What AI costs, and how to judge the return

The licence is usually the smallest line. Per-seat AI tools commonly sit in the tens of dollars per person per month, while the staff hours to redesign a process, test the output, and check results in the first month typically cost several times that. Statistics Canada found 10.6 percent of businesses cited cost as a barrier limiting AI use in Q2 2026, ranking behind cybersecurity and privacy concerns at 13.4 percent.

Four recur: the staff time to define and document the process, the checking time in the first month while you learn the error rate, the ongoing verification that never fully disappears, and the cost of switching if you later change tools. The checking time is the one most often left out of a business case, and it is the one that continues indefinitely rather than ending at launch.

Take the hours the task consumed before, subtract the hours it consumes now including checking, multiply the difference by a loaded hourly rate covering salary plus employment costs, and compare that to the total spend including staff time. The subtraction only works if you measured the before, which is why the baseline measurement matters more than any other single step.

For a narrow task using an off-the-shelf tool, a few months is a reasonable expectation, because the spend is small and the time saving starts immediately. Projects requiring system integration or data cleanup take considerably longer and carry more risk of returning nothing at all. Be sceptical of any payback claim that does not state which of those two shapes it describes.

No independent survey of Canadian AI consulting rates could be found while researching this page in September 2026. Every published rate card located was written by a firm selling the service, which makes it marketing rather than research, so no benchmark is quoted here. Ask any consultant for their rate directly and compare against what the work would cost you internally.

The AI Marketing Boost package on this site is published at a one-time setup fee of 1,500 USD plus 1,400 USD per month, month to month with no annual commitment, as listed on the AI Boost page. That figure is stated here as this firm's own published price rather than as a market rate. Other providers price differently and comparing several quotes is sensible.

Fixed price suits a first engagement because it caps your exposure while both sides are still learning whether the work pays. Hourly suits genuinely exploratory work where the scope cannot be defined honestly in advance. A retainer makes sense once something is running and needs maintenance. Starting on an open-ended retainer before the first workflow proves out puts the risk entirely on you.

Often yes, for a first test. Free tiers of mainstream assistants are enough to find out whether a task is a realistic candidate before any money is committed. The caution is privacy: free consumer tiers are the wrong place for anything containing personal information, and that restriction matters more than the feature differences between tiers.

Budget an amount you can lose without regret, because a meaningful share of first projects return nothing. MIT found 95 percent of generative AI pilots produced no measurable profit impact. Sizing the first attempt as a genuine experiment, with a stopping date, keeps a failed test cheap and preserves the budget for the second attempt, which benefits from what the first one taught you.

Two pressures push in opposite directions. Per-unit model pricing has generally fallen, while usage tends to rise as more people find uses, and vendors move capability into higher tiers. Assume your bill grows with adoption rather than shrinking with prices, and review subscriptions quarterly. Unused seats are the most common form of waste in a small company's software spend.

Lead with the current cost of the task in hours and dollars, name the single task rather than a strategy, state what you will spend, and give the date when you stop if the numbers do not move. Scepticism is usually a reaction to open-ended commitments rather than to AI. A capped experiment with a defined exit is far easier to approve.

That is the normal outcome, and it is worth understanding rather than hiding. Two hours saved across five people does not become a payroll reduction; it becomes capacity. The gain is real only if that capacity goes somewhere you can name, such as more jobs quoted or faster response times. Decide in advance where the recovered hours will go, or the saving stays theoretical.