What happens if an AI agent promises a customer something we can't deliver?
Last updated
The customer holds you to it, because to them the agent was your company. What you can change is whether you find out from the customer or from your own record, and whether the promise was ever made against something nobody stood behind.
A buyer who was told that integration ships in March does not ask whether an agent or a person said it. They plan around it. The sentence was sent in your name, and the cost of walking it back lands on whoever owns the account when the buyer comes back to it.
Why these go unnoticed
A broken promise has no failure moment. Nothing errors when an agent says something your company cannot do. The promise sits in a thread, the buyer remembers it and nobody else does, and the first sign is a renewal call or a support ticket months later. Promises that nobody closes do not expire; they rot quietly.
Three shapes cover most of them:
| what the agent said | what was missing |
| a feature the product does not have | a line anybody stands behind |
| a date | anything in place to keep it |
| we always, we never, unlimited | room for the exception the buyer will find |
What to have in place before it happens
- One place the agent reads from, where every line has a named owner and a date it was last checked. An agent repeating your product sheet is repeating whatever that sheet said the day somebody last edited it.
- A rule that an unusable line is escalated, never answered around. When the line is overdue, disputed or has no owner, the agent says it is checking and tells the buyer so. That is an open promise, written down, on a list somebody reads.
- A record of every promise as it is made: what was said, which agent, what it relied on and who owns that. When the customer says you told us March, there is an answer within minutes rather than a hunt through transcripts.
When it has already happened
Find the sentence, the line it rested on, and its owner. If the line was wrong, fix the line rather than the agent: every agent reading it was about to repeat the same thing. Then decide with the account owner what you will honour — that is a commercial decision, and no tool makes it for you.
What this does not do
Reynforce does not stop an agent from sending anything. It answers the agent's question before it promises, records what was promised either way, and keeps the open ones visible. Whether your agent waits for the answer is a choice in how it is built.
/can-an-ai-agent-promise-a-delivery-date.md covers dates on their own. /who-is-liable-when-an-ai-agent-makes-a-promise.md covers who carries it.