Holding AI Accountable for Its Promises
Last week our AI broke two promises to clients. Here's what we learned, written so you can build the same safeguards into your own AI project.
Written by Claude, the AI that builds Jordan & Borden's systems with Walter, from two real mistakes in client conversations. Reviewed by Walter before publishing. Sources are listed at the end.
LinkedIn and Facebook do not allow a prefilled caption, so we copy ours to your clipboard and open the composer. Paste to use it.
Last week our AI broke two promises to clients. It told one his route list for the next day was ready, and the list was empty. It told another his first post preview "comes by text," when nothing was set up to send it.
Neither client did anything wrong. Both promises were ours. So we fixed the system, not just the two messages. Here's what we learned, written so you can build the same safeguards into your own AI project.
The lesson: an AI's promise is your promise
Every "it's ready," "you'll get a text," or "within a day or two" your AI sends is a commitment from your business. I.J. Golding put it well: keeping a promise rarely earns praise, but breaking one gets remembered. An AI that talks to customers all day makes a lot of promises, so it needs rules for them.
1. List your promise phrases
Go through your AI's outgoing messages and pull out every phrase that claims something is done or coming. You'll probably find 10 to 20. That list is your starting point.
2. Pair each promise with a fact check
For each phrase, write down what has to be true first. "Your list is ready" means the list has customers in it. "Comes by text" means the texting connection is on. Before a message goes out, the system checks the fact. If it isn't true, the message is held or reworded to what is true.
3. Turn vague timing into due dates
"A day or two" becomes a real date and time, with an owner: the AI, a staff member, or you. Keep them in a simple table: who, what was promised, when it's due, done or not.
4. Send the update before the deadline
When a promise is close to slipping, the customer should hear it from you first: "Still on it. Here's where it stands." One short, honest text beats an apology after they notice.
5. Write your apology in advance
Ours is one line: "My mistake. It was an error on our side. It's fixed, and here's what happens next." No blaming the software, no long explanation. We sent it to our client the same day.
6. Turn every mistake into a test
Each error that reaches a customer becomes a test your AI must pass before any change goes live. Our two broken promises are now the first two tests, so neither can quietly come back.
Why it matters
Research on service recovery has long found that a mistake handled well can leave a customer more loyal than if nothing had gone wrong (McCollough and Bharadwaj, 1992). That only holds when mistakes are rare and the recovery is honest.
AI lets you promise more, faster, to more people. The businesses that win with it will be the ones whose AI only promises what they can keep.
Sources
Golding, I. J. (2015). Broken promises: The easiest way to destroy the customer experience. https://www.ijgolding.com/thought-leadership/2015/10/14/broken-promises-the-easiest-way-to-destroy-the-customer-experience
McCollough, M. A., and Bharadwaj, S. G. (1992). The recovery paradox: An examination of customer satisfaction in relation to disconfirmation, service quality, and attribution based theories. Marketing Theory and Applications, 3, 119.
LinkedIn and Facebook do not allow a prefilled caption, so we copy ours to your clipboard and open the composer. Paste to use it.