← Back to Insights

AI You Can Trust: What to Check Before It Acts for Your Business

Corbin CookAI Myths, Trends & Trust
Owner shaking hands across a meeting table while a colleague reviews a checked-off approval list on a tablet.

Most owners we talk to are already using AI somewhere. A proposal gets a first draft. A messy email gets cleaned up. Meeting notes turn into a task list. That part feels safe because a person reads the result before anything happens.

The harder question comes next: when do you let it do something without a person in the middle? Send the reply. Book the appointment. Mark the invoice as followed up.

That's where trust actually gets tested, and it's a fair thing to be careful about. The good news is that trust here isn't a gut feeling about "AI" in general. It's a short list of checks you can run on one specific workflow.

Everyone's using it. Fewer people trust it.

The gap is real and it's measurable. In a 2025 global study led by the University of Melbourne with KPMG, which surveyed more than 48,000 people in 47 countries, 66% said they use AI regularly. Only 46% said they were willing to trust AI systems. The same study found many people rely on AI output without checking its accuracy (KPMG summary).

Read those two findings together and you get the actual risk for a small business. It isn't that AI is too powerful. It's that people use it, half-trust it, and skip the checking anyway.

The fix is to decide ahead of time what the system is allowed to do and how you'll know it's doing it well.

The businesses that own the mistake

Here's a public example worth knowing. In 2022, a customer asked Air Canada's website chatbot about bereavement fares. The chatbot said the discount could be claimed after the trip. The airline's actual policy said otherwise. When the customer asked for the refund, Air Canada argued it couldn't be held responsible for what the chatbot said.

The British Columbia Civil Resolution Tribunal disagreed. In Moffatt v. Air Canada (2024), it found the airline responsible for the information on its own website, chatbot included, and ordered it to pay the difference plus interest and fees (tribunal decision).

The dollar amount was small, a little over $800 Canadian. The lesson isn't. If an AI tool speaks or acts for your business, customers will treat it as your business. "The software said it" won't be a defense with a client, and it may not be one anywhere else.

Four levels of "letting AI act"

It helps to stop thinking of AI as on or off. Most useful workflows sit somewhere on a short ladder:

  1. Suggest. AI points something out. "These five invoices are past 30 days." A person decides what to do.
  2. Draft. AI prepares the work. A person reviews and sends it.
  3. Act within limits. AI does the routine version on its own, inside rules you set, and hands everything else to a person.
  4. Act and report. AI handles the job end to end and you review a summary and a sample.

Most first workflows should start at level one or two. That isn't timidity. It's how you collect the evidence to move up. Each step up the ladder should be earned with a record, not granted because the demo looked good.

The checks to run before AI acts on its own

Before a workflow moves to level three or four, you should be able to answer each of these in plain language. If you can't, it isn't ready yet.

1. What is it allowed to touch?

Name the exact actions: send this type of email, book into these calendars, update this field in the CRM. Then name what it can't do. Changing prices, issuing refunds, and promising dates are common lines to draw. Give the tool access only to what those actions need.

2. Where does its information come from?

The Air Canada chatbot gave an answer that didn't match the policy. Your version of that risk is an AI reply built from an old price sheet, last year's service list, or a policy nobody updated. Pick one source of truth for each thing it's allowed to say, and make someone responsible for keeping that source current.

3. Who owns the result, by name?

"The team" doesn't own anything. A named person reviews the output, handles the exceptions, and gets the call when something goes wrong. Name a backup for when that person is out.

4. Where are the stop signs?

Write down when the system has to stop and hand off. Common ones: an upset customer, anything over a dollar amount you choose, anything that can't be undone, anything involving health, legal, or financial details, and anything that doesn't fit the rules. A good workflow knows what it shouldn't handle.

5. Is there a record?

Every action should leave a trail: what it did, when, and based on what. You'll need that the first time a client says, "Your system told me something different."

6. How will you know it's slipping?

Pick a few measures before go-live and check them weekly. Useful ones are how often a person corrects the output, how often it hands off, how many customers ask for a human, and the outcome you built it for, like reply time or follow-ups completed. If corrections start climbing, it moves back down a level.

7. What happens when it fails?

Tools go down and connections break. Decide what happens then. Who gets alerted, what the customer sees, and how the work gets done by hand until it's fixed. Customers should always have an easy way to reach a person.

None of this is exotic. If you want a deeper reference, the National Institute of Standards and Technology publishes a free AI Risk Management Framework, built around four functions: govern, map, measure, and manage. It's written for larger organizations, but the seven checks above are the same ideas at small-business size.

A hypothetical, to make it concrete

Picture a 14-person home services company. It's a made-up example, but a common situation. The office manager spends part of every morning following up on quotes that went out and never got an answer.

The owner wants AI to send those follow-ups. Here's how the checks might shape it:

  • Allowed to touch: a follow-up email on quotes three or more business days old, using approved templates. It can't change a price, offer a discount, or promise a start date.
  • Source of truth: the quote record in the CRM. If a quote isn't in the CRM, it doesn't get a follow-up.
  • Owner: the office manager, with the operations lead as backup.
  • Stop signs: any reply that asks a question, pushes back on price, or sounds unhappy goes straight to the office manager. So does any quote above an amount the owner sets.
  • Record: every follow-up is logged on the customer's record.
  • Measures: quotes followed up on time, edits the office manager makes, replies handed off, and quotes that turn into booked jobs.
  • Failure plan: if the connection breaks, the office manager gets an alert and the morning list goes back to manual.

For the first few weeks it runs at level two. AI drafts each follow-up and the office manager approves it. Once the edits drop to almost nothing over a stretch the owner agreed to up front, it moves to level three for standard quotes only.

The return here would show up first as capacity: hours back for the office manager and fewer quotes slipping through. It only becomes added revenue if more of those quotes turn into jobs, and you'd measure that, not assume it. We break down that distinction in AI ROI for small businesses.

The new job is oversight

There's a common fear that letting AI act means losing your people. In practice, a well-run workflow does the opposite of removing judgment. It takes the repetitive part off someone's plate and leaves them with the part that needs a person: the unhappy customer, the odd request, the call on whether to bend a rule.

Someone also has to watch the numbers, keep the source information current, and decide when the system earns more room. That's a real role, and it usually belongs to someone who already knows the work well.

If the person who ends up answering every exception is the owner, that's a different problem worth solving first. Our piece on owner dependency and AI implementation covers it.

Questions to ask this week

  • Where are we already letting AI send, book, or update something without a person reading it first?
  • For that workflow, could we answer all seven checks in plain language?
  • What's the one thing we'd never want it to say or do for us, and is that written down?
  • Who would notice first if it started getting things wrong?

Where to start

The checks are easier to run once the process underneath is solid. Before You Automate walks through the people, process, and data side, and From Blueprint to Working System covers what a complete build should include.

If you want a starting point for your industry, grab the free AI Leverage Report. It shows where AI tends to pay off in businesses like yours.

And if you're weighing whether one of your workflows is ready to act on its own, book a free 30-minute AI Leverage Call. We'll look at the workflow with you and talk through which level it should start at. When it's time to build, the AI Implementation Blueprint is where we map the limits, owners, and measures before anything goes live.

Corbin Cook
Corbin Cook
Founder of SMB Strategy Consultants. Helps owner-led service, tech, and SaaS businesses find the highest-return use of AI and implement it in the right order.

Ready to talk through your business?

Book a free 30-minute AI Leverage Call. We'll walk through your operations and tell you honestly where the biggest return is likely to be.

Book Your Free AI Leverage Call