An AI agent can’t take responsibility. That’s the core of what it can’t do, and everything below is a variation on it. It can execute work brilliantly, but when things go wrong, you’re the one standing in front of the judge, not the agent. So build a human checkpoint wherever money, contracts, or customer relationships are involved.
This isn’t an anti-AI piece. Renforza places agents as the third candidate on your shortlist. That’s exactly why we want the boundaries sharp: an agent in the wrong spot costs more than no agent at all.
Who pays when the agent makes something up?
You do. Two court cases make that painfully concrete.
The first is Air Canada. The airline’s chatbot promised a passenger a retroactive bereavement fare, a discount that didn’t exist. Air Canada argued the chatbot was a separate entity responsible for its own statements. The court found that argument as creative as it was doomed, and the company had to honor the invented discount, as Forbes reported.
The second comes from Germany. The Higher Regional Court of Hamm ruled that a company is directly liable for what its chatbot says. No intermediate step, no mitigating circumstances: what your bot says, you said.
The lesson isn’t that chatbots are dangerous. The lesson is that the law doesn’t distinguish between an employee making a promise and software making a promise. With one difference: the employee usually knows when they’re unsure. The agent sounds exactly as confident when it’s wrong as when it’s right, and that’s precisely what makes it risky in customer contact.
What else can’t an agent do?
Beyond carrying responsibility, four structural limits stand out.
Recognizing an exception as an exception. An agent follows patterns. A situation that falls outside every pattern doesn’t register as an alarm. It registers as just another case, and gets processed like one. A human senses something’s off precisely because it doesn’t resemble anything.
Weighing context that isn’t in the data. That a customer had a death in the family last month. That a supplier is about to go under. That a colleague is working their last day. People factor that in automatically. An agent only knows the systems it’s plugged into.
Carrying relationships. Trust is accumulated history between people. An agent can administer a relationship at best, not hold one. A field report from Project NANDA at MIT also notes that many tools learn too little from feedback. That is not evidence about human relationships, but it is a relevant technical limitation.
Stopping when it really matters. A good employee halts the work when the instructions don’t add up, and makes a call first. An agent does what it’s told, even when “what it’s told” has become nonsense through an upstream error. Hence the rule of thumb: the more irreversible the action, the closer the human should sit.
Where does the human checkpoint belong?
Three places, no exceptions:
- Money. Anything that pays, credits, discounts, or quotes prices. The agent can prepare and stage. A human presses the button or samples the outgoing stream.
- Contracts and commitments. Anything a customer can derive rights from. After Air Canada and the Hamm ruling, every agent answer to a customer is legally your answer.
- Customer relationships under pressure. Complaints, escalations, bad news. Klarna said its AI assistant performed work equal to roughly 700 FTE. The CEO later acknowledged that an excessive focus on cost had affected quality and reinvested in reachable human support. The model remained AI-first, with humans for complex cases.
Notice what’s not on the list: sorting invoices, moving data, summarizing email, scheduling meetings. There, the checkpoint is a periodic spot check rather than a signature per action, and that’s where an agent earns its keep.
Is this temporary or permanent?
The execution limits keep shifting. Agents get better at exceptions and context, and what’s impossible today may be partly possible next year. But the core limit isn’t technical. Responsibility is a legal and human construct, and no model version will convince a courtroom that the software itself is to blame. Don’t plan for the checkpoint to disappear. Design your processes so the checkpoint is cheap rather than absent.
For employers, that’s the sober summary of this whole piece: the question isn’t whether you’ll still need people, but exactly where. Repetitive, rule-based work can go to the agent. The spots with money, signatures, and emotions stay human. How to draw that line in practice is also something we cover on our page for employers.
Not sure where the line falls in one of your own processes? That’s exactly the kind of question our agents intake is built for. We’ll look at it with you, point out where the checkpoint belongs, and tell you honestly when a process is better left entirely to people.
