SAP Canada: 66% are piloting agents. 97% are not ready.

You can pilot an agent. You cannot pilot a control environment. You've probably watched a department stand up a chatbot, call it agentic and then discover nobody owns the exception when it writes the wrong record. Canadian firms are confusing the two, and Cathy Tough just said the quiet part.

Tough is Country Manager at SAP Canada. SAP and Oxford Economics put 200 Canadian director-and-above leaders on the record, part of a 2,600-person, 13-country study. Mid-market firms of 500 to 1,500 people. Enterprises above that.

Sixty-six per cent of Canadian organisations are already piloting agentic AI. Ninety-seven per cent say they are not fully prepared to deploy and govern it. Ambition is not a control environment.

The average Canadian organisation in the study expects to spend CA$39.4 million on AI this year, for a 20% return that they think nearly doubles to 38% in two years – another CA$20.2 million per company. Those dollars were collected in US dollars and converted at 1.3863 on 2 September 2026. Treat them as the respondents' forecast, not as cash in the door. Eighty-two per cent of the 200 see moderate to high potential in agents. Potential is free.

The wreckage they already have

Among organisations using agentic AI, 64% hit higher-than-expected integration effort, 54% saw reliability or quality issues as usage scaled, 48% had agents take incorrect actions and 44% got inconsistent outcomes for the same process. Seventy per cent of leaders admit (agree or are unsure) they are deploying agents faster than governance can keep up. Forty-six per cent have no human-in-the-loop on autonomous workflows.

Strategy is still a food fight. Eighteen per cent take a centralised approach. Thirty-nine per cent are department-led. Thirty-five per cent are piecemeal. Sixty-five per cent believe they are data-ready. Seventy-three per cent then admit incomplete or inconsistent data. Seventy-one per cent are reworking because the outputs are junk. That is not a model problem. That is a filing-cabinet problem with a chat interface.

Seventy-eight per cent of leaders are not convinced their upskilling is keeping pace. Shadow AI shows up at least occasionally at 56% of organisations. Fully ready for AI governance: 17% on internal skills, 16% on processes. Less than one in five. Tough's line is that the technology is ready and the challenge is getting the business ready. AI without business context creates activity without outcomes, and at worst introduces risk.

Anthony Robinson, chief technology officer at SAP Canada, offered the operating model they want on the tin: humans direct, assistants coordinate, agents execute, in a governed auditable environment with humans in the loop at the moments that matter. Fine. Most of the 200 cannot staff that sentence. Skills work that starts on the process you are about to hand an agent is how you stop the 48% "incorrect action" number becoming a regulator's exhibit.

The 18% with a centralised approach are the only ones who can even attempt Robinson's model. Department-led and piecemeal (39 and 35) will produce five agents that cannot agree on a customer record. That is the 44% inconsistent-outcomes row. Do not hire another model. Hire someone who owns the process object.

Data-ready at 65% versus incomplete data at 73% is the Canadian version of "we have a lake." Belief is not a catalogue. Seventy-one per cent reworking junk outputs is the tax. Pay it in remediation or pay it in brand. Upskilling that 78% of leaders already distrust will not fix the lake. It might stop shadow AI at 56% from being the only functioning enablement programme in the building.

What this means for L&D

Treat CA$39.4 million as a budget line they expect, and CA$20.2 million as a hope attached to a 38% return they have not earned. Tough said activity without outcomes is the failure mode. Print that next to the spend. Put 66% and 97% on one slide. Ask who owns the human in the loop. Do not brief CA$39.4 million as value. It is spend they expect.

Sources: SAP Canada News Center, 10 September 2026.