This is the 4th edition of my bots and agents series. So far we’ve built a dashboard and reporting agent, improved AI performance with better context and looked at the best ways to build AI skills and plugins. Today we’re looking at the 5 things you need to keep your AI bots and agents safe with your numbers.
How do you make AI not hallucinate? This is how Brandon did it:
Two weeks ago, in our Mastery office hours, Brandon (a member of my AI Finance Club) walked us through his variance memo skill.
Some variances crossed his threshold. But the memo didn't explain them.
Most AI will generate something wrong to fill the gaps.
But Brandon's variance memo skill doesn't.
It writes the draft, flags the item, and then tells the user "you need to check on this before you publish the memo."
It's exactly safeguards like this that will protect your numbers, protect your reputation, and make sure you are confident in deploying AI.
So today, I have 5 things you can do, including safeguards like Brandon's to make sure AI does not harm your numbers.
Reminder: 5 Capabilities
If you've followed this series, you'll already have a basic understanding of all of these.
|
|
- Connectors link your AI to the other tools you use every day: Google Drive, OneDrive, email, even NetSuite or Xero.
- AI in Microsoft Office or Google Workspace where you add AI as an add-in, so Claude, ChatGPT, Copilot or Gemini work inside the file you have open.
- Projects are for persistent context (knowledge and files you upload to the project) that your AI will read in every chat that uses that project.
- Skills are saved instructions where you teach a task once and AI runs the same steps every time. As we’ve discussed in previous newsletters, you can chain skills together into ‘super-skills’ or build them into plugins like we did in my last newsletter.
- Scheduled tasks where a workflow is carried out automatically. More information below:
So how do we make sure we’re keeping our numbers safe when using all of this different capability which has the ability to generate different results every time?
Here're are 5 things that relate to all of the capability above.
5 things to do to keep your numbers safe
1. Tell AI what to do when data is missing (Connectors)
|
|
Connect Google Drive (or OneDrive etc)
Then write the prompt. Say who you are, where the data is, what you want, and that file names will change. Here's mine.
"Flag missing information" is the important line. When I did this, Claude came back with 11 flags across 56 P&L files. One of them: a country's personnel cost was exactly the same in February and March, to the euro. Claude flagged it as probably copied forward from the month before.
|
|
One line, in one file out of 56. Would you have caught that at 7 a.m.?
For a scheduled run I’d take it one step further. And add this to the end of the prompt (more on scheduling at the end).
Tip – If you’re using Claude: Open Settings, then Connectors, then Google Drive, and scroll to Tool permissions. Each type of action has its own setting: reading files, editing files, deleting files. Leave reading on "Always allow". Set editing and deleting to "Always Ask", so Claude asks you every time.
2. Ask for formulas and an audit tab (Microsoft Office & Google Workspace)
|
|
In Excel, click Add-ins, search for Claude (or ChatGPT), install it and sign in.
Open your file and open a new AI chat. My example is a master revenue file with one tab per month.
I ask for formulas on purpose. A hard-coded number is a number you have to trust. A formula is a number you can check. And the audit tab means I don't have to go through everything line by line.
|
|
Then read the audit tab, and read what your AI says it didn't check. Mine said all checks passed. But further down, Claude admitted it had only checked the totals, not every number inside the tables.
And one result looked odd. Customer retention in the first month should start at 100%. Mine started at 92%. Claude explained why: two customers had no revenue that month. That's something you can check in your own file in two minutes.
Tip – If you’re using Claude – Once the Excel file passes your checks, turn on "Let Claude work across files" in the Claude settings, in both Excel and PowerPoint. This lets Claude in PowerPoint read the Excel file you have open. Then ask it to build your board deck from that file.
3. Read the skill before you let it run (Skills)
A skill runs the same steps every time. That's great when the steps are right. When they're wrong, it repeats the mistake every month.
My example is branding. In a new chat, ask AI to create a branding skill for all documents, presentations and visual outputs. Attach your brand guidelines, or a deck that follows them. Then click Save skill.
|
|
Before you trust it, open it and read it. Mine listed the exact color codes and fonts it would use. Hold that list next to your real brand guide and check they match.
Claude also told me it had changed something I never asked for: it fixed the layout so the dashboard fits on a phone screen.
Small change, harmless this time. But read notes like that. They tell you what AI changed on its own.
Tip – For any skill that touches numbers, write a rule like this inside the skill itself: "If something is missing or doesn't tie, stop and flag it. Never guess." Like this, it's in every run, not only in the prompts you remember to write.
4. Tie the numbers back to the source, then run it twice (Projects)
|
|
One of the first projects I train my members to build is called the "AI CFO Clone"
It has three files: your LinkedIn profile as a PDF, a company snapshot, and your writing style guide.
But you can also connect that project with the Drive folder with your financials, plus your e-mails.
This way you can ask for a daily CFO dashboard with the top actions, plus a scan of the last 24 hours of emails. Written in a way that’s relevant to your role, your company, and the way you speak.
|
|
You can then turn it into a skill. Just ask ChatGPT or Claude to “Help me build a skill for this” that runs whenever you ask "How is my day?".
Now the check, in two parts.
- Tie it back. Pick three numbers from the dashboard, for example this month's revenue, the budget and closing cash. Open the source file and find each one. If you can't find it, or it's different, the dashboard isn't ready.
- Ask the same question twice. On the same day, using the same files, and the same skill. The answers should match. Mine didn't. The first time, it said five things needed my attention today. The second time, four. Nothing had changed in between. If it can't count the same thing twice, it isn't ready to run alone.
If either check fails, it's not ready to be used and you need to improve it.
5. Decide on purpose what AI can do without asking (Scheduled tasks)
Now, and only now, you schedule it.
|
|
Click Scheduled in the sidebar, then New task. Choose to set it up yourself rather than describing it to Claude. Name it "CFO Brief" and write this in the instructions.
Connect your AI CFO Clone project / skill. Set Frequency to Daily and the time to 07:00, then Save.
Then do two checks before you let it run without you.
- Check what it's allowed to do without you. When you set up a scheduled task, there's a setting for approvals. Mine was on "Automatically approve". That means Claude uses your connected apps without stopping to ask you first. That's fine if the task only reads files. It's not fine if it can send an email or delete a file. If yours can, switch it to ask first.
- Make sure the task names exactly what to run. My task said "build the same dashboard", but it never said which skill to use. So every morning, Claude has to guess. Don't make it guess. Add one line: "Run the skill cfo-day-overview. Put anything flagged at the top, under 'Check before you use this'."
That last part is Brandon's idea again, now built into the job that runs when you’re not there.
And don't stop at your own dashboard. The same five checks work for the month-end memo, the weekly cash report and the board pack.
Anything you'd like to find waiting for you when you log in on a Monday.
The One Thing to Remember
The best bots & agents are self-checking. But they only work if you check them yourself first 😉
Add your audit tabs, and your check rules into your prompts.
Every agent that runs automatically needs a way to alert you to what it’s not confident about.
It’s the only way for you to build confidence in AI, and keep your numbers safe.
Best,
Your AI Finance Expert,
-Nicolas
P.S. The best way to learn from me directly is to join my weekly masterclasses. 60 mins + Q&A where I'll get all your AI in Finance questions answered. Join me here.
P.P.S. Want to see the folder setup and AI agent in action? Watch 5 Skills to Become an AI CFO in 2026