How To Teach Your AI Agent About Your Company

Direct Source Verification: This story is aggregated from Forbes (forbes.com). Full reporting rights and copyright belong to the primary publisher.
Frederik Riskær Pedersen, Founder and CEO of Maplith, a company brain for your AI agents.

Frederik Riskær Pedersen, Founder and CEO of Maplith, a company brain for your AI agents.

gettyFor 16 years, I ran a translation company that grew from a handful of people into an international business. The most important thing I learned about knowledge is that it rarely lives where you think it does.

One customer shows it well. We had to fetch their files from their own system and deliver them in their own layout, and we could only invoice some of the repetitions. None of that was written down anywhere. It lived with one project manager. Anyone else would have done it the standard way and got it wrong.

That tailoring was not a side issue. It was the product. If the work had been standard, the customer could have run the model themselves. This is why a company’s rules never get standardized away: The exceptions are often exactly what customers pay for.

I think of this every time someone asks me why their AI agent isn’t delivering. The models are good enough. The problem is that the agent starts from zero every time. It does not know which systems you run, how they connect or which rules you follow without ever writing them down. So it guesses, and a confident guess in your accounts is worse than no answer at all.

Here are five things you can do this month before you let an agent work inside your business. None of them needs a new tool.

Book two hours with the three or four people who make most of the everyday decisions. Ask them one question: “What would a new hire ask you in their first week?” The answers are your unwritten rules: which customers get special terms, when a refund is approved, what can be invoiced and what cannot. Write each one as a short sentence, specific enough to check: “This customer gets files back in their own template.” “Refunds above 300 euros are approved by the finance lead.”

List every system you use and what each one is the source of truth for. You will find places where two systems describe the same thing differently. “Returned” in the ERP can mean the customer has sent the item back, while in the warehouse it means the item is back on the shelf.

The instinct is to make the systems agree. In my experience, that rarely works because each system was set up for the team that uses it, and that is why it works for them. Write the translation down instead. The same goes for numbers. In my company, the price did not sit in the same place from one customer to the next. Note where it lives for whom and which value to trust for which question.

Most useful answers need two systems joined: the order in the web shop and the invoice in accounting, or the support ticket and the customer in the CRM. Find the field that ties them together and test it by hand. Take 20 random orders and look for them in the other system. If you find 19, you can build on that link. If you find 12, you cannot, and an agent will rarely tell you it is guessing. The test takes an afternoon and saves months of wrong reports.

Split actions into four levels: read, propose, act after approval and act alone. Start with everything that changes something requiring approval from a named person. Set limits in money and volume, not in trust. When the agent has proposed the same action a hundred times without a mistake, move that one action up a level. That decision should be yours, never the agent’s.

For every action, record who asked for it, who approved it, what happened and when. It sounds bureaucratic, but it is the only way to answer when a customer, an auditor or your board asks, “Why did this happen?” Without a log, you also cannot tell when it is safe to give the agent more freedom.

None of this is about artificial intelligence. It is about making knowledge that lives in people’s heads, and in systems that do not talk to each other, available and trustworthy. Companies that do this get more out of their AI. They also become less fragile when a key person goes on holiday or leaves.

So start small. Book the two hours this week, write down 10 rules and test one link.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

Original Source
https://www.forbes.com/councils/forbestechcouncil/2026/11/20/how-to-teach-your-ai-agent-about-your-company/
Visit Forbes ↗
SHARE STORY:
𝕏 f in

Related Coverage in Technology