Dynamic Pricing With AI: What It Can and Cannot Do

Pricing software that used to belong to airlines is now a plugin. A restaurant can vary covers by night, a landscaper can price by season and route density, a shop can clear stock on a schedule nobody has to remember. The capability arrived faster than the judgment about how to use it, and regulators have started to notice: China's Supreme People's Court published guidance this week, reported by Xinhua, stating that businesses using algorithms to charge different prices for the same thing without reasonable justification can be liable where consumers are harmed.

The three kinds, and why the distinction matters

The first two vary the price of a circumstance. The third varies the price of a person, and customers read that as being charged what you think they will tolerate. Two people comparing receipts is all it takes, and in a small market they will compare.

Where it earns money in a small business

Perishable capacity. An empty table, an unbooked hotel room, an idle van on a Tuesday. That capacity expires worthless, so a lower price to fill it is pure gain and a higher price at peak is the same logic in reverse. Hospitality, accommodation, tickets, studio and clinic time all sit here.

Seasonal trades. Landscaping, snow clearing, HVAC, pool servicing. Demand swings by a factor of five across the year and most operators still quote one number. Pricing the shoulder season deliberately moves work into weeks that were empty.

Stock with a shelf life. Anything dated, seasonal, or about to be superseded. A markdown schedule that starts earlier and moves in smaller steps usually beats one large clearance at the end.

Job-cost variability. A callout twenty minutes away and a callout ninety minutes away are not the same job. Pricing for travel, access difficulty and urgency is dynamic pricing that customers find entirely reasonable, and it is the version most trades under-use. Our note on AI quote generation covers building that into the quote itself.

Where it backfires

Regulars. Your best customers are the ones who buy often enough to notice the price moving. A scheme that maximises revenue from strangers can quietly annoy the twenty people who keep the lights on in February.

Screenshots. Any price a customer finds unfair travels further and lasts longer than the margin gained. The reputational cost is real, hard to quantify in advance, and lands on a small business harder than a national chain.

B2B relationships. Business buyers compare notes at industry events and expect a rate card. Surge pricing a commercial client who has bought from you for six years reads as opportunism whatever the model says.

Emergencies. Raising prices during a storm, a flood, or a breakdown is legally constrained in some places and reputationally constrained everywhere. Set a ceiling that holds during exactly the conditions when a model would want to push through it.

The legal direction of travel

The Chinese guidance is not Canadian law and does not bind anyone here. It matters as a signal, because it is one of the first times a senior court has put algorithmic price discrimination alongside deepfakes and infringing content in a single set of liability rules. Regulators tend to converge on this kind of question over a few years rather than in isolation.

In Canada, the settled enforcement risk sits in how prices are presented rather than in whether they vary. Drip pricing, where mandatory fees only appear at checkout, has been an active area for the Competition Bureau. Anything that keys off a protected characteristic is a separate and more serious problem, and can be produced accidentally by a model using a proxy such as postal code, which is the mechanism we described in AI bias in a small business. None of this is legal advice, and a customer-based scheme deserves a lawyer's eye before launch rather than after a complaint.

How to run it without wrecking trust

1. Vary circumstances, not people. Everyone booking a Saturday at 7pm sees the same number. This single rule removes most of the legal risk and nearly all of the reputational risk while keeping most of the revenue gain.

2. Publish the logic. "Peak pricing applies Friday and Saturday evenings and on long weekends" converts a suspicious price into an understood one. It also stops staff having to improvise an explanation at the counter.

3. Set a floor and a ceiling by hand. Bounds are the control that survives a model behaving oddly, and they are the first thing missing in most implementations. A pricing system optimises the number you gave it, which is the general failure we set out in your AI will optimise for the score you set.

4. Start with rules, not a model. If a spreadsheet could express your pricing logic, use rules first. They are explainable to a customer, auditable by you, and they establish a baseline that any later model has to beat.

5. Review the outliers weekly. Look at the ten highest and ten lowest prices the system produced. That is fifteen minutes and it catches the problem before a customer does.

Where to start

Export a year of sales with dates, times and quantities, and ask a general AI assistant to find the patterns: which days sell out, which are consistently soft, how far ahead people book, what a slow Tuesday actually costs you. That analysis used to be a consulting project and is now an afternoon. The related forecasting work is covered in demand forecasting with AI.

Then change one price band and hold everything else steady for a month. Compare revenue and unit volume against the same month last year, not against your expectation. Most businesses find the gain comes from filling the soft periods rather than from charging more at peak, which is a better outcome anyway because nobody complains about it. Our note on the pricing gap covers the margin side of the same question.

Frequently Asked Questions