Most teams check an AI agent carefully before launch and then stop looking. The agent answers hundreds of messages, nobody complains loudly, and the assumption becomes that it works. Problems tend to be quiet: an outdated price repeated with confidence, a tone slightly too casual for a complaint, a case that should have gone to a person and did not.
The remedy is not a large monitoring project. It is a small, boring routine: once a week, someone reads a handful of real conversations and scores them against the same few criteria. This article describes how to set that up so that it actually happens.
Why a sample, and why weekly
Reading every conversation is not realistic, and it defeats the purpose of having an agent. Looking only at conversations customers complained about gives you a biased picture, because most mistakes are never reported. A random sample is the middle path: cheap enough to keep up and honest enough to show what typical quality looks like.
Weekly is a good starting rhythm because it is frequent enough to catch a problem within days, and it matches how most teams already plan. In the first weeks after launch, or after any change to the agent's instructions or sources, sample more. Once the picture is stable, you can lengthen the interval, but do it deliberately and write down why.
Decide who reviews
The reviewer should know the subject well enough to tell a right answer from a plausible one. For customer support this is usually an experienced support colleague. For internal questions it may be someone from HR or IT.
Two rules help:
- The reviewer is not the person who built or configured the agent. People who made the setup tend to read answers generously. A fresh pair of eyes notices more.
- The review has a fixed slot and a named backup. "We will look at it when we have time" means it stops after three weeks. Thirty to sixty minutes in the calendar, with a deputy for holidays, is enough.
Also decide who receives the result. One person, typically the owner of the agent, should be responsible for acting on it.
How big should the sample be
There is no universal number, and we will not pretend there is one. Choose a size that one person can read carefully in the time slot you set, and that gives you a few conversations from each important type of request. For a small team, that may be ten to twenty conversations a week. If the agent handles very different topics, make sure each topic appears in the sample, not only the most frequent one.
Some practical points:
- Pick at random, ideally with a simple script or a filter in your tool, rather than letting the reviewer choose. People pick the interesting ones.
- Add the risky ones on purpose. Besides the random part, include conversations where the agent handed over to a person, where the customer expressed frustration, or where money, contracts or personal data were mentioned.
- Keep the method stable. If you change how you sample every week, you cannot compare weeks.
Treat the conversations as personal data and show the reviewer only what is needed. If you have not yet decided what to log and keep, our article on what an AI agent activity log should record is a good place to start.
Four criteria that cover most problems
Keep the scoring sheet short. Four criteria are enough for most customer-facing and internal agents.
1. Correctness. Is the answer true, current and based on an approved source? Check specific claims: prices, dates, conditions, steps. An answer that sounds right but cannot be traced to a source counts as a failure.
2. Scope. Did the agent stay within what it is allowed to do? It should not give legal advice, promise refunds it cannot grant, or comment on topics outside its remit. Scope failures often show up as helpful-sounding improvisation.
3. Tone. Was the answer polite, clear and appropriate for the situation? An upset customer needs a different register than someone asking about opening hours. Tone is subjective, so write down the standard once (a few example sentences are better than adjectives) and use it for every review.
4. Handover to a person. Did the agent pass the case on when it should have, and did it do it well? A good handover tells the customer what happens next and gives the colleague the context so nobody has to ask the same questions again. The opposite error matters too: an agent that escalates everything is not saving anyone time. The general triggers are described in when an AI agent should stop and pass the case to a person.
For each criterion use a simple scale, for example pass, minor problem, serious problem. Avoid ten-point scales: nobody can explain the difference between a six and a seven, and the numbers give a false sense of precision.
Write one line for every failure
A score alone does not help the person who has to fix things. For every minor or serious problem, the reviewer writes one short note: what was asked, what the agent said, what would have been right. Add a guess at the cause if there is an obvious one, such as "source document is out of date" or "instruction does not say what to do with refund requests".
These notes are the most valuable output of the whole routine. After a few weeks they show patterns that no single conversation reveals.
When to intervene right away
Most findings can wait for the weekly summary. Some cannot. Agree in advance which ones need action the same day, so the reviewer does not have to ask permission:
- a wrong statement about money, contracts, safety or legal obligations,
- personal data shown to the wrong person or used where it should not be,
- a missed handover in a situation the rules clearly mark as one for a person,
- the same serious error appearing in more than one conversation.
The immediate action should be proportionate. It can mean correcting the customer, tightening an instruction, removing an outdated source, or, in the worst case, switching the agent to "suggest only" mode until the cause is fixed. Decide who has the right to do that and how it is reversed.
Feed the results back into the agent
Scoring without change is paperwork. Each week, the owner of the agent goes through the notes and sorts the findings by what needs to change:
- Sources. The answer was wrong because the document was outdated, missing or contradictory. Fix the document, not the symptom.
- Instructions and rules. The agent was allowed to do something it should not, or lacked guidance for a recurring situation. Add the rule in plain language.
- Handover rules. The agent escalated too late or too often. Adjust the triggers and the information passed on.
- People and process. Sometimes the problem is not the agent, for example when colleagues give conflicting answers to the same question. That is a finding too, and it belongs with the team, not with the configuration.
Keep a short change log: what was found, what was changed, on which date. The next week, check whether conversations of the same kind improved. If a change made things worse or had no effect, you will see it and can reverse it. This also connects naturally to a wider monthly look at whether the agent is worth keeping, as described in the AI agent monthly review.
What to report upwards
Managers do not need every note. A one-paragraph summary is enough: how many conversations were reviewed, how many had a problem in each of the four areas, what was changed and what remains open. Report counts and trends honestly, including weeks when quality dropped. A report that is always green is not believed, and rightly so.
A short checklist
- A named reviewer who did not configure the agent, with a fixed weekly slot and a deputy.
- A random sample plus a few deliberately risky conversations, handled as personal data.
- Four criteria - correctness, scope, tone, handover - with a simple three-level scale.
- One written line for every failure, including the likely cause.
- A list of findings that require same-day action and a person allowed to act.
- A weekly change log and a check, the week after, that the change worked.
An AI agent does not stay good by itself. Sources age, products change and customers ask new things. A small weekly review costs an hour and gives you what a launch test cannot: evidence of how the agent behaves in real conversations, and a steady way to make it better.
