Begin with the work, not an AI verdict.
01 / Relevance
In the second quarter of 2026, 40% of surveyed Canadian employer businesses said
that AI was not relevant to the business. That result is real. It is not an
employee-use census, and it cannot tell an operator whether AI is useful, safe, or
repeatable in a specific piece of work.
The practical starting point is smaller than an entire company. Ask whether people
regularly get information wrong or struggle to remember what happened before;
whether they spend more than an hour a week moving information between places; and
whether the business had the data and warning signs to avoid a problem but failed to
connect them in time.
Those questions do not preselect AI. The honest answer might be better
documentation, search, integration, ordinary automation, process repair, or no
change at all. They locate the operating problem before a tool is asked to solve it.
Source context: Statistics Canada’s Q2 2026 analysis of business AI use and Jeremy Thomas’s July 2026 interview record.
Run it twice before you call it value.
02 / Second run
The first run of an AI workflow is a poor referendum on the idea. It is also a poor
excuse for keeping a bad idea alive. A useful question is simpler: was the second
run easier or better? Was the third?
Measure the whole loop: the time required to prepare the input, the meaningful work
completed by the system, the correction and verification burden, the exceptions it
created, the work quietly handed to another team, and the final outcome. A polished
middle does not count if the cleanup queue simply moves downstream.
Then make one deliberate change: clarify the input, add evidence, define the
acceptable output, narrow the workflow, or route a risky case to a person. Healthy
implementation shows some learning—fewer corrections, less preparation, faster
review, clearer ownership, or better handling of known exceptions. If the same
defect returns after real changes, redesign the work or stop.
Source context: Jeremy Thomas’s July 2026 interview record; complementary-capability context from Statistics Canada’s analysis of AI adoption and productivity.
Give the human reviewer a job.
03 / Oversight
“A human reviews it” is not a control plan. Who is the person? What evidence can
they see? Can they stop the action? What happens after they find a problem? Without
answers, the human is a line in a slide deck, not meaningful oversight.
The minimum is modest: a named owner accountable for the consequence and visible
evidence behind the output. Higher-consequence work may also need authority to
pause or reject, an escalation route, enough time to investigate, and a way to feed
known failures back into the workflow. The right level of control depends on
consequence, uncertainty, reversibility, and observed failure—not on a universal
checklist.
A useful reviewer does more than approve or reject. They explain why an edge case
mattered, what signal should have revealed it, and whether the system should learn,
narrow its scope, or hand that class of decision back to a person. The aim is not a
larger approval log. It is fewer recurring defects.
Source context: Jeremy Thomas’s July 2026 interview record; NIST’s guidance on human-AI interaction and the Office of the Privacy Commissioner of Canada’s generative-AI principles.
Core question
If the tool works in a demo but fails in the business, what part of the operating
system was missing?