AI Bias in a Small Business: Where It Shows Up

Almost everything written about AI bias is about hiring algorithms at large corporations, which makes it easy to file as somebody else's problem. If you have thirty staff and no screening software, it sounds like a topic for a different kind of company. It is not, and the version that affects you is quieter, closer to the ground, and considerably easier to check for than the coverage suggests.

Where it actually turns up

The third row deserves attention in a country where a large share of customers speak English or French as a second language. A transcription tool that handles some accents worse produces shorter, thinner summaries for those calls, and that difference flows into every record built from them without anyone deciding it should.

The hour-long check

You do not need a data scientist. You need the same input twice with one thing changed.

Take a real customer enquiry. Swap the name for one from a different background. Run both through whatever drafts your replies. Put the two outputs side by side.

Do that with five or six examples across the situations your business handles. Look for differences in warmth, length, formality, and how much effort the reply puts into being helpful. If those vary and nothing but the name changed, you have found the problem and it took an hour.

The legal part is more concrete than you think

People treat this as an ethics conversation, which it is, but the harder edge is legal and it already applies to you.

Human rights legislation governs how you treat customers and employees regardless of what produced the decision. A system generating the outcome is not a defence, in the same way that a template producing it would not be. Employment standards and accessibility rules work the same way.

So AI has not created a new legal category. It has created a way to produce outcomes at speed and volume in areas where you were already responsible, which is why hiring is the sharpest case and why we treated it separately in the risk in AI resume screening.

Why switching tools rarely helps

The instinct on finding a problem is to conclude you picked a bad model and go looking for a better one.

That usually disappoints, because the underlying patterns are broadly similar across the major models. They learned from overlapping material and inherit overlapping tendencies. You may move the problem rather than remove it, and now you have also paid a switching cost.

What actually works is changing the process. Add a review step for the category where the difference appeared. Give the tool a fixed template so structure cannot vary by recipient. Or stop using it for that specific task while keeping it everywhere else. That is the same consequence-based approach as practical AI risk management: bound the thing that matters rather than searching for a perfect tool.

Ask the tool about itself

A cheap addition to the check. When something produces a draft, ask what assumptions it made about the person.

You will occasionally get a revealing answer: it assumed a level of technical knowledge, a formality register, or a familiarity with your industry that nothing in the enquiry supported. Those assumptions are where differential treatment starts, and they are invisible in the finished text. It is the same technique as asking AI what it is likely to get wrong, pointed at a specific question.

Do it once a year

This is not a permanent project. Run the paired test when you adopt something new, and once a year afterwards, because models change underneath you without notice and a check from eighteen months ago describes a system that no longer exists.

An hour annually, plus whatever you fix. That is a proportionate response for a small business, and it is considerably more than most are doing, which means it is also a straightforward answer if a customer or a regulator ever asks what you have done about it.

Frequently Asked Questions