# Your AI agents are guessing about your business

> Your agents can read your data, but they do not know how your business runs, so they guess at your numbers and never say so. Why it happens, and a ten-minute test to see whether yours are guessing.

By Brandon Gross, Co-founder. Published November 2, 2026.  
Tags: ai agents, ai context, ontology.  
Canonical: https://kenseme.com/blog/your-ai-agents-are-guessing/

![Two panels. On the left, your best new hire, wearing a Day 1 badge, asks what the refund rule is before approving one. On the right, an AI agent approves a $740 refund, and a note asks whether it was over the $500 limit: it did not check.](/blog/media/your-ai-agents-are-guessing/brilliant-new-hire.svg "Your best new hire asks how you do it here. Your AI agent just does it.")

AI agents can help you scale your business, and your own role with it. But they share one problem. They do not know how your business operates, and they do not tell you when they are guessing. They just guess, confidently. This post shows what that looks like, explains why it happens, and ends with a ten-minute test that shows whether your agents are guessing too.

## A brilliant new hire, every day

Picture a new hire far smarter than any you will ever hire. They speak nearly every language. They can pass the bar exam, the CPA exam and most professional certifications you could name. They know more, across more fields, than almost anyone on earth. They never sleep, never have a bad day and never tire of your questions. That is an AI agent.

Now give it a real task. A customer asks for a $740 refund, and the agent approves it in seconds, polite and sure of itself. It never asked whether refunds over $500 need a manager. It did not know your rule, and it did not say it did not know.

Your best human hire would have asked. On day one they were sharp too, and they still could not tell you who counts as an active customer or how refunds work here. So you onboarded them: a glossary, an org chart, the policies that matter, a walk through how work really moves. What they knew about your business grew every week. A few weeks in they answered correctly, because they knew how your business runs. When they were unsure, they asked.

The agent has the same problem your best hire had on day one: context. Everything it knows came from public information, not your business processes and definitions. Your pricing rules, your fraud signals and the way you segment customers were never in it. That knowledge is yours by definition. The agent knows only the part of your business you are working on with it right now, and only what you have taught it along the way. Walk it through your refund rules today and it can apply them today. Ask it about pricing tomorrow, or turn to the agent in the next department, and it starts from nothing. Some agents keep a little memory between conversations, a preference here, a fact there, but it never holds a complete picture of your business, or of how it ought to reason about it. Whatever it learns, it learns in tiny steps.

![Two charts of how much each one knows about your business over time. A new hire climbs steadily from day one to week three and asks when unsure. An AI agent climbs a little in each conversation, falls back to almost nothing when the next one starts, and guesses when unsure.](/blog/media/your-ai-agents-are-guessing/new-hire-vs-agent.svg "A new hire builds on what they learned. An agent keeps little from one conversation to the next.")

## So it guesses

When an agent does not know what you mean, it does not stop and ask. It guesses. That is the nature of the technology: it predicts the most likely answer from the statistics of everything it has read, not from how your business actually works. It guesses at your numbers, and it guesses at how your business operates: which customers qualify, which orders need approval, which exceptions go to a person.

Ask an agent for last quarter's profit and you will get a number. Gross or net? Before or after the one-time charges finance always takes out? Which business units count? Each is a reasonable reading. Nobody told the agent which one you meant, so it picked the one it had seen most often. The number looks right. It may be wrong, and nothing in the answer says so. A few answers like that and people stop trusting the agent, and the scale never arrives.

![The question last quarter's profit has four reasonable meanings: gross profit, net profit, profit before one-time charges, and core business units only. The agent picks gross profit and returns 4.2 million dollars, with no sign that it guessed.](/blog/media/your-ai-agents-are-guessing/one-question-four-meanings.svg "One question, four reasonable meanings. The agent picks one and does not say which.")

Profit is not special. Every industry runs on numbers like it, and on rules for how work gets done.

- **Retail.** Does "same-store sales" include the store that closed for a month? Can this order ship from a store, or only from the warehouse?
- **Healthcare.** Is a readmission within 30 days or 90? Which patients need a follow-up call before they go home?
- **Banking.** Is a loan delinquent at 30 days past due or 90? When does a flagged payment go to a person instead of being declined?
- **Insurance.** Does "claim date" mean the loss, the report or the payment? Which claims can be paid without an adjuster?
- **Manufacturing.** Is "on time" measured against the date the customer asked for or the date you promised? When can the line swap in a substitute part?
- **Software.** Does annual recurring revenue include setup fees? Which customer issues go straight to an engineer?

In every case the agent answers anyway, with a clean number or a confident action, and no sign that it guessed.

A smarter model does not fix this. The model already knows more than anyone you could hire. What it lacks is your business, described in a form it reads every time it starts work. That briefing is an ontology.

## What goes in the briefing

An ontology is a description of your business: the things in it, how they relate to each other, and the rules they follow.

1. **Things.** The nouns your business runs on: customer, product, store, supplier, order.
2. **Relationships.** How those things relate to each other: a customer places an order, a supplier ships a product, an order is picked at a store.
3. **Rules.** What must be true: every order has exactly one customer, refunds over $500 need a manager, a product needs a price before it can be sold.

Each definition carries the meaning your people already agree on. "Profit" says which profit, and how finance calculates it. "Active customer" says active within what window. A simple test for whether something belongs: if you can say it in a short sentence, it can go in the ontology.

It is the same onboarding you give a person. The glossary becomes definitions. The org chart and the tour become relationships. The policies become rules.

## More than definitions

Definitions stop the guessing. The relationships and rules do something more. Together they give the agent a model of how your business works, and an agent can reason against a model. It can follow the relationships from one fact to the next and check each step against your rules.

That lets it answer questions nobody wrote down in advance. Which of our largest customers have orders waiting on a supplier who shipped late last month? Which refunds this week skipped the manager approval they needed? No one defined those questions ahead of time. The agent can answer them because it knows what a customer, an order, a supplier and a refund are, how they relate and what must be true.

That is the power of an ontology. A glossary answers the questions you expected. A model of your business answers the ones you never thought to ask.

## Built once, kept current

Your team does not have to write the briefing from scratch. Most of it already exists, in your policies, process documents, glossaries and data dictionaries, and in the heads of the people who do the work. The job is to draw it out and get your experts to agree on what it means.

You onboard every new person from scratch. A briefing is built once and shared. Every agent that reads it starts from the same understanding, so the sales agent and the finance agent mean the same thing by "customer." Each time a definition is added or corrected, every agent that reads the briefing knows it the next time it starts work, so your agents get smarter as the briefing grows instead of starting over.

![One briefing, drawn from your documents, your data and your experts, and kept current. It holds things, such as customer, product and order; relationships, such as a customer places an order; and rules, such as refunds over 500 dollars need a manager. A sales agent, a finance agent and a service agent all read the same briefing.](/blog/media/your-ai-agents-are-guessing/one-briefing-every-agent.svg "One briefing, built once and kept current, read by every agent.")

It also has to stay true. A new product line arrives, a threshold moves, finance changes how it reports. The briefing is only useful if someone owns it and updates it when the business changes, the way a good team keeps its handbook current.

This is the work Kenseme does. It reads what you already have, from process documents and glossaries to data dictionaries, industry standards and the structure of your data, and fills the gaps through structured interviews with your experts. Kenseme drafts the definitions and the relationships between them, your experts settle what they mean, and Kenseme binds the result to your data, so your agents stop guessing at what your business means and can reason across it to answer questions no one thought to ask.

## What to do next

Pick one question you would most like an agent to answer well, ideally one where the right answer could grow revenue, cut costs or open a new market. Ask three people on your team to write down, separately, what the key terms in it mean and how that work gets done. Then compare. Wherever their answers differ, your agents are guessing. Those differences are your starting point: the knowledge to capture first.

If you want to see this on your own data, together we pick one business area, the one where a right answer is worth the most to you: more revenue, lower costs or a new market. Two weeks later you can ask a real question and check the answer. [Book a demo](/demo.html).