasoning and research.

I currently have the below built out using Opus 4.8, which is good. I tested Fable with dummy data and it blew me away. It catches things I would have missed and documents it exactly how I would think a director-level revenue person would…

We created an AI agent that has our ASC 606 policies, customer-specific memos, and historical contract reviews.

1. Reviews draft quotes and flags potential issues to the sales rep before it’s even discussed with the customer.

2. Creates an ASC 606 summary for each contract that documents key terms, rev rec considerations, etc

In Summary…

Make sure your IT team carefully rolls out Fable 5. It’s very expensive and has privacy concerns you want to watch out for.

But…Fable 5 is a major leap on the benchmarks that matter for finance and accounting. Once the privacy concerns are resolved, revisit use cases in finance and if/when Fable should be used (versus a less expensive model)

Footnotes:

• I am opening up sponsorships for the first time for this newsletter (CFOpilot). Reach out now and get amaaazing prices! Email onlycfo@onlycfo.io

• Subscribe and forward this newsletter to your team.

Subscribe now

Subscribe now

For this post, I asked Alex Altman (Head of Finance at Coram) to explain how he built a small army of AI agents to automate a lot of his review work and catch things he’d normally miss across finance, accounting, and sales ops. He’s built an agent system that’s keeping his team lean and more productive than ever.

How I Built a Team of Finance AI Agents

—by Alex Altman

At Coram I own finance, accounting, legal, HR, ops, inventory, and shipping. It’s a lot to cover while we’re growing quickly. Instead of immediately hiring more people, I started building AI agents for every major function that was higher risk and took up my time.

It was fairly easy to set up with immediate ROI:

• Output has improved (numbers are more accurate, fewer mistakes, etc)

• I’ve freed up a lot of my time for other things

The agents give me better visibility and confidence into the areas I own but don’t run day-to-day, and they take direct work off my plate where I am the one doing the work.

Each agent watches a specific part of the business, runs on its own schedule, and only talks to me when something needs my attention. They all live in Slack and they’re all connected to our tools - ERP, CRM, AP, procurement, Google Drive, and more.

Here’s what I’ve built so far, and where it’s going.

How these actually get built

Before the agents themselves, the practical part - because this is more accessible than most people think, and there are two paths.

Path 1: MCP connections. You connect Claude directly to your tools through MCP. A finance person can do this immediately, no engineering required. The tradeoff is that each agent queries those APIs live every time it runs, which is slower, can get expensive, and makes it harder to trace a flag back to its source. Great for getting started and proving value.

Path 2: Centralized database. Once you find this valuable, you move the underlying data into a database like Supabase, and the agents read from that instead of hitting each API live. Everything lives in one place, every agent works off a consistent source of truth, and when something gets flagged you can drill straight through to the underlying data. This is what makes the whole thing sustainable at scale.

The build mechanics are the same regardless of path. For each agent you:

• Write a skill file that defines what the agent is responsible for, the data sources it watches, and exactly what should get flagged to you

• Set up a scheduled prompt so it runs on its own cadence

• Tell it to notify you in Slack, or wherever works for you (that instruction lives in the skill file too)

And if the above sounds complicated…it’s not. You just tell Claude and it will do it for you!

One rule that matters: give each agent a clear, narrow responsibility. Don’t try to make one agent do everything - the instructions get muddy and the results get worse. And expect to tune the flags over time, because too many flags is just noise and too few means you’ll miss things.

Building My AI Agent Finance Team

I have segmented my finance AI agents into four categories based on what they do.

1. Watchers (AP, AR, GL, Cash) monitor transactions and operations in real time

2. Analysts (variance engine, financial analyst) compare results against expectations

3. Investigators dig into flags, filter signal from noise, and decide what to escalate

4. Controller routes escalations to the right person and confirms resolution

The AP Agent (watcher)

The AP agent reviews every bill and every credit card expense that hits our accounts. For each one it does three things:

1. Checks Google Drive to see whether the vendor has a contract on file, and flags any new vendor without one

2. Compares the GL coding against how this vendor has been coded historically, and flags anything different

3. Compares the size of the bill against historical bills from the same vendor, and flags outliers

The simplest way to set this up: upload your AP and expense policy, and tell the agent what you want flagged to you.

I get an alert every morning with the day’s bills, what passed, what was flagged, and what the agent recommends for each one.

The first week it ran, it caught a bill from a vendor we’d been paying for months that didn’t have a contract saved in Drive. Filing discipline I’d been meaning to enforce, now enforced automatically.

The AR / Sales Ops Agent (watcher)

The AR agent watches the other side of the business. For every closed-won deal in our CRM it checks three things:

1. That the deal record is clean - a number of fields are supposed to auto-populate from the signed quote, but they don’t always populate correctly, so the agent verifies the numbers tie

2. That an invoice has actually been raised for the deal

3. That any overdue invoices have a documented response or collections plan

It runs every three hours and gives me a daily summary.

It’s a sales ops agent as much as a finance agent, because CRM cleanliness is a constant battle. The agent doesn’t fix the data itself, but it surfaces every issue so I can route it to the right person.

Right when I set this up, it caught four closed deals where line items hadn’t flowed through correctly from our quote-to-cash pipeline. Without catching it, we would have had incorrect revenue, COGS, and ARR downstream. I fixed the deals manually before quarter-end close and sent the underlying bug to our engineer. That one catch alone paid for the time it took to build.

The GL Review Agent (watcher)

This agent watches the books directly. Every transaction that hits the GL - journal entry, expense, bill, anything - gets evaluated against historical patterns and the prior period. It only surfaces transactions that look off: a coding that doesn’t match how similar transactions were handled before, an amount outside the normal range, anything unexpected. The clean, normal transactions never reach me.

This is the cleanest example of how I think about using AI in finance. The accounting itself happens deterministically through the systems I’ve already built. The monitoring happens probabilistically, with the agent watching for anomalies. I look at the flags when there are flags, and otherwise I leave it alone.

The Cash Agent (watcher)

The cash agent watches the thing every founder actually cares about: where the money is and where it’s going. It pulls our bank activity, AP, and collections together and maintains a rolling 13-week cash forecast, so instead of someone rebuilding a cash model every couple of weeks, it’s always current. It flags the things that move the picture - a big outflow week coming up, collections slipping behind, runway dropping faster than expected.

What makes it powerful is that it rolls up into the controller alongside everything else. If the cash agent sees collections falling behind, that’s not just a number on a forecast - it becomes a routed action item to the right person to chase the specific overdue accounts driving it. It connects the “what’s happening to our cash” question to the “who needs to do something about it” answer, which is usually where cash forecasts fall apart.

The Analyst Agents

The agents above all look at transactions as they happen. The analyst agents do something different - they look at results against expectations.

I actually have two, because they do different jobs.

1. Variance Engine Agent. It compares actuals to plan every Monday morning across the metrics that matter (Net New ARR, Gross Margin %, Runway, Burn, etc) and surfaces anything beyond a defined threshold. Those thresholds live in a simple config file I can edit (any metric more than X% off plan, any opex line over $Y in variance, departments more than Z heads behind hiring plan, and so on), so I can tune what’s worth flagging without touching the underlying logic.

2. Financial Analyst Agent. This one doesn’t compare to plan - it studies the actuals themselves and asks “what’s odd here?” MoM and QoQ deltas, gross margin drift, DSO trends, customer concentration, inventory days, mix shifts. It runs monthly after close and produces a handful of numbered observations, each with a quantified change, a hypothesis, and a suggested follow-up. This is the “what does Alex ask his analyst at 7pm on a Sunday” content.

Where the variance engine asks “are we tracking to plan,” the financial analyst asks “what should we be paying attention to that we didn’t think to ask about?”

Both feed into the investigation agent.

The Investigation Agents

Flagging something and explaining it are two different jobs. The analyst agents are good at noticing that hardware margin compressed. But figuring out why requires its own toolset and its own playbook - pull SKU costs, look at average discount and ASP, check for allocation errors. Every domain is like that. So the investigation agent isn’t one generalist. It’s a set of specialists, each with its own analytical playbook, and it routes a flag to whichever specialist is right for it.

The proof it works: on its sixth run, the working capital AI agent self-reported a data gap. It said “inventory data missing from working capital pack” and stopped. That’s exactly the right behavior - a specialist using its tools, finding the data wasn’t there, and being honest about it rather than making something up.

The Controller Agent

This is the layer that turns a pile of agents into an actual team. The watchers and analysts all generate flags, the investigator decides which ones matter, and the controller takes it from there - it routes each escalation to the right person and owns the follow-through. CRM cleanliness issues go to sales ops, billing issues go to our accountant, AP issues go to our admin, and so on. It tags the person in Slack, makes the ask clear, and then actually checks back to confirm the work got done instead of letting it disappear into a backlog.

Once this went live, I stopped being the router. Before, I was the one reading every agent’s output, deciding who needed to act, and chasing it down. Now I only get pulled in when something genuinely needs me. The controller handles the coordination layer that used to eat a surprising amount of my week - and nothing slips, because following up is the agent’s job.

The Finance AI Agent Stack

Once you put all of these AI agents together (watchers, analysts, investigators, and controller), you create an AI team that does meaningful work. Your numbers and accuracy improve. Things get done faster because the agents work 24/7. There are more controls as agents ensure your policies are followed.

Start building AI agents now and you’ll quickly see the ROI. They won’t get everything perfect the first time (just like a new hire), but continue to refine the skills and they will get better.

Footnotes:

• I am opening up sponsorships for the first time for this newsletter (CFOpilot). Reach out now and get amaaazing prices! Email onlycfo@onlycfo.io

• Subscribe and forward this newsletter to your team.

Subscribe now

Subscribe now

CFOs Are Reviewing AI Model Costs

The average business is spending 13x more on AI tokens than in January 2025, according to Ramp’s spend data 🤯. AI is quickly becoming the largest vendor expense on your P&L and it may soon (if not already) crush your gross margins as well.

Internal AI model adoption is being pushed harder than ever (tokenmaxxing) and everyone is releasing more AI products with worse gross margins. Growth on AI spend has gone exponential in the last few months.

The size of our AI model (chatgpt/claude) bills have gone from nothing to pretty massive in a short period of time. And now every CFO should be paying close attention.

In this post, I will share some of the things that companies can pretty easily do to slash their AI bills.

AI Model Pricing

The two charts below show the different AI model pricing for the two way communication that a user has with AI:

1. Input tokens: Your prompt, convo history, attached files, tool definitions, etc

2. Output tokens: What is returned (text, file, etc), tool calls made, code returned, etc

The big take away is that the “smartest” AI model is 5x+ more expensive than the cheapest model.

You should also understand that there are three ways that AI model companies will bill enterprise licenses:

1. Seat-based pricing: flat fee paid annually upfront for the privilege of simply accessing the platform (~$18/user/month)

2. Claude/ChatGPT usage: billed monthly in arrears based on actual usage (on top of the seat-based pricing) from the rates noted in the charts above

3. API calls: Paid upfront on a credit card. This is for your product that you sell or for internally built apps that are hosted.

How We Cut AI Model Costs By 40%

TLDR: You don’t use a heart surgeon to take your temperature. Similarly, you shouldn’t have your CFO reviewing T&E expense reports.

This is the primary problem at most companies.

They are using the most expensive model (or even the 2nd most expensive model) when they could be using a cheaper one that provides results that meet their needs. There is HUGE potential savings by simply using the right model.

Let’s say you have a $100K/month bill for Claude:

You currently use Opus 4.7 (the most expensive model) for everything, but if the company was to only use the model it actually needs then it could look something like the below:

Opus is rarely needed for every task. So if users shifted to using Sonnet and Haiku when appropriate then the company would save nearly 50% ($576K per year)…

We want everyone using AI and adopting into all their workflows, but they don’t need the most expensive AI model for every task. When the Anthropic bill was small, no one cared. But as the costs have grown, there is significant ROI in making sure you are using the least expensive model needed to get the job done.

Most folks on an enterprise usage plan are paying premium rates (the heart surgeon) for basic responses (checking your temperature).

How Do You Actually Do it?

There are two different paths based on how AI is used and priced:

1. Enterprise subscriptions/usage

2. API calls

#1 is where almost all your internal AI usage comes from.

Enterprise Subscriptions/Usage (internal AI usage)

You have fewer controls here because users can choose which model they use, but there are a few things you should do that will save a lot of money.

Below are three things that every company should do:

1. AI Admin Controls

The Claude/ChatGPT admin should set the default model to the middle tier (e.g. Sonnet for Claude). If you do that, then 90%+ of users will just use that model. While they usually won’t be using the most expensive model, they will also almost never downgrade to the cheapest model. But just doing this may save you up to 40% (price difference between Opus and Sonnet).

Benioff said they will spend $300M on Claude this year. And I have also heard from various people that everyone just defaults to Opus for everything. They could probably save $100M+ by using cheaper models when appropriate…

User Usage Limits: Obviously you should put controls on users so they don’t create something that just blows up AI usage. Admins can put dollar limits by user (e.g. engineers get $5K per month while finance gets $1K per month)

The below was just reported by Axios 😱. Maybe not entirely true (or missing some details), but you MUST have spend controls on your AI stuff so this doesn’t happen to you.

2. Admin Manual Tracking

Admins get dashboard reporting that show what each user is spending and what models they are using. If someone is burning a ton of tokens using the most expensive model on easy tasks, then the admin should have a conversation with those users and encourage them to use something else.

Don’t decrease AI usage…but choose your AI models more wisely.

3. Employee Education

Make sure employees understand the cost difference between these models. Many just don’t know. Forward them this article :)

There are also several other things (besides different AI model costs) that you can teach employees so they spend significantly less on tokens. For example:

• Use a new chat window for new tasks. Each time you prompt in the same chat window it re-reads the entire chat history. This gets expensive quick (especially if you are on a premium model). I have seen people just default to the same chat window every time they go to AI…

• Be upfront with everything you need in one prompt. Back-and-forth exchanges compound quickly. Every reply re-reads the full history. Give Claude the full context and ask for exactly what you want in one shot.

• Update the AI model in your automations, agents, skills, etc. A lot of scheduled tasks are structured and repetitive (often fine for a cheaper model). Users can define which AI model to use in agents/skills/automations. Test out a lower tier model and see if it works fine.

API Calls (internal apps and your AI products)

API calls happen when you aren’t in the ChatGPT/Claude app. It is when you build an internal app or customers use your product and it calls an AI model for something.

There are more options here to control what AI models are used in the response, which is great because you will want to control it to improve AI gross margins.

Model Routing: directs AI requests based on a defined criteria to the right AI model

The goal is to use the cheapest passing model for each task. “What is the minimum requirements we are OK with to check someone’s temperature and can they do just as good a job as the heart surgeon?”

I talked to one CFO that saved a few percentage points on their AI gross margins after they switched from the most expensive AI model embedded in their product to a model routing system that switched a lot of tasks to a cheaper model.

While a well-structured model routing process is going to have the highest ROI, there are other ways to also cut down costs on AI API calls. Some examples I have done:

• Analyze your actual usage: Anthropic’s console shows token usage by API key. You can also tag calls by feature or workflow and see exactly what is driving costs. Do this so you can figure out how to make the expensive stuff more efficient

• Batch API: If a task doesn’t need a real-time response, you can use Anthropic’s Batch API and save 50%. A lot of stuff doesn’t need to be real-time

• Prompt caching: If you’re sending the same large context on every API call (a system prompt, a reference document, a knowledge base), then you’re paying full input price every single time. Prompt caching stores that content so repeat calls cost ~90% less on the cached portion. Make sure your engineering team built this…

Final Thoughts

AI spend has become very large in the last few months. It’s time for CFOs to pay attention. We want to drive AI adoption. It will likely be the difference between companies that survive and those that die in the AI world.

But…

It doesn’t mean we should just light money on fire and throw all of our investors’ money to Anthropic/OpenAI. It’s time to add more process and educate employees so you can cut your largest vendor bill in half.

Footnotes:

• Email me and tell me what you are building. Or reach out with questions!

• Subscribe and forward this newsletter to your team.

Subscribe now

AI Stuff From The Week:

The below tweet blew up because I think it feels so real for all finance leaders right now… This was me reviewing the April usage bill lol.

And it’s why I spent the last two weeks pushing all the things I wrote in this article!

Hearing many cases of people double-clicking into the ROI of their AI spend. It’s not a question if AI is valuable (it obviously is), but rather is how much they are spending and how they are spending it have decent ROI.

Subscribe now

AI Agents for Accounts Receivable

I built an accounts receivable AI agent in <10 minutes. It worked surprisingly well right out of the gate. And it’s only getting better as I fine-tune it.

AR/collections is a near-perfect use case for AI…incredibly boring, repetitive, and requires context from multiple places.

A lot of finance folks think it’s too hard or it requires a lot of technical expertise to build this, but here are some facts about my AI agent:

• Built it myself (with ChatGPT) in <10 minutes

• Didn’t need to build an app for it

• No separate API calls or complex integration work (all within ChatGPT/Claude)

• Was immediately useful. And is getting better every day.

• I control what gets just drafted to my email versus auto replies. Everything started as email drafts that we reviewed first.

Save your AR/collection team hours every week and keep reading 👇

The Collection/Billing Problem

Invoices get automatically sent from the ERP. And then the ERP has automatic “dunning” reminders for unpaid invoices that look something like the below.

Hello Customer - Your invoice [xxxx] is past due by 30 days. Please provide payment as soon as possible.

But humans have historically handled all the subsequent back and forth with customers, which can take a lot of time…

• Customer: “Can you send me the contract and all the outstanding invoices?”

  • Accountant: Open ERP, find relevant invoices, download invoices, attach them to email. Then open up the CRM, download the relevant contract and then write an email response.

• Customer: “Can you send me your W9?”

  • Accountant: Grab latest W9, attach it, and email a reply

• Customer: “We will pay the invoice next Friday.”

  • Accountant: Read email, updates the collection tracker, turns off automatic dunning notifications for that customer, email a reply that says, “Thanks! I’ll look out for it”

This is where my AI Accounts Receivable agent comes in….

How I Built My Accounts Receivable AI Agent

The Context Layer

The main reason building this only took me 10 minutes is that I already had ChatGPT connections (via MCP) to all the necessary tools. If your AI model doesn’t have context, then you or your IT team needs to do that immediately. You are greatly behind already if you haven’t done this step yet…

The Initial Prompt

I built the AR agent in ChatGPT’s Codex. Below is my initial prompt.

Prompt: I want to automate billing/collections emails with AI draft replies. The model should already be connected to my email, CRM, ERP, and Slack. Let’s build it so it runs via Codex (not a local or hosted app) and it triggers every 10 minutes via automations so email drafts get pushed to Gmail frequently. There should be controls, guardrails, and email best practices and styling rules that I provide. I want to run the prompt as efficiently as possible. Suggest what to add so I follow best practices.

After the above prompt, ChatGPT asked me a few questions to make sure it was doing the right thing and then it created an “Automation” (see below)

There is no separate app. No need to do additional API calls or other wiring. It’s all done via ChatGPT and the connections that should already be in place.

It’s a simple and very effective way to get started. And probably works perfectly if collections is handled by one person. While you may outgrow this solution eventually, it can work for a long time and it will help create the framework if/when you build an internally hosted app (which adds a lot more complexity). Or it will help you decide what you need if you buy a dedicated tool from a vendor.

How It Works

The AI Wiring

• Trigger: “Cron” automation in ChatGPT runs every 10 minutes (while your laptop is open) and fires off instructions.

• Automation Instructions: Below are the simple instructions that live in the Automation tab. This initial prompt used call all the other detailed instructions that live in the project folder (which are on my desktop).

💡 Pro Tip: In the Automations settings you can change your AI model. For something that doesn’t need to be super smart (like collection emails) you can change the model to a less expensive one.

• Remaining Instructions: You could put all your instructions in one super long file in the Automation prompt, but that isn’t efficient and it’s hard to track/update. So it’s usually better to break it out. Below are a few different files I have for this:

  • Email Style (.md file): I provide some general instructions like “Keep it concise and informal.” And then I wrote, “Use the ‘approved replies’ email tag in Gmail for examples to reference when drafting emails”. This means I can improve AI drafting by crafting email responses that I like and tagging the email to “approved replies”. The AI is now instructed to use those emails as examples.

  • Guardrails (.md file): This is where I document all the rules I want the model to follow. Like “Never send emails automatically. Only draft them into Gmail.” or “Only use the ERP and the facts I provide as a source of truth when drafting emails”.

  • Runbook (.md file): This file defines when to draft a reply. Like only draft a reply if it relates to collections or billing (don’t draft replies to spam or non-related emails).

Improving AI Responses

Below was a pure AI response automatically drafted to Gmail on an example I set up. The AI model got invoice information from the example in QuickBooks and it grabbed the W9 and invoices from a Google Drive folder (via MCP connector) after confirming the email address was a contact in our ERP.

The above AI draft already saved me a bunch of time. It listed out invoices (and attached them all), attached my W9, and drafted the response. But I wanted to change one thing…so I edited it to show the days overdue next to each invoice. I then added this response to the “approved replies” Gmail label so that the AI model will use this as an example for the next similar email.

And it did. The next AI email draft response included days overdue…

And if there are other rules, guardrails, style, etc that I want to permanently add later, I will just go into the ChatGPT project and tell ChatGPT what I want and then it will update the necessary files. Easy.

Should You Have AI Auto Reply?

Not at first. Definitely not.

And for certain things, I would never have AI respond automatically. But it certainly can make sense in many circumstances (like a simple request for a W9, invoices, outstanding balances, etc) with the right guardrails in place.

But even if you don’t want to take that risk yet (I totally understand), you can have AI push the email drafts right to your email so replying is 10x faster.

Final Thoughts

The setup I have described is fast and works great when one person is managing collections. But unless you have a large volume of customer billing, one person + AI can probably manage collections for a really long time now.

However, when you need to move beyond one person managing AR emails, then this set up breaks since it is dependent on one person’s laptop and their ChatGPT account.

Building an internally hosted app adds a lot more complications (API calls, security, team-based roles and access, etc). I see a lot of people trying to jump immediately to internally hosted apps when they are working with AI because they think they have to. And then they stop because of the effort and continued maintenance required with complicated hosted apps.

If you decide you need an full-blown app then make sure it’s worth your time. This is the real build vs buy territory. And building an app (and maintaining it) is often still not worth your time. There are lots of great vendors who provide AR AI agents for companies that need more functionality and features.

Start simple.

Footnotes:

• Reply to this email if you have any AI use cases you want to share! I’d love to hear what you are building and doing in finance or accounting

• Subscribe and forward this newsletter to your team. It will help them leverage AI and be more efficient

Subscribe now

Subscribe now

My 6-Step Process to Build an Accounting AI Agent

I have been deep building with AI for the past year. Every person (leaders included) needs to be building with AI. It’s important to understand how to use AI, what to expect of your team, and how to rebuild your org for the age of AI.

I built our first AI agent to tackle accounting reconciliations (focused on prepaids initially)

Here is the process I ended up with…

An “agent” is essentially the worker that navigates all of these systems and steps to complete a task.

Step 1: The “Skill” Itself

Skill: A self-contained workflow an agent runs. A system prompt that defines what to do. MCP connectors to the systems it needs. A trigger that kicks it off. A config that drives the behavior. You build it once and it runs forever.

Every skill has a SKILL.md file. The skill defines the inputs, workflows, rules, output, edge cases, etc. My prepaid skill is ~200 lines.

You’re not writing a prompt every time you run it. You’re writing it once. And then updating it if/when something changes.

Step 2: The Config File

My config (yaml file) contains:

• Prepaid GL account codes

• Vendor → coding map

• Materiality threshold

• Partial-month convention

• Routing: who gets pinged, in which channel

When something changes, you edit the config (not the skill).

This is one thing I was doing wrong before. I was dumping everything into a prompt (or a skill). Every time something changed I would have to update the skill and re-do it.

Having a separate config file with the stuff that might change breaks things up so it’s easier to maintain. It’s like why you don’t put your entire model and all the components in one giant tab or make Excel model formulas huge.

Putting everything in one skill file also makes it so you can’t reuse your skill in other recons, entities, etc because it contains the exact detail for one prepaid recon process. To make it scalable/reusable you separate out those details into the config file. Then you can reuse the skill and just update the config file.

Step 3: MCP Connector to the ERP

MCP (Model Context Protocol): A standard way for an AI agent to plug into your software systems (Deel, NetSuite, Google Drive, etc) so it can read data, do work, and take actions in those systems on your behalf.

A surprisingly large number of companies I talk to still use AI like a simple chatbot.

The biggest unlock comes from giving AI visibility to everything…Only then can AI agents perform meaningful work end-to-end. The primary way to do that is via MCP connectors.

If AI is not connected to all your systems, then it lacks context so it will require a lot more manual human effort for each handoff and coordination. AI needs to “see” all the context and have the ability to take actions (even if read-only) to be really useful.

My prepaid agent queries my ERP GL directly via the MCP. No CSV exports/imports.

The skill I created reads the config file to know which prepaid accounts to pull. The agent calls the ERP connector and pulls current period activity for those GL accounts.

If you don’t have an MCP connector to your ERP, that’s the first thing to fix.

Step 4: MCP Connector to the Shared Drive

The workpapers don’t live in your ERP. They live in GDrive, Box, SharePoint, etc.

The skill needs to:

• Read last month’s workpaper

• Write the new workpaper to a new file

• Enable the AI agent to create and edit the new prepaid workpaper

An MCP connector for the shared drive handles this.

Step 5: Slack Is the Trigger and the UI

Slack is the UI for my AI agent.

Slash command: I kick off the recon by typing “/prepaid April”. Or I just chat in Claude “I want to kick off the prepaid recon for the April” and it will find it.

The agent will find the prior workpaper (location is defined in the config file), copy that workpaper to the current month-end close folder. It will then use that as the starting point for adding new prepaids, adding a new month of amortization, etc. The agent is trained just like a human would be…it brings over Excel formulas just like the prior month so the amortization calculates correctly.

When the agent finishes, it posts a digest with the workpaper, the JE CSV, and a tag for the preparer.

Below is an AI generated Slack example of what I see (done with AI so you don’t see my actual Slack…)

If the AI agent flags something I need to fix, then I’ll do one of two things:

1. Open up the workpaper and fix the errors/flags (e.g. add the Cloudflare term from the example above) and then I add a ✅ to the thread which signals to the agent that I am done so it re-reads the files to see if I fixed the flags. It updates the files and posts a follow-up message for the reviewer

2. If the issue will be repeating, then I will update the skill or config files to automatically fix it going forward

Step 6: The Human Part

The skill doesn’t auto-post the JE. We have a human reviewer who then uploads the JE to the ERP from the file the agent created.

We don’t let agents post journals to the GL. At least not yet…

In Summary…

What previously took lots of manual steps (copy workpaper to new location, open up ERP, download detail, update workpaper, create JE upload file, etc) is now 95% completed by our AI agent.

From ~2 hours of human work to ~5 minutes.

It did take longer the first time I ran the reconciliation, but then the team took it and kept iterating and improving the skill and the config to automate more and more. Eventually, you get some major time savings.

And then repeat that across your other recons and you will shave off days from your month-end close…Smaller accounting teams, faster close, less errors, etc

There is a lot more detail behind what we are doing with AI agents in accounting/finance…Subscribe👇 to follow along.

Subscribe now

Footnotes:

• Have a cool AI use case in finance/accounting? I would love to feature it. Email me at onlycfo@onlycfo.io

Share