What it is
Your plan for AI probably counts the time a task will no longer take. Count the new work AI brings as well:
- Checking. Someone reads the output before it is used.
- Logging. Someone records what was generated and who approved it.
- Handling exceptions. Unusual cases have to reach a person, and someone has to decide them.
- Owning the process. Someone sets where AI sits in the process and what the rules are.
- Looking after the tools. Someone maintains the tools, prompts, settings and links to other systems.
Each of these needs time in someone’s week and a name next to it.
What the evidence shows
The problems this practice deals with have been studied:
- AI output needs checking. On a task the AI did badly, consultants using AI were 19 percentage points less likely to reach the right answer. Someone has to catch those cases.
- Checking becomes the job. In a survey of 319 knowledge workers, the critical work that remained moved to checking, combining and overseeing AI output. Among teams that build AI agents, 74% judge their agents mainly by having people review them, according to a study of agents in production.
- A person plus AI is not better by default. Across 106 experiments, a person working with AI did worse on average than the better of the two working alone. Checking takes skill and time.
- When nobody owns the checking, skills wear down. At one accounting firm, staff stayed answerable for automated work they no longer fully understood.
- The cost can land on the person who receives the work. In a survey of 1,150 US desk workers by BetterUp Labs, which sells coaching, and Stanford researchers, 40% said they had received AI output in the past month that looked finished but was not. The receiver, often in another team, has to fix it. The survey is a signal: how respondents were recruited is not described.
- Cutting before counting. Commonwealth Bank of Australia announced 45 roles would go to an AI voice bot. The union said call volumes were rising, with overtime offered and team leaders answering calls. The bank later said the roles were not redundant.
- Savings can vanish. In Denmark, widespread use of AI chatbots left no trace in earnings or recorded hours two years on.
How to do it
- For each task you plan to hand to AI, list what a person will still do: check, approve, log, handle exceptions.
- Estimate the time for each, per week. Ask the people who will do it.
- Subtract that from the expected saving. Plan on the net figure.
- Name a process owner for each process AI touches. The process owner sets the rules and decides exceptions.
- Name one owner for the AI tools themselves. See name one accountable AI owner.
- Write the new duties into job descriptions and workloads. Do not leave them as favours.
- Review after the pilot. Checking time often falls as the tool earns trust. Sometimes it rises.
What to measure
- Time spent checking and correcting AI output, per task.
- The share of cases that become exceptions.
- Net time saved: the time saved minus the time added.
Where it fails
- Counting only the saving before checking time. The business case looks better than the result.
- Nobody owns the exceptions. Each team handles them differently and the number grows.
- Checking becomes a formality. People approve without reading. Sample the approvals from time to time.
- IT becomes the queue. With nobody looking after the tools, every change waits for IT.
By company size
- Micro and small: one person may hold all the roles. Still write down the hours.
- Medium: one owner for the tools, and a named owner for each process AI touches.
- Mid-market and enterprise: a small team that looks after the AI tools, and process owners in the business.
See it applied
Company cases
What happened without it
- Air Canada Keeping a chatbot’s answers in line with current policy is ongoing work. Here it was not done, and a dispute over a few hundred dollars ran for months.
- Butler Snow Five unchecked citations led to a review of more than 2,400 citations in 330 filings by 28 outside lawyers, paid for by the firm.
- Deloitte (Australia) Review before delivery did not catch the wrong citations. The report had to be corrected twice.
- Levidow, Levidow & Oberman P.C. The made-up decisions were filed without being checked. The court said the duty to check stayed with the lawyers.
Sources
- Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks and Wilson (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 1121. What this study did and found
- Pan, Arabzadeh, Cogo and others (2026). Measuring agents in production. ICML 2026. arXiv 2512.04123. arxiv.org
- Niederhoffer, Kellerman, Lee, Liebscher, Rapuano and Hancock (September 2025). AI-generated “workslop” is destroying productivity. Harvard Business Review. BetterUp Labs, one of the authors’ employers, sells coaching. hbr.org
Related practices
- Name one accountable AI owner Give one named senior person the responsibility for AI: priorities, budget, risk and results. Shared responsibility tends to mean nobody decides. Do first · Effort: low (days, one person)
- Measure time saved per task against a baseline Record how long a task takes and how good the result is before you add AI. Without that starting figure you cannot tell a real gain from a hopeful one. Do first · Effort: low (days, one person)
- Redesign the workflow instead of adding a tool to it Giving people an AI tool inside an unchanged process changes little. Decide again who does each step, in what order, and where a person checks. Do next · Effort: high (months, several teams)
- Budget for people and process, not just tools Licences are the small part of the cost. Plan most of the time and money for training, redesigning work, and helping people through the change. Do first · Effort: low (days, one person)