Update (Nov 2, 2026): Pricing has changed. Runexus now offers one product, All-Access: $39.95 per website, one time. See pricing
Building a Safe AI Agent for a Local Business
A safe AI agent for a local business answers only from facts the owner has approved, says "I don't know" and hands off to a person when a question goes beyond those facts, never commits to prices or promises on its own, and logs every answer. A tribunal has already ruled that a business owns what its chatbot says.
In 2022 a man booked flights after his grandmother died and asked Air Canada's website chatbot about bereavement fares. The chatbot told him he could buy a full-price ticket and claim the bereavement discount afterwards. Air Canada's own bereavement page said the opposite. When he applied for the refund, the airline refused, and argued in part that the chatbot was responsible for its own words.
The tribunal disagreed. That case, and a public chatbot in New York that gave business owners advice that broke the law, are good guides to what an AI agent for a local business should and shouldn't be allowed to do. This article starts with what went wrong, then turns it into a design.
What happens when a business's chatbot gets it wrong?
The business is responsible for it. In Moffatt v. Air Canada, 2024 BCCRT 149, decided Feb 14, 2024, British Columbia's Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation and ordered it to pay C$812.02, made up of C$650.88 in damages plus interest and fees. The tribunal's reasoning is the part every business owner should read:
"While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot."
The amount was small. The principle isn't: what your agent says, you said.
The second case is a public one. In March 2024, The Markup reported that New York City's MyCity chatbot, built to help business owners, told them they could take a cut of workers' tips and that landlords could refuse tenants who use rental assistance. Both answers were wrong under city law.
Neither bot was missing information. Air Canada had a correct bereavement page. New York's rules are public. The bots produced fluent answers that went beyond what they had been given, and nothing stopped them.
How do I stop an AI chatbot from making things up about my business?
Have it answer only from a set of facts you've approved, and say "I don't know, here's how to reach us" when a question goes beyond them.
In practice that means three design choices:
- A fixed source. The agent answers from a written list of approved facts (services, areas, hours, contact routes, policies), not from whatever it can infer from your website or the open web.
- Permission to not know. Model vendors recommend the same thing in their own documentation. Anthropic's guidance on reducing hallucinations, for example, says explicitly allowing a model to admit uncertainty "can drastically reduce false information." For a business agent, "I don't have that information; call us at..." is a correct answer, not a failure.
- Fixed answers for the questions that matter most. "Do you service Carrier units?" should return the approved answer every time, word for word, not a fresh paragraph generated on each request. Creative phrasing is fine for a greeting. It isn't fine for a refund policy.
Apply that to the Air Canada case: if the bot could only state the approved bereavement policy, or hand off when unsure, it couldn't have invented a retroactive claim window.
What should an AI agent never do for my business?
An AI agent should never quote pricing exceptions, give legal or medical advice, or commit to anything you haven't approved; those go to a person, and conversations should be logged.
A scope table makes this concrete. Here's one for a fictional two-van HVAC company:
| Request | Agent does | Why |
|---|---|---|
| "Do you serve Allen?" | Answers from the approved area list | Approved fact |
| "Are you open Saturday?" | Answers from approved hours, including holiday overrides | Approved fact |
| "Can you match a competitor's quote?" | Hands off to a person | Pricing exception |
| "Is this gas smell dangerous?" | Says to leave the house and call the gas company's emergency line, then hands off | Safety; not a judgment the agent should make |
| "Can you guarantee it's fixed today?" | Declines and offers the booking contact | Commitment the owner hasn't approved |
| "How much for a new system?" | Gives the published range if one exists; otherwise hands off | Quotes need an inspection |
The risk scales with the business. A wrong answer about a restaurant's patio hours costs a table. A wrong answer at a clinic, a law office or a financial adviser can cause real harm, and those agents should refuse anything that looks like professional advice.
The NIST AI Risk Management Framework is a useful outside reference if you want a structured way to think this through. NIST describes it as "intended for voluntary use," and its core idea fits here: decide what the system may do before you deploy it, and keep checking.
Which edge cases break simple agents?
Most failures happen at the edges of an approved fact, not in the middle of one:
- After-hours questions. "Can someone come tonight?" needs an answer that reflects your real after-hours policy, including "no" if that's the truth.
- Holiday hours. If the approved hours have no holiday override, the agent will give normal hours on Thanksgiving. Add overrides as dated facts and remove them when they expire.
- Out-of-area requests. A customer one town past your list should get a clear "we don't currently serve that area," not a hopeful "we may be able to help."
- Quotes with too little information. "How much to fix my AC?" can't be answered honestly. The agent should ask for what's needed or hand off, not name a number.
- Questions about similar-sounding things. If you're approved for two equipment brands, a question about a third shouldn't get a yes because the names sound related.
Write each of these down before launch and test the agent against them, the same way you would test any other system before it talks to customers.
How should changes to the agent go live?
Through versioned releases you approve, not live edits. A draft change to your hours shouldn't change what the agent says until someone publishes it. When the agent gives a wrong answer on a given date, you should be able to see which version was live and which fact produced the answer. If a release turns out to be wrong, you roll back to the last good one.
Keeping one approved fact list is also why the agent and your website should be updated together. If the website says Saturday 8 to 2 and the agent says 9 to 1, customers can't tell which to believe. Neither can a search engine reading both.
What does Runexus Site Agent do today?
On the Core plan, Site Agent publishes the facts you've approved as a machine-readable answer interface, hosted by Runexus and linked from your website. Draft changes stay private until you publish them, so nothing reaches the public by accident. It isn't a general-purpose website chat widget. It doesn't take payments, negotiate or book jobs.
The free quick snapshot checks the readiness side of this. If your site links a Runexus Site Agent interface, the scan sends it two fixed test questions: "What is the business name?" and "Can you guarantee a ranking or recommendation result?" A safe agent answers the first from approved facts and declines, or says it doesn't know, on the second. The test goes to your own endpoint only; it doesn't ask any AI assistant about your business.
Publishing an endpoint makes your facts available to systems that look for them. It doesn't guarantee any outside platform will use it, and a business that hasn't sorted out its basic facts should do that first: What an AI-Ready Local Business Profile Needs covers that groundwork, and How Answer Engines Build a Local Shortlist explains where an agent fits in the larger picture.
Is an agent worth it for a small business?
Only if you'll maintain it. An agent with stale hours is worse than no agent, for the same reason the Air Canada bot was worse than the static page it contradicted. If you can keep a short fact list current and review the questions it couldn't answer once a month, a narrow agent is a reasonable addition. If you can't, put the effort into your website and listings first.