AI with a license, not AI with the keys.

Many AI policy templates focus on what staff may paste into a chatbot. That matters, but it misses the risk of an AI that can act: send an email, confirm a booking, quote a price, or answer a customer. LocalPixel’s working rule is simple: let the AI draft, summarize, classify, and propose. Keep a person responsible for anything that spends money, commits the business, or reaches a customer. Here is how that rule becomes a practical first draft for a ten-person team.

Why a policy about pasting into ChatGPT is not enough

Templates often start with data leaving the business, and that still needs a clear rule. AI can also sit inside workflows with permission to act, which adds the separate risk of an unauthorized decision. KPMG Canada found only 29 percent of employees say their employer has a comprehensive AI-use policy, and 40 percent do not know what controls exist. Stanford’s 2026 AI Index counted 362 publicly reported AI incidents in 2025, up 55 percent in a year.

Putting a person somewhere in the process does not automatically make it safer. A 2024 review in Nature Human Behaviour of 106 experiments found that human-plus-AI combinations, on average, performed worse than the better of the human or AI alone, especially on decision tasks, though they outperformed humans alone. That result does not tell a small business exactly how to divide the work. Our recommendation is to make the division explicit: let AI prepare and let a named person make consequential decisions.

The three tiers of decision rights

Sort every action an AI could take in your business into one of three tiers. A small team can do the first pass in an afternoon, then use it as a working policy that is reviewed as the workflow changes.

Tier 1

Licensed to act alone

Internal, reversible, and nobody outside the business sees it.

Summarize a thread before a handoff. Draft a reply into a queue. Tag an inbound request by type. Pull background on a company before a call. Flag an invoice that looks late. Keep these internal and reversible, with a person able to correct a bad result. That does not remove privacy or misrouting risk, so the approved-tools rule still applies.

ExamplesDrafts, summaries, classifications, research, and flags. Output lands in front of a person by design.

The testIf the AI is wrong, is the fix a click? If yes, license it. If the fix is an apology, it is not tier one.

Tier 2

Needs a signature

The AI proposes, a named person confirms, the system acts.

Sending a customer reply. Confirming a booking against the calendar. Issuing a quote from the price book. Posting a review response. Applying a credit. The AI can do the reading and typing; a named person reads the result and approves it. Keep a record of that approval where the action warrants one.

ExamplesAnything that reaches a customer, touches money, or commits the business to a time or a price.

The testWould you want your name on it if it were wrong? Then your name goes on it before it goes out.

Tier 3

Never delegated

Decisions about people, money at scale, and anything with a legal edge.

Hiring, firing, and pay. Pricing strategy rather than a single quote. Complaints and refunds above a threshold you set. Anything that resembles legal, medical, or financial advice. Deleting or merging records. Here the AI may inform but it may not even propose, because a plausible proposal is the problem: it gets approved because it looks finished.

ExamplesConsequential decisions about a person, and anything where a regulator or a court would ask who decided.

The testIf a customer would be angry to learn a machine was involved at all, it is tier three, whatever the efficiency case.

What the incidents teach, and why they are small business problems

In February 2024, the BC Civil Resolution Tribunal decided Moffatt v. Air Canada. The airline’s chatbot invented a bereavement refund policy. Air Canada argued that the chatbot was a separate legal entity responsible for its own words. The tribunal rejected that submission, found the airline had not taken reasonable care to ensure the chatbot was accurate, and awarded $812.02 in damages. The amount was small. The lesson is that if an AI tells a customer something, your business has said it.

Similar failures show up in very different settings. A Chevrolet dealership’s website bot agreed to sell a Tahoe for one dollar and described the offer as legally binding. DPD disabled its chatbot after it swore at a customer and wrote a poem criticizing the company. In BC, a lawyer was ordered to personally pay costs after filing two cases that ChatGPT had invented; in 2026, an Ontario lawyer was ordered to pay $31,150 in costs for the same error. In each case, a person gave the output more authority than it deserved.

The Canadian rules already in force

You do not need to wait for a new AI statute to set boundaries. The federal and provincial privacy regulators’ principles for generative AI say accountability for decisions remains with the organization and people should be able to understand a decision and request human review. Quebec’s Law 25 has specific notice and human-review requirements for exclusively automated decisions. The Competition Bureau treats exaggerated AI claims and deceptive AI-generated content as deceptive marketing. The federal voluntary code for generative AI includes human oversight and monitoring, while the NIST AI Risk Management Framework is a common international reference. The three tiers help make ownership visible, but they do not replace checking the rules that apply to your business.

In the loop versus on the loop

Two phrases get used interchangeably and mean different things. Human in the loop means a person confirms each action before it happens: the tier-two signature. Human on the loop means the AI acts while a person monitors, reviews samples, and can intervene. That may suit tightly bounded, reversible internal work. Tier three is neither; the person makes the decision. I would not treat a sample review of sent customer replies as equivalent to reading them before they go.

A one-page policy to adapt and review

• The AI may draft, summarize, classify, research, and flag. Its output lands in a queue a person sees.
• AI-generated or consequential customer outputs, money decisions, and commitments to a date or price need a named person’s approval. Keep a record where the action warrants one.
• A pre-approved fixed transactional message may run automatically only under a bounded rule with a verified trigger, correct data, and applicable consent. Review that rule when the workflow changes.
• Decisions about people, pricing strategy, refunds above a set threshold, and anything legal, medical, or financial are not delegated, and the AI does not draft them.
• Customers are told when they are talking to an AI, at the start, every time. On the phone, in chat, in email signatures.
• Customer personal information goes only into tools the business has reviewed and listed. Nothing else, including personal chatbot accounts.
• Every AI step has an owner who reviews a sample monthly and can turn it off in an hour. If nobody owns it, it does not run.

Write the rule plainly: the AI can prepare the work; a person makes the decision. If you cannot say who decides, the workflow needs more work before AI joins it.

What to do Monday

List every place AI touches your business today, including the informal ones. Put each on a tier. Anything acting alone that belongs in tier two gets a queue and an approve button this week; anything in tier three gets switched off. Then write the six lines above into your operations doc with an owner’s name beside each AI step. If you want to see whether your workflows are clear enough for the tiers to hold, the AI Readiness Check scores AI usefulness, including whether there are rules for what AI may and may not do. The automation vs AI guide covers how we decide which steps get rules, AI, or a person in the first place.

License the drafting.Keep the keys.

Common questions

What is a human in the loop approach to AI?

Human in the loop means a person reviews and confirms what an AI proposes before it takes effect, rather than the AI acting on its own. In practice: the AI drafts a reply, quote, or classification, a named person approves or edits it, and only then does the system act. It differs from human on the loop, where the AI acts and a person monitors afterwards, which is only appropriate for reversible, low-stakes steps.

What are some good AI policies for small businesses?

A useful small business AI policy answers three questions rather than listing banned tools: what the AI may do alone, what needs a person’s approval, and what is never delegated. Add disclosure to customers, a list of approved tools for personal information, and a named owner for each AI step. Those points make a useful staff summary. They do not cover every privacy, employment, security, or sector requirement; review the full policy for your business before adopting it.

How do I create an AI policy for my company?

List every place AI touches your business, including informal use. Sort each into three tiers: licensed to act alone, needs a signature, never delegated. Write one rule per tier, add a disclosure rule and an approved-tools rule, and name an owner for each AI step who reviews a sample monthly. Review it whenever a new AI step is added and whenever the workflow changes.

Can you give me an example of AI governance?

A quoting workflow where the AI drafts the quote from the price book and past jobs, the draft lands in a queue, the owner reads every number and taps approve, the approval is logged with a name and time, and the customer’s email footer states that quotes are prepared with AI assistance and reviewed by a person. That is governance: decision rights, a record, and disclosure, with no committee involved.

What are the four pillars of AI governance?

Frameworks vary, but most enterprise versions name accountability, transparency, fairness, and safety or security, and the NIST AI Risk Management Framework organizes the work as govern, map, measure, and manage. For a small business the same ideas reduce to who decides, who is told, what data is allowed, and who can switch it off. The three tiers help clarify who decides. Disclosure, tool review, and an accountable owner are further starting points; the level of risk determines what else you need.

See how we build AI-assisted workflows with clear review points.