How to measure the ROI of an AI assistant: the recovered-hours method

“Our clients gain 30% in productivity.” That sentence means nothing: 30% of what, measured how, on which tasks, at what adoption rate? The gains from an AI assistant are real, the research shows it, but they are local: they depend on your tasks, your volumes and your people. Somebody else’s percentage does not predict yours.

Our rule, set at the scoping of every engagement: every gain is measured at the client’s site, before and after, on indicators chosen in advance. This article gives the full method, the one we apply, so that you can demand it of any vendor, ourselves included.

ROI is a comparison. If you do not know how long a quote, a procedure search or a meeting summary takes today, you will never know what the assistant changed: you will be comparing impressions with memories, and impressions are forgiving.

The baseline does not have to be a project in itself. Two weeks is enough: pick the three tasks the assistant will target, ask the people involved to note the time they spend (a simple shared sheet will do), and record three or four data points per task: frequency, average duration, people involved, and the rework or error rate where it can be counted. That small inventory serves twice: for the ROI, and for prioritizing the use cases, because the best target is rarely the one you assumed.

The heart of the calculation fits on one line: hours recovered per period, multiplied by the loaded hourly cost (salary plus benefits and payroll costs) of the people recovering them. To that you add, where they can be measured cleanly, the value of errors avoided (rework, penalties, scrap) and the value of speed that turns into revenue (a quote filed on time is a quote that can win).

Here is what the calculation looks like on a hypothetical scenario, presented as such: a procedures and document search assistant at a 40-employee SME, with 20 regular users.

Published reference points help frame expectations, provided you quote them for what they are: measurements taken elsewhere. In a controlled experiment published in Science, Noy and Zhang measured that with a generative AI assistant, professionals completed writing tasks roughly 40% faster, with quality judged higher. In customer service, the study by Brynjolfsson, Li and Raymond, published in the Quarterly Journal of Economics in 2025 across 5,172 agents, measured 15% more issues resolved per hour on average, and roughly 30% among the least experienced agents.

Two cautious readings are called for. These studies cover specific tasks, not whole jobs: a 40% gain on writing does not make a company 40% more productive. And the effect varies with people’s experience: the assistant levels upward, it does not replace expertise. That is consistent with what BDC reports on its side: according to its research, Canadian SMEs that use AI are roughly 24% more productive than the rest, a correlation that invites action, not an individual guarantee.

Two further reference points complete the picture. McKinsey estimated, in what has become a classic study (2012), that information workers spent 19% of their week, nearly a full day, searching for and gathering information: if your people recognize themselves in that, document search is probably your first profitable use case. And Statistics Canada published an analysis in April 2026 that sets the record straight: the raw productivity gap between AI-using firms and the rest (16.8%) shrinks to 5.1% once complementary investments in cloud, R&D and training are taken into account. In other words, AI alone is not enough; it is the complete project, data and training included, that produces the gain.

An AI assistant produces real benefits that resist being priced: an expert’s knowledge that no longer retires with them, the employee who stops interrupting three colleagues to find a procedure, the new hire who becomes self-sufficient in weeks rather than months, the compliance risk that recedes.

Our advice runs against the instinct of spreadsheets: name those gains, track them qualitatively, but do not convert them into dollars. Every forced conversion (“knowledge retention is worth $80,000”) weakens the whole calculation in a board’s eyes. A sober, defensible ROI convinces more than an inflated, contestable one.

ROI is calculated against the full costs, not against the first invoice: implementation, hardware (or rental), annual upkeep (in the order of 10% to 20% of the implementation cost), training, and your own people’s time during the rollout, which is free to nobody.

On the other side, two reliefs can be calculated too: public support, which reduces the capital invested (our Funding page lists the programs), and, for a local solution, the absence of per-user licences that would otherwise climb with adoption. The irony is worth noting: the more a per-seat cloud assistant succeeds, the more it costs; the more a local assistant succeeds, the more each recovered hour costs the vendor next door.

The overall ROI of “an AI project” is an abstraction. What can actually be steered is one indicator per use case, chosen before deployment and reviewed at every wave. Here are the pairings we use most often.

This week: choose the three candidate tasks, start the baseline (two weeks, a shared sheet), and settle each use case’s indicator with the people who will do the work. Before any commitment: require your vendor (ourselves included) to put the before-and-after measurement and the decision points in the proposal, in black and white.

If you want to hold your three candidate tasks up against what we have seen elsewhere, the scoping call is there for exactly that. And if the arithmetic shows your project does not pay for itself, we will tell you: an engagement that disappoints costs our reputation more than one we turn down.