What AI agents get wrong, and how to catch it
The specific failure modes of AI phone agents, why they happen, and the review habits that catch them before a customer does.
Every vendor in this category would rather talk about what their agents get right. That is a mistake, because the businesses who succeed with this technology are the ones who knew the failure modes on day one and built a review habit around them.
Here is the honest list.
It invents things that sound right
Language models produce plausible language, not verified fact. Asked a question your documents do not answer, a poorly configured agent will generate an answer with the same confident tone it uses for facts it actually has. A price that is close to your real price. A policy that sounds like something a company would have.
This is the single most damaging failure, because the caller has no way to tell the difference — and neither will you, unless you read the transcript.
How to catch it: search your transcripts for numbers. Prices, timeframes, measurements. Then check each one against a document you actually wrote. How to reduce it: instruct the agent explicitly to refuse and take a message rather than estimate, and test that instruction by asking it something you never told it.
It mishears, and repeats the error downstream
Speech recognition is very good and not perfect. Postcodes with letters that sound alike, unusual surnames, house numbers over background noise, accents underrepresented in training data. An error at this layer propagates: the wrong address goes into the booking, and a van arrives at the wrong street.
How to catch it: require read-back for anything consequential — addresses, phone numbers, dates, spellings — and check the transcripts show it happening. Read-back converts a silent error into one the caller corrects.
It is agreeable when it should not be
Models are trained to be helpful, and helpfulness under pressure looks like agreement. A caller who pushes — "you can do it for two hundred, right?" — can walk a weakly configured agent into confirming something it should have declined.
How to catch it: search transcripts for discount and exception language. How to reduce it: state the boundaries as absolutes in the instructions, and name what to do instead of conceding.
It misses tone
A caller who is upset but polite reads as fine to a transcript-driven system. Humans hear it immediately. This is why escalation should trigger on more than the phrase "speak to a human" — repeated questions, a second call in a day, and explicit frustration should all route to a person.
It is worst in week one
Accuracy is not a fixed property. It starts low, because the agent has not met your edge cases, and improves as you correct it. A business that reads transcripts daily for the first month ends up somewhere very different from one that switched it on and looked again in June.
A review habit that works
- First 30 days: read every transcript. It is fewer than it sounds, and it is where all the learning is.
- After that: read every escalation, every conversation tagged with a negative outcome, and a random sample of ten percent.
- Weekly: search for numbers, prices and commitments and verify a handful.
- Whenever you find an error: fix the instructions the same day, and note what you changed.
None of this is unique to AI. You would not put a new hire on the phones unsupervised and check in after a quarter. The difference is that the transcript makes supervision easy — which is the actual advantage, and the reason to insist on a product that records everything.
Next: the readiness test for putting one on your main line, and why latency decides whether calls work at all.