Skip to content

93% use AI. 2% report a return.

Leaders read "we use AI" as "it pays". They are different claims, and the gap is measurement and ownership, not the model. Fix both before the 2027 budget round.

In brief

  • KPMG Canada's November 2025 survey of 753 business leaders found 93% said they use AI, yet only 2% reported a return on their generative AI investment.
  • Stronger firms adopt AI first, so without a baseline and a comparison group, AI collects the credit for their strength while saved hours dissolve unallocated.
  • Treat every material AI workload as a value hypothesis with a named business owner, a pre-rollout baseline, a comparison group and a scale-or-stop date.

Two per cent of the 753 Canadian business leaders KPMG Canada surveyed in November 2025 reported a return on their generative AI investment. Ninety-three per cent said they use AI.

KPMG asked again six months later. This time 70% of Canadian respondents said AI delivers meaningful value. Only 3% had achieved measurable returns. Inside many firms the pattern is familiar: licences active, usage climbing, and no number that survives finance's second question.

Adoption is not evidence of value. We treat every material AI workload as a value hypothesis: a named business owner, a baseline taken before rollout, a comparison group and a date by which it scales or stops.

Adopters were already stronger, and AI took the credit

Three explanations account for the gap, and they reinforce one another. The first is selection. Researchers Jiang Li of ISED and Huju Liu of Statistics Canada found AI adopters were 16.8% more productive than non-adopters after basic controls. Once they accounted for pre-adoption productivity and complementary investments such as R&D, cloud, data analytics and staff training, the premium fell to 5.1% and was no longer statistically significant.

Firms using data analytics were 15 percentage points more likely to adopt AI at all. Put plainly, strong firms adopt AI, and AI collects credit for their strength.

The second is measurement. KPMG noted that many firms already assess returns with metrics such as productivity gains and scale of adoption. Scale of adoption is an input. When it sits in the return column, the survey is partly measuring enthusiasm. The third explanation, time, is the objection most leaders raise, and we return to it below.

Together they point to one practical conclusion. Returns arrive when AI is paired with workflow change and the data behind it. They are visible only to organisations that set up measurement before rollout.

McKinsey's 2026 State of AI survey, of 1,719 respondents in 97 countries, shows what the exceptions do. High performers, who attribute at least 5% of EBIT (earnings before interest and taxes) to AI, have stayed at about 6% of respondents for a second year.

Nearly three-quarters of those high performers fundamentally redesigned workflows, against about a quarter of everyone else. They were also twice as likely to have defined processes for measuring the impact of AI initiatives. Both are management choices, open to any firm.

Saved minutes vanish unless someone decides where they go

The most common pattern is real individual time savings that never reach the income statement. ATB Financial's Tara Lockyer described it in June 2026: huge individual productivity uptake, and real difficulty harnessing it at team or line-of-business level. She said ATB had not solved it yet. That is an Alberta Crown corporation saying publicly what many firms find privately.

Saved minutes are not saved dollars. Time freed for an analyst becomes value only if something specific happens next. The team absorbs more volume without hiring, external spend falls, or the work shifts to something with a measurable payoff. If nobody decides which, the minutes dissolve into email. In our own builds, the conversation about where freed capacity goes is often harder than the build, and it is the one that decides whether a return appears.

Unsettled definitions, not messy schemas, keep the data unready

Data is the other place value leaks. "The data is not ready" usually means something more awkward than messy schemas. Two departments define a customer, a job or a completed order differently, and nobody has the authority to settle it. An AI tool built on top of that dispute produces answers each side can reject.

The StatCan link between data analytics and adoption is a reminder that the hard part sits upstream of the model. That work is organisational: someone has to own each definition that a workflow depends on.

Patience buys an exploration budget, not an exemption from evidence

The obvious objection is time. KPMG Canada's Stephanie Terrill argues that new technologies take time to show identifiable returns, and most respondents expected returns within one to five years. That is plausible. General-purpose technologies have often taken years to appear in productivity figures.

A related objection holds that early measurement kills options. Some uses are exploratory, closer to R&D than to process improvement. Forcing a payback model on them in month two is how organisations end up with only safe, small projects. BDC's February 2026 survey of 1,500 SMEs found 78% of AI users satisfied with their return, which suggests many owners feel value they have not formally measured.

We accept part of this. A bounded exploration budget, explicitly exempt from return tests, is healthy. But BDC's own data argues for structure. Satisfaction was 85% among firms with a concrete AI plan, against 66% without one.

It was 86% among firms that trained staff on AI, against 53% among those that did not. Satisfaction is not return, yet the firms that planned were the ones most pleased with the result.

Patience is also running short. In a vendor-commissioned Harris Poll for Dataiku, 72% of 685 CIOs in eight countries said their AI budget was likely to be cut or frozen if targets were not met by the end of 2026. If those CIOs act on it, exploration without evidence will be the first spending to go.

No national benchmark will arrive before 2027

National statistics will not fill the gap. Statistics Canada measured business AI use at 19.2% in the second quarter of 2026, up from 6.1% two years earlier.

Agriculture and forestry sat at 4.5%, and construction at 9.2%. But the survey behind those figures has been discontinued, and its successor is not due to publish business data until 2027. Western Canadian firms in resources, construction and agriculture will have no outside benchmark for some time. They need their own.

Seven steps put a defensible number on one workload

Here is the sequence we recommend for every AI workload above a trivial size.

  1. Write the value hypothesis in money. Name the profit-and-loss (P&L) line that should move: cost per claim, days sales outstanding, contractor spend, overtime hours. If you cannot name the line, it is an exploration project and belongs in that budget.
  2. Take the baseline first. Measure the current workflow for four to six weeks before rollout: cycle time, cost per unit of work, error or rework rate. A baseline reconstructed afterwards is an argument, not a measurement.
  3. Name a business owner. The person who owns the P&L line owns the metric. The AI or IT team owns delivery, not the verdict.
  4. Decide where freed capacity goes. Absorb volume, cut external spend or redeploy people to named work. Write it down before go-live.
  5. Build a comparison. Stagger the rollout by team or region so some groups act as a control for a few months. It is the cheapest way to separate the tool's effect from the selection effect StatCan found.
  6. Fix the data the workflow touches. Data analytics was one of the complements that absorbed the premium in the StatCan study, but you do not need an enterprise programme first. Settle the definitions and access for this workflow, and give them an owner.
  7. Set the scale-or-stop date. On that date, the owner presents the result against the baseline, and the workload scales, changes or stops.

One workflow fits on one page

A worked example shows how little paperwork this needs. Take AI-assisted invoice matching in accounts payable. The hypothesis is that fewer exceptions need manual handling, so cost per invoice falls. The controller owns it. The baseline is exceptions per thousand invoices and handling minutes per exception, measured for a month. Freed capacity absorbs volume growth instead of backfilling the next vacancy. One region goes first, another waits a quarter. At the end of that quarter, the controller either shows the gap or stops the work.

Report results in three tiers and keep them separate.

Only the financial tier of AI measurement can claim a return
TierExample measuresWhat it can claim
ActivityLicences active, prompts per userPeople are trying it
OperationalCycle time, rework rate, cases per full-time equivalent (FTE)The workflow changed
FinancialCost per case, external spend, revenue per repThe business gained

Source: Wennovate framework.

The first two tiers are useful leading indicators, provided they are labelled as such.

Put it to work: Three questions for your next AI budget review

  1. Which of your AI workloads has a baseline taken before rollout, and which are reported on activity alone?
  2. For each material workload, does the person who owns the P&L line sign off the result, or does the AI team?
  3. Where did last quarter's saved hours go: absorbed volume, lower external spend or named redeployment?

The 2% judges how firms count, not what AI can do

The 2% figure is not a verdict on AI. It is a verdict on how most organisations have chosen to deploy and count it, and the costliest mistakes are counting mistakes. Reporting licences and prompts as a return is the first; how usage targets distort behaviour is the subject of Part 2 of this series. Letting the AI team grade its own work is the second, because the verdict belongs to whoever owns the P&L line.

Counting hours saved without saying where they went is the third, since unallocated time is not a saving. The national 60% adoption target is no benchmark either: an adoption rate says nothing about whether a given firm's adoption pays. And waiting for an enterprise data platform before measuring anything only delays the one workflow that could have produced evidence this year.

Before the next budget round, decide which workloads you will put a baseline, an owner and a stop date on, and which ones you are funding on faith.

Next in this series: Mandate the outcome and let teams choose the tool

Running AI across your teams? Book a 30-minute discovery call with the Wennovate team. No pitch deck required. Start with a scoping call

The Studio Letter

One letter a month from the CEO: what moved, what stalled, what we learned.

Get the Studio Letter