In February 2024, a Canadian tribunal ruled on a case that started with Air Canada's website chatbot. A customer whose grandmother had just died asked it about bereavement fares. The bot told him he could book at full price and claim the bereavement discount afterward, within 90 days. That wasn't Air Canada's policy. When the airline refused the refund, the tribunal ordered it to pay the difference, about CA$812. Air Canada had argued that the chatbot was responsible for its own words. The tribunal rejected that, and a small payout turned into international headlines.

Whatever your AI tool says to a customer, you said it.

AI makes things up. That's normal and manageable.

For any business using AI, hallucination is a side effect of how these tools work, and you shouldn't plan around an update making it go away.

At its core, a language model doesn't look facts up. It predicts the most plausible next word, then the next, then the next. Most of the time "plausible" and "true" line up. Sometimes they don't, and the model gives you a fake statistic or a policy you never wrote in the same confident tone it uses for everything else. Tools that search the web or read your documents can still misread or mix up what they find, so the same checks apply.

Better models make fewer mistakes, but they still make them. In 2023, in Mata v. Avianca, two New York lawyers and their firm were sanctioned $5,000 jointly after filing a brief that cited six court cases ChatGPT had invented. The cases had real-sounding names, docket numbers and quotes. None of them existed.

So keep using AI, with two rules. Anything you send gets a human check first, which takes about two minutes per item. Anything that talks to customers on its own, like a chatbot, needs tight limits on what it's allowed to say and someone reading its transcripts. Air Canada's chatbot was free to explain a refund policy on its own, and it got the policy wrong.

What to check for

You don't need to fact-check every word. AI is good at tone and structure. It's bad at specifics. Put your attention where the errors actually happen.

Factual claims

Any number, date, price or product spec the AI states should be treated as unverified until you check it. Think "SEER2 rating of 17," "most homeowners replace their roof every 20 years," or "the 2025 model includes a 10-year parts warranty."

Check these against the source: the manufacturer spec sheet, your price list, the actual study. Your memory doesn't count. AI errors are convincing because they're close to right, and "close to right" is exactly what memory confirms.

Service area and availability

This one bites service businesses constantly. Ask AI to write a landing page or reply to a lead, and it will cheerfully say you serve Gilbert when you stop at Chandler, offer 24/7 emergency calls when you don't, or do commercial work when you're residential only.

The AI is filling gaps with what a business like yours usually offers. The customer reading it doesn't know the difference, and you'll be the one explaining on the phone why you won't come out.

Legal and regulatory references

Anything touching compliance gets a hard stop. That includes privacy laws like California's CCPA (which can apply even to out-of-state businesses with California customers, once they pass its thresholds: as of September 17, 2026, about $26.6 million in annual revenue, or data on 100,000 or more California residents or households), Arizona's data breach notification law, contractor licensing rules, HIPAA, fair housing language in real estate listings and statutes of limitations in a legal intake email. Requirements vary by state and change often, and AI training data is always somewhat out of date.

If the AI cites a specific law, a license requirement or a deadline, verify it against the official source or ask your attorney. If you can't verify it, cut it.

Anything that sounds too good

AI leans optimistic. It will promise "same-day service," call your results "guaranteed," or tell a patient a procedure has "minimal recovery time." Those are claims your business has to stand behind. If a sentence makes you think "well, usually," rewrite it to say usually.

Bias in AI output

AI models learn from huge amounts of human writing, and they pick up our habits along with it. Most of it is subtle. Some of it will cost you customers.

Patterns worth watching for in service businesses:

  • Gendered assumptions about decision-makers. Follow-up emails addressed to "the husband" about the HVAC quote, or scripts that assume the woman handles scheduling and the man approves the spend. Plenty of your buyers will notice, and some won't call back.
  • Income and zip code assumptions. Try it: ask AI to write outreach for two zip codes and compare the results. Watch for one version pushing financing and budget options while the other gets premium upsells, based on nothing but an address. In real estate this is more than a tone problem. Steering buyers by neighborhood can be a fair housing violation.
  • Oversimplified industries. AI tends to flatten legal and medical topics into tidy, reassuring summaries. A blog post that says "most personal injury cases settle quickly" or "this condition is easily treated" can set expectations your practice can't meet.
  • Default "customer" personas. Generated personas and sample customers often skew toward one age, background and family structure. If your actual client base looks different, your marketing will read like it's talking to someone else.

Don't panic about any of this, and don't go hiring a bias consultant. What works is someone reading the output and asking, "Would I say it this way to this specific customer?" Most of these problems are obvious once a person actually reads the draft, which is the same reason you shouldn't fire your copywriter just yet.

Building a lightweight AI quality control process

Big companies build AI governance committees, but a small business can get most of the benefit from three habits.

1. A two-minute checklist before anything goes out

Anything AI-drafted that goes to a customer, patient, client or employee gets a quick pass. Print this and tape it next to the monitor:

  • Are all numbers, prices and dates checked against a real source?
  • Does it only claim services and areas we actually cover?
  • Does it mention any law, license, regulation or deadline? If so, is it verified?
  • Is anything promised that we can't always deliver?
  • Does it make assumptions about who this person is?
  • Would I be fine with this being screenshotted and posted online?

Most drafts clear all six in under two minutes.

2. One person owns the weekly spot-check

Pick one person, usually whoever already reviews outgoing work (often your office manager). Once a week they pull five or ten AI-assisted items at random. That might be chatbot transcripts, review responses, quote emails or social posts. Auditing AI output this way is just error checking: they read each item looking for mistakes.

This catches what the two-minute check misses, especially for anything automated that runs without a human approving each message. But a spot-check finds mistakes after customers have already seen them. It would not have saved Air Canada. If you have an AI chatbot or auto-responder on your website, prevention has to come first:

  • Limit it to your own approved documents: your price list, your service area, your written policies. That's what a RAG setup is for.
  • Have it hand off to a person for refunds, discounts, policy exceptions, price commitments, and anything medical or legal.
  • Don't count on a disclaimer. The tribunal held Air Canada responsible for what its bot said, and a line saying "this assistant may make mistakes" won't change who owns the answer.

Then the weekly transcript review is how you find out whether those limits are holding.

3. Write down every error you catch

Keep a shared doc or a Slack channel. When someone catches a mistake, log it in one line: what the AI said, what was true, where it happened and what you changed.

After a month you'll see patterns. Maybe it keeps inventing a warranty term, or it always gets your service radius wrong. Each pattern points to a fix. Add the correct info to your prompt, put your real price list in the tool's instructions, or tell it plainly: "We do not serve Mesa. Never say we do."

Keep patient, client and customer personal details out of these tools unless your agreement with the vendor covers them. For medical practices, that means a signed Business Associate Agreement, which usually comes with an enterprise or API contract, not a self-serve "Team" or "Business" plan. Law firms owe clients confidentiality for anything pasted into an AI tool, so check its data terms first (more on that in AI data privacy for Phoenix small businesses). Your price list and service area are fine to paste in. A patient's chart is not.

Most tools don't learn from your corrections on their own, so the improvement comes from your prompts getting better as you work through the log.

Where to start

Pick the single AI workflow closest to your customers: the chatbot, say, or the quote follow-ups. Tape the checklist up this week and assign the spot-check owner. That's enough to start.

If you'd like a second set of eyes, book an AI audit. We'll look at how your team actually uses AI day to day and show you where errors are most likely to reach a customer, plus where your review process needs tightening. You'll get a list of what to fix.