AI agents now answer a large share of customer conversations: chat bots in the help centre, AI replies in the inbox, voice agents on the phone line. They are fast and never tired. They also make the same mistake every time until someone notices.
That makes QA for AI agents different from QA for people in one important way. A person who gets something wrong gets it wrong on some calls. An AI agent with a gap in its instructions gets it wrong on every conversation that touches that gap.
Use the same rules as your people
The simplest principle is also the most useful: hold AI agents to the same standard as your team. If a person must confirm identity before sharing account details, so must the bot. If a person must not promise a refund without approval, neither may the bot.
Using the same rules has three benefits:
- Fair comparison. You can see whether conversations handled by the AI agent pass as often as conversations handled by people.
- No blind spots. Rules written for people often cover exactly the risks that AI agents fall into, such as promising things that aren’t true.
- One queue. Failures from people and from AI agents land in the same review queue, so nothing slips between teams.
What goes wrong with AI agents
AI agents fail in recognisable ways. Add checks for these on top of your usual scorecard:
- Made-up policies. The agent states a policy, price or deadline that doesn’t exist. Check: the agent does not state a refund window, fee or deadline that differs from your published policy.
- Missed hand-offs. The customer asks for a person, or the conversation needs one, and the agent carries on. Check: when the customer asks to speak to a person, the agent offers to transfer them.
- Missed vulnerability. The customer mentions something serious, such as bereavement, illness or financial difficulty, and the agent answers the literal question only. Warning: the customer mentions a vulnerability. Send these to a person.
- Unauthorised actions. The agent promises a refund, credit or exception that needs a person’s approval. Check: the agent does not promise a refund, credit or exception.
- Circular conversations. The customer asks the same question several times and never gets an answer. Check: by the end, the customer’s original question has been answered or handed to a person.
- Wrong confidence. The agent answers a question it should have declined, such as legal, medical or financial advice. Check: the agent does not give personal financial, legal or medical advice.
How much to check
Because AI agents repeat themselves, patterns appear quickly. That is good news: you don’t need to wait months to spot a problem. It is also a reason to check every conversation rather than a sample. A gap that shows up in 1% of conversations can still affect hundreds of customers a week, and a small sample will rarely catch it.
Check more closely after any change to the agent’s instructions, knowledge base or model. Most AI agent problems are introduced by a change, not by drift.
Keep people for the judgement calls
Automatic checks are good at confirming whether something happened. They are less good at deciding whether an unusual answer was reasonable. The best setup sends clear results straight through and puts unclear ones in front of a person, with the relevant lines already highlighted.
For AI agents specifically, reviewers should also feed what they find back to whoever maintains the agent. A failed check on a bot conversation is often a missing article in the knowledge base or an instruction that needs rewording.
Label the speakers
When you import AI agent conversations, make sure the AI’s turns are labelled as the AI, not as a human agent. Otherwise your reports will blend the two, and you won’t be able to compare them. Most helpdesks record which replies came from a bot; check that your QA setup keeps that label.
Where Assay fits
Assay checks AI agent conversations against the same rules as your people. When importing from Intercom, replies from the AI agent are labelled separately, so you can see how bot conversations compare. Every result points at the line that decided it, and anything uncertain goes to a person.