Your AI agent just went off script.
That is not a hypothetical. If you have an AI agent—a piece of software that talks to your customers by chat, text, or phone on your behalf without a human in the loop—handling your follow-ups, booking reminders, or inquiry responses, it will eventually hit a situation it was not built for. The question is whether you find out because you caught it, or because a customer tells you they are never calling back.
This guide is not about fear. AI agents genuinely do things well: they answer at 2 a.m., they never forget to follow up, and they handle the repetitive stuff so your team handles the real stuff. But "handles the repetitive stuff" is the key phrase. The moment a conversation stops being repetitive, the agent needs to know that—and so do you.
Here are the three warning signs to watch for, and what to do when you see them.
---
Warning Sign #1: The Customer Is Expressing Frustration or Emotion
An AI agent is pattern-matching against a set of instructions. It is very good at recognizing keywords like "appointment," "price," or "cancel." It is not good at recognizing when a person has gone from mildly annoyed to genuinely upset—and the difference matters enormously in a service business.
Think about the situations where you would naturally lower your voice and slow down: a customer whose repair job came back wrong, a patient anxious about a procedure, a homeowner who just found out the damage is worse than expected. Those conversations require a human being who can read tone, absorb frustration, and make a judgment call.
An agent that keeps cheerfully offering booking links to a customer who is expressing anger is not helping. It is accelerating the problem.
What to watch for:
- Messages that include words like "frustrated," "upset," "ridiculous," "wrong," "never again," or "lawsuit"
- Repeated messages from the same contact in a short window—a sign they feel unheard
- Any message where the customer describes a specific harm: something broke, someone got hurt, something was missed
What to do about it:
Build a keyword list into your agent's rules. When any of those words appear, the agent stops, sends a holding message ("Let me get someone from our team on this right away"), and routes the conversation to a real person. Check whether your platform supports this kind of escalation trigger—good platforms will. If yours does not, that is worth fixing before you scale.
---
Warning Sign #2: The Conversation Requires a Professional Judgment Call
This one is subtle, and it trips up a lot of well-intentioned setups.
A customer asks your HVAC agent whether they should replace or repair their unit. A patient asks your clinic's chat agent whether their symptoms sound like something serious. A homeowner asks your roofing agent if the damage qualifies for an insurance claim. These are not FAQ questions. They are judgment calls that require a licensed professional, or at minimum someone with real expertise and accountability.
An AI agent will attempt to respond if you let it. It can generate something that sounds confident and complete. It may even be right some of the time. But "some of the time" is not good enough when the stakes involve a customer's health, a major purchase decision, or a claim that ends up in dispute.
Your agent's job is to get information, confirm appointments, and keep the customer in the loop—not to practice your trade on your behalf.
What to watch for:
- Questions that start with "Should I…," "Is it safe to…," "Does this qualify for…," or "What do you think is wrong with…"
- Any conversation where the customer appears to be making a significant financial or health-related decision based on the agent's reply
- Questions about warranties, liability, or coverage—anything that could become a contractual issue
What to do about it:
Train your agent to recognize these patterns and respond with something like: "That is a great question for one of our technicians—let me have them reach out to you directly." Then make sure the handoff actually happens. An escalation that goes nowhere is worse than no escalation at all, because the customer was told someone was coming.
---
Warning Sign #3: The Agent Is Giving Inconsistent or Contradictory Answers
This one is harder to catch because it does not show up in a single conversation—it shows up across conversations, or when a customer circles back with "but the text said…"
AI agents can drift. If your agent is drawing from multiple sources—a FAQ doc, a script, a product sheet, a calendar feed—it can sometimes synthesize information in ways that produce small contradictions. Your website says one thing, your agent says another, and now the customer has a screenshot that disagrees with what your tech told them in person.
Even without multiple sources, agents can be inconsistent when the same question is phrased differently by different customers. This may not be a catastrophic failure—it can simply be the nature of how these systems tend to behave. But for a service business, consistency is trust. If customers can get different answers to the same question depending on how they ask it, you can end up with a credibility problem that is hard to trace back to its source.
What to watch for:
- Customer complaints that reference something the agent "told them" that conflicts with your actual policy
- Situations where your team has to walk back something the agent said
- Internal confusion about what the agent is actually telling people—if your own staff does not know, your customers definitely do not
What to do about it:
Audit your agent's conversations regularly. Many platforms keep conversation logs; check whether yours does and turn that feature on if it is available. Set aside time—even thirty minutes every couple of weeks can be a reasonable starting point, though busier operations may need more—to read a sample of recent threads. Look for hedging language ("I think," "I believe," "it may be"), outright errors, or anything that would embarrass you if a customer forwarded it. When you find a problem, update the source material the agent is drawing from, and test it before it goes back to customers.
---
The Bigger Point: Guardrails Are Not a Weakness
Some business owners worry that adding handoff rules makes their AI agent seem less capable. The opposite is true. A customer who gets transferred to a real person at exactly the right moment feels taken care of. A customer who gets a cheerful automated response to a complaint they filed three times feels ignored.
The goal is not a fully autonomous agent. The goal is the right message, delivered by the right party, at the right moment—your agent handling everything it can handle well, so your team has bandwidth for everything else.
That balance is something you tune over time. You start, you watch the conversations, you fix what is off. No setup is perfect on day one, and anyone who tells you otherwise is selling you something.
---
What to Do This Week
- Pull up your agent's conversation logs and read twenty recent threads from start to finish. Look for frustration language, judgment-call questions, or anything that sounds inconsistent with what you actually offer.
- Make a short keyword list of phrases that should always trigger a human handoff—frustrated, upset, lawyer, claim, hurt, wrong, and any industry-specific terms that signal a serious situation. Add that list to your agent's escalation rules.
- Test your escalation path end to end. Send a test message that includes one of your trigger words and follow it all the way through. Does a real person actually get notified? Does the customer get a holding message? Does the handoff happen fast enough to matter?
- Pick one category of question your agent should never answer directly—diagnosis, pricing disputes, warranty claims, or whatever applies to your trade—and write a redirect response the agent will use every time it sees that pattern.
- Schedule a recurring calendar block to review agent conversations. Thirty minutes every couple of weeks is a reasonable place to start; adjust based on your volume. Skip it for a month and you are flying blind.
Your agent is a tool. A well-configured, well-supervised tool that your customers never feel is ignoring them. Keep it that way.