AI agents that analyze company data are now good enough to write the query, run it and suggest an action. Yet fewer than half of what they find leads to anything that helps the business. Cheap analysis has not made good analysis cheap, because the two things that make a finding useful are still supplied by people.
The AI cannot guess what your words mean
Picture a shop owner who connects an AI assistant to the till records and asks which products are making money. The records have a column called "amount", another called "cost" and a third called "discount". Whether "making money" means before or after staff time, returns, card fees and the free gifts that go with some orders is something only the owner knows. The assistant picks one meaning and gives a confident ranking. The owner has no reason to doubt it, so the owner acts on it.
That is why lack of business context is the most common reason agent findings disappoint: 55% of executives name it. It also explains a test in which seven different AI models answered the same 410 business questions. Swapping in a better model improved accuracy by up to 18 points, while giving the models the company's definitions improved it by 20 to 31. Paying more for a smarter model is the expensive fix, and writing down what your terms mean is the cheap one.
Checking is the other shortage
Suppose a retailer's agent flags 40 unusual stock levels every morning. At five minutes each, checking them takes more than three hours, so the person ends up checking a few and trusting or ignoring the rest. That is roughly where most companies are: most firms still check their agents' work by hand, and fewer than a quarter manage to review everything their agents produce.
The more an agent can do without being asked, the more it matters that its findings arrive ranked, with the definition it used stated at the top. A finding that says "late means more than two days after the promised date" can be checked in ten seconds. A finding that does not say what it means by late has to be rebuilt from scratch before anyone can trust it.
The argument comes before the AI
The fix in the 1990s was not only software. Someone had to decide what "revenue" meant, and that is a conversation between finance and sales, not a technical task. Agents force the same conversation. A company cannot hand work to an AI until it has decided what its words mean.
The few companies where more than 70% of agent findings lead to action struggle less than the rest with giving their agents business context. As more companies use agents, the person who writes the definitions will have more say over what the AI recommends than the person who chooses the model.
You can start on your own scale this afternoon. Write one page that defines the five terms your reports rely on most, such as active customer, margin, late order, churn and a qualified lead. Paste it at the start of every analysis you ask an AI for, and ask it to say which definition it used in each answer. When two answers disagree, the page will show you why.