The recipe this CFO runs his practice on
|
|
This is Eduardo Guevara. He is a fractional CFO based in Paris. Earlier this month, in our AIFC Connect session, he walked my community through the system he runs his practice on.
And the first rule of that system surprised a lot of people.
He does not connect AI to his inbox.
Most people do the opposite. They discover connectors, plug AI straight into Outlook or Gmail, and super quickly it can read everything. It feels like the obvious first step, because e-mail is where everything is.
But Eduardo did something different. And when I looked at what Eduardo built, I didn't see an IT setup. I saw a kitchen.
I love to cook (especially BBQ steak and seafood). And every good cook knows an amazing dish is about a lot more than what pan you use.
|
|
You pick good ingredients. You get help with the chopping. Then the chef cooks.
Shop. Prep. Cook.
Eduardo does all three with AI.
- Shop – He’s selective about the data he gives to AI (ensuring it is high-quality)
- Prep – He has a ‘sub-agent’ that does routine work at a lower cost (more on this later)
- Cook – He has a master AI that helps with the finished output
Using this method, you can avoid the 3 mistakes of:
- Giving AI too much data.
- Consuming a lot of usage.
- Producing inaccurate outputs.
And you can use it for a lot more than just e-mails… (more finance examples at the end).
Step 1: Shop – Choose what comes into the kitchen
|
|
Here is the problem with giving AI everything.
It will rarely come back and say "I'm not sure about this, the pricing document you gave me is from March."
It will just answer – and this is what you need to avoid.
These models were trained to give you something. As I explain in my AI clone masterclasses, they were rewarded for answering. So even when there is no correct answer, they produce one anyway. And it always sounds confident.
It's a bit like walking into a restaurant and saying "give me something I like". The server doesn't know if you like meat or fish. So they bring you something. It looks good, but it’s not what you wanted.
So you have to decide what goes into the kitchen.
Now you could give AI access to everything. Email, meeting notetakers, shared files, task apps, ERP etc.
This is easy to do. But the problem is you have a lot of e-mails, a lot of meetings and a lot of files. So, giving AI access to all of it is how you get higher costs and worse answers.
Eduardo doesn’t give AI access to everything.
Instead, he has a second e-mail address, and he forwards only the relevant e-mails to it. He then connects that e-mail via MCP to his AI tool.
Yes, this adds another step in his process. But it is worth it in exchange for the clean data (he gets a lot of rubbish e-mails.
There is a second reason why he does this. It’s called ‘Prompt Injection’
If an AI accidentally mistakes and e-mail for an instruction. Your outputs could be wildly inaccurate.
So don't send it access to the whole market. Give it a shopping list: Specific mailboxes. Specific folders. One project app. One data source.
Pro tip – You don't have to go as far as Eduardo. You can build the filter into your prompt. "When searching my e-mails, only use the ones labelled [label]." Or "When using Drive, only use [this folder] and ignore everything else." It won't be as safe as a separate inbox. But it's much better than letting AI burn through your usage because it has to guess all the time.
Now careful, because I disagree with the extreme version of this.
A lot of people tell me "my data isn't good, so I cannot start". That is wrong. Use AI to clean and structure your data.
Messy is fine. Inaccurate is not.
Step 2: Prep – Give the chopping to a smaller AI
|
|
No restaurant has the head chef peeling potatoes all day.
With AI, your kitchen helper is a sub-agent.
Say you have a 90-minute call transcript and there are 12,000 words in it.
Instead of giving all 12,000 words to your most powerful model, you give it to a smaller one first, and you say "pull out only the decisions, the numbers and the owners".
It comes back with 400 words.
The powerful model never sees the other 11,600. So it's cheaper. And more importantly, none of the rubbish is taking up its "brain space" (the context window) in the chat.
Same thing on finance work. 50 invoices become 5. 200 GL codes become 20. Every variance above 50k, with the owner next to it.
You can do this today with two chat windows and copy-paste. Claude Sonnet in one, pasted into Claude Opus in the other. It doesn't need to be clever.
Advanced Tip – If you use Zapier, Make, n8n or Power Automate, you can chain it without the copy-paste. The first step (or node) is the small model. The second step is the big one.
This is also where Claude Cowork and ChatGPT Work are really good. Most people use them when a job feels big and complicated. But, sometimes I would do the opposite.
Cowork in Claude (or Work in ChatGPT) is strongest on the repeatable bit, where the steps are obvious and there are a lot of them. Files. Folders. E-mails. Line matching. It works on your computer, with connectors into things like Google Drive.
So it makes a very good kitchen helper, to carry out repeat steps that you can then feed into a more intelligent model.
Step 3: Cook – Only now bring in the expensive model
|
|
This is exactly how I teach my community. I use Claude's Sonnet model most of the time. But, when I need deep thinking, where it has to take time to reason, I switch to Opus. Same thing in ChatGPT, with Terra for everyday work and Sol for the hard thinking.
Where a lot of finance pros go wrong is they open the most powerful model first, put everything into it, and ask it to sort out the mess and do the analysis in one go.
This is like asking your head chef to do the shopping and the chopping too!
In France we call a prepared station 'mise en place' – This is where everything washed, chopped, in little bowls, before service starts.
That's what your tasks should look like before the expensive models see it.
This last step is where you use your human brain, and the more powerful AI brain to produce higher quality work and make better (more accurate) decisions.
3 finance recipes to try this week
1. Invoice and bank reconciliation
I tried this in Claude Cowork one weekend with about 200 invoices and my bank statements. It did not get everything. But even at 70-80%, having the first cleanup is really useful.
- Shop: the bank export and your AP folder.
- Prep: Sonnet matches what it can and flags what it cannot.
- Cook: work with Opus on the 20-30% it could not match, and why.
2. Variance analysis
- Shop: one GL export, from one period, in one entity.
- Prep: every variance above your threshold, with the account, the number and the movement.
- Cook: commentary on the few that matter for decisions.
3. Revenue recognition on contracts
I showed this in an office hours session with a trial balance and a set of revenue contracts. Change the completion percentage, and you see the impact on revenue and the P&L straight away.
- Shop: the signed contracts and the TB.
- Prep: the dates, the amounts and the completion percentage per contract, in one table.
- Cook: work through the recognition judgement with the stronger model.
The One Thing to Remember
AI is never going to second guess your data (unless there’s an obvious mistake).
AI is never going to tell you “this will be an expensive and inaccurate way to run this task”.
It is going to find a way to answer you – confidently – and it doesn’t care about how much consumption it uses to give you an answer that could be wrong.
But now you have a recipe for higher-quality outputs.
All you need to do is cook 😉
Best,
Your AI Finance Expert,
– Nicolas
P.S. The best way to learn this from me directly is my weekly masterclass. 60 minutes plus Q&A, where I answer your AI in finance questions. Join me here.
P.P.S. For more ways to save your AI consumption and keep costs down, here is my YouTube Video – How To Save 97% of Your Claude Tokens As a CFO