You can't stop AI agents hallucinating — not completely, and anyone who tells you otherwise is selling something. What you can do is set an AI agent up so that when it does make something up, the mistake is caught before it reaches a customer, and corrected so it happens less next time. That turns hallucination from a reason not to use agents into something more like a known staffing problem: real, manageable, and mostly handled by how you set the job up.
We run agent teams every day — on our own business, and a full demonstration team we use with clients. What follows is what actually works.
Why do AI agents make things up?
AI agents make things up because they're built to produce plausible answers — and when they don't know something, plausible and true drift apart. Ask an agent about a customer it has no information on, and it will happily give you a fluent, confident, wrong answer — unless the way it's set up makes "I don't know" the easier path.
It helps to think of it like a keen new hire who would rather guess than admit a gap. You wouldn't fix that person with a stern talking-to. You'd fix it with clear sources, a review step, and a culture where saying "I couldn't find that" is fine. Agents are the same, except the fixes are written down and enforced by the system rather than by habit.
In practice, the most common failure we see is an agent answering from the wrong real information — something true that belongs to a different context. Early in one of our builds, a sales agent answered a task using another company's playbook that happened to be within its reach. Every word was based on something real; all of it was wrong for the job.
Keep agents working from your data, and only yours
The first fix is about sources. An agent should work from your files, your CRM, your tracker — never from its general knowledge of the world. Ours read the company's own documents before every task, and their job descriptions say, in effect: if the information isn't in these sources, you don't know it. Say so.
The second half of that fix is a boundary. After the wrong-playbook incident, every agent got a named list of places it may read from — and a standing instruction to refuse any task that seems to need something outside them. An agent that answers "I can't see that information" is doing its job well. You want that answer to be normal, not embarrassing.
Make every claim checkable
Our outreach rules require that every claim about a prospect trace to something a person could verify. A draft can say "your bakery's counter is usually full at closing" only if there's a source behind it. And each draft carries its evidence with it — a short note at the bottom listing what was observed and where — so whoever approves it can see the words and the basis side by side.
The same rule catches a subtler kind of making-things-up: an agent flattering its own progress. One of ours once reported a batch of work as newly finished when most of it already existed from an earlier run. Because every run writes a record of what it actually produced, the claim fell apart the same day, and the fix was a permanent rule: reports must separate "done this run" from "already existed". Keeping the quality bar high is largely this — replacing trust with the ability to check.
Keep a human between agents and the outside world
However careful the setup, the last line of defence is simple: nothing customer-facing sends without a person approving it. Our agents can't send email, publish posts or spend money — they don't have the credentials. Finished work waits in a queue with an approval label until a human clears it.
That one design choice changes what a hallucination costs. The worst case stops being "a customer received nonsense" and becomes "you declined a draft" — thirty seconds of your day. And every decline teaches the team something: the patterns behind your no's get written into the rules the agents read, so the next batch of drafts starts from a higher floor.
The risks that remain, and what they cost
These layers make hallucinations rare and cheap. They don't make them impossible, and it matters to say so. Internal work that isn't gated — a research note, a draft analysis — can still contain a wrong inference, which is why the records and spot checks exist. And the gates cost speed: a human approval step is a queue, and agents on tight evidence rules are slower and more conservative than the flashy demos you may have seen. We think that trade is the right one for a real business. It's the same trade you already make with people — freedom inside the building, sign-off at the door.
If hallucinations are the thing that's kept you from trying agents, that's a good instinct — and a solvable problem. Book a call and we'll show you a working agent team, evidence rules and all, or read how an agent team actually works first.