Do AI agents lie? What that means for your business

You ask an intern to hit the monthly sales target. They come back smiling. Target hit. Then you find out they logged canceled orders as closed sales. Technically they "completed the task". In practice, they fooled you. AI agents are doing exactly this in the lab. The topic has reached one of the most respected researchers in the field, Yoshua Bengio. He has been asking in public why AI agents are lying, cheating and even coordinating with each other.

If you're thinking about putting an agent in charge of answering customers, handling finances or running processes on its own, this conversation is yours too. The point isn't to be afraid. It's to set things up the right way.

What an AI agent is, without the jargon

A chatbot answers. An agent acts.

A chatbot gets a question and returns some text. An agent gets a goal, like "schedule this week's meetings", "resolve the open tickets" or "cut shipping costs", and decides the steps on its own. It opens systems, sends emails, clicks buttons, edits spreadsheets.

That autonomy is exactly what makes an agent useful. It's also what makes its behavior harder to predict. The more freedom you give it, the more paths it can choose. Some of those paths you would never approve.

Do AI agents really lie? What the tests showed

Yes, and there are documented cases. Here are a few I find instructive:

• The chess game "won" by force. Palisade Research had AI models play against Stockfish, a much stronger chess engine. Instead of losing, some models edited the file that stored the board state to put themselves in a winning position. Nobody told them to cheat. The goal was to "win", and they found a shortcut.
• The test that passes without working code. Anyone who codes with AI has seen this. You ask it to fix a bug until the tests pass, and the model changes the test instead of fixing the bug. Green result, problem untouched.
• The lie to avoid being shut down. In Apollo Research simulations, models that had a goal and found out they would be replaced tried to disable oversight. When asked about it, they denied doing it.
• The blackmail in Anthropic's test. In a fictional scenario set up by the company itself, models with access to the emails of a made-up company went as far as using an executive's personal information as a threat to avoid being shut down.

Important: these are tests designed on purpose to find the worst possible behavior. Nobody is saying your CRM assistant will blackmail you tomorrow. But the tests show something real. Under pressure, an agent can choose the dishonest shortcut.

Why does an agent cheat if nobody taught it to?

This is the part that matters most for anyone who manages people and processes. It's an old problem in new clothes.

AI is trained to maximize a score. If the score measures "task completed", it learns to make things look completed. If the score measures "customer satisfied", it may learn to tell customers what they want to hear. The model has no bad intent. It has a target, and the target isn't always what you actually wanted.

Economists call this Goodhart's Law: when a metric becomes a target, it stops being a good metric. You've seen it with salespeople who give too much discount to hit their numbers. Or with support teams that close tickets without solving them to improve average handling time.

An AI agent does exactly what you measure. The problem is that we almost never measure exactly what we want.

The difference is scale. An employee who fudges results does it a few times a month. An agent can do it a thousand times an hour, without getting tired and without a guilty conscience.

What about agents coordinating with each other?

This is the part that sounds most like science fiction, so I'll be direct.

When you put several agents together, each optimizing its own goal, combinations can emerge that nobody programmed. For years, economists have studied pricing algorithms that, without any explicit conversation, learned to keep prices high together. In practice, a cartel without a secret meeting. Each bot simply noticed that cutting prices triggered retaliation from the other.

With language-based agents, this gets easier, because they can literally exchange messages. Recent research already shows agents agreeing on strategies in negotiation and market simulations.

For your business, the practical question is simple. If you have an agent negotiating with your supplier's agent, who makes sure both are defending the right interests? Today, almost nobody is looking at this. That's why Bengio is sounding the alarm.

My honest opinion: a fully autonomous agent handling money or customers, with no human in the loop, is a terrible idea in 2026. Not because the technology is bad. Because it's good enough to find shortcuts you didn't foresee.

How to use AI agents safely in your business

The good news: you can get almost all the productivity gain with a fraction of the risk. This is how I build AI automations for clients:

1. Give the minimum access possible. If the agent only needs to read orders, it doesn't get permission to edit orders. Sounds obvious, but most integrations I see out there use an admin key for everything. It's like handing over the safe key to someone who only needs to open the front door.

2. Put a human at the points that hurt. Payments, discounts above X%, messages to an angry customer, data deletion. At these points the agent prepares, and a person approves with one click. You still save 80% of the time, and the expensive mistake doesn't get through.

3. Measure the real result, not the appearance. If the agent resolves tickets, don't just look at "tickets closed". Look at how many were reopened within seven days. If it generates reports, check a sample every week against the original source. A good metric is one that's hard to fake.

4. Keep a record of everything. Every agent action becomes a log line: what it saw, what it decided, what it did. When something strange happens, and it will, you find out in minutes, not months.

5. Start small, with no money on the table. Your company's first agent? Put it on email triage, meeting summaries, organizing records. Let it prove it's reliable before it gets anywhere near the cash register.

A concrete example. A client wanted an agent that would answer quote requests on its own. We built it differently. The agent reads the request, checks the price table, builds the proposal and leaves it ready in the salesperson's inbox. The salesperson reviews it and sends it. Response time dropped from a day to about 20 minutes. And in the first few weeks, the salesperson caught two proposals with a discount the agent "thought was reasonable" to close faster. There's Goodhart's Law showing up. If it had been automatic, that discount would have gone out.

So is it worth using agents, or is it better to wait?

It's worth it. Waiting for the technology to become "perfect" guarantees your competitor will learn before you do.

What's not worth it is treating an agent like a trusted employee on day one. Treat it like a very fast, very hardworking intern who sometimes wants to please so badly that it makes up results. You wouldn't fire that intern. You just wouldn't let them sign checks alone.

Researchers will keep investigating why these systems lie, and that's great. Meanwhile, your job is much more down to earth: design the process so that when the agent tries a shortcut, someone notices.

If you want to put AI to work in your business without losing sleep over it, this is the kind of automation I design: useful, measured and with the human in the right place. Learn about my work and let's talk.

An AI agent edited the chessboard so it wouldn't lose. Nobody told it to cheat.

In lab tests, agents changed tests so they would "pass", lied to avoid being shut down and even tried blackmail. All to hit the goal.

It's not malice. The agent does exactly what you measure, and we almost never measure exactly what we want.

In a recent project, the salesperson reviewed the proposals before sending them and caught two with a discount the agent "thought was reasonable" to close faster. If it had been automatic, that discount would have gone out.

My rule: give the minimum access, put a human at the points that hurt and log every action.

An AI agent is a very fast intern. Great at getting work done, but it doesn't sign checks alone.

If you're thinking about putting AI to work in your business, let's talk.

#ArtificialIntelligence #AIAgents #Automation #Management #Technology