Browse
Log in / Register
Share
April 27, 2025
Artificial intelligence is sold as a way to claw back hours in a workweek. But a new study out of Canada shows that the time AI gives back is rarely clean time. Instead, employees are spending part of that gain checking AI’s work and cleaning up after it.
The report, The State of AI at Work in Canada, comes from Punchcard Systems, an Edmonton-based digital innovation consultancy, with polling by Angus Reid Group. It is based on a survey of 1,100 working Canadians. A separate automotive-focused slice of the dataset looked at 50 workers in that industry.
The headline number is blunt: 94% of AI users say at least some AI-generated output needs meaningful correction or has to be partly redone before it can be used. In other words, almost nobody thinks AI output is ready to go untouched.
That does not mean AI is failing to save time. About two-thirds of users say the technology does save them time, and those users estimate an average of 6.1 hours per week. The other 35% say AI saves them no time. The survey did not ask whether people counted the time spent fixing AI output in those estimates — and that omission matters more than it appears.
If workers are telling themselves that spot-checking AI work is part of the “real work,” their 6.1-hour savings number is overstated. If they are not counting correction time at all, they are understating the true cost of using AI. Either way, the gain is smaller than the headline suggests. That is the catch.
Employer support, meanwhile, looks thin. Across all industries, 64% of workers said they receive no formal AI support. Only 7% have received formal training on using AI in their daily role. Just 16% have received guidance on AI risks, privacy, security and limits. That suggests most workers are left to figure out A on their own — and to make their own mistakes.
The automotive sector stands out as a particular concern. Among 50 automotive workers surveyed, 60% said their employer provides no formal AI support. That is a small sample, so treat it as an indicator rather than proof. Even so, the finding is awkward for an industry that otherwise embraces automation.
The broader adoption numbers confirm that AI is already embedded in work life. 52% of workers said they use AI for work. That includes 32% who use general-purpose AI assistants or websites. The other users may not think of what they use as AI — meeting summaries, code suggestions, email drafting tools and customer-service bots all count.
Any U.S. employer reading this should recognize the pattern. Canadian and American workplaces are not identical, but the generative AI adoption cycle looks similar: employees start using tools on their own, get no formal training, save some time on rough drafts or research, then spend extra unreported time verifying and correcting what the model produces. The only difference may be the size of the company or the industry.
The real problem is not the AI output quality. It is visibility. When AI use goes underground, managers lose sight of how work is actually being done — and how much time it takes to do it well. A worker might produce a report in half the old time and then spend the other half fixing the AI’s invented sources, weird formatting, or inaccurate numbers. The manager sees the shiny final product, not the duct-tape effort underneath.
That does not mean AI should be banned or discouraged. It means employers need to be more deliberate. The survey offers a clear set of gaps: little training, little guidance on limits, little formal support. Closing those gaps matters more than choosing which AI tool to buy.
First, employers should audit current AI use before imposing new policies. Ask employees what they are already using and where the friction is. Most will admit to using ChatGPT or similar tools. The ones who don’t are either not using them or hiding them.
Second, training should be role-specific. Generic workshops on “what is generative AI” are less useful than sessions showing how to verify AI output in a given job. The priority should be on when to trust, when to check, and where AI tends to fail in that department.
Third, teams should build guardrails for high-stakes output. For anything going to customers, regulators or public release, there should be an explicit AI-output review step. That review is not bureaucracy; it is the new cost of doing business with a probabilistic tool.
Fourth, time savings should be measured honestly. Organizations that see “6 hours saved” as a pure win are missing the hidden correction tax. A better metric is task-cycle time: how long from start to launch, including verification. If AI shortens that, keep it. If it only shifts the work from drafting to debugging, the gain is not worth the risk.
The underlying issue is that AI does not remove human judgment. It changes where that judgment gets applied. Instead of writing a first draft from scratch, a worker now has to evaluate a generated draft, decide what to keep, fact-check claims, repair tone, and ensure the final version still sounds human. That is still work — important work, but work.
The Canadian report is a useful reminder that hours saved and hours spent are different numbers. For U.S. employers, the lesson should be obvious: don’t let the headline efficiency gains fool you. Find out where AI is actually helping, where it is quietly creating cleanup tasks, and where workers are operating without a net.
Punchcard Systems launched the report to provoke exactly that conversation. The survey was in the field Sept. 2–11, so the data reflects a period before many companies announced their latest AI layoffs or policy updates. As such, it may already understate how quickly AI is becoming normal in daily work.
One more number from the survey speaks loudly: 64% of workers still get no formal employer AI support. That is the number to fix. Until that changes, every time-saving minute AI delivers will carry the same quiet caveat — someone has to check behind it before the work is actually done.
