About narration
Uses speech voices provided by your browser or device. Availability, quality and online processing depend on that service.
Start with a task that can be checked.
A procurement team considering AI has many possible starting points. Drafting a routine supplier email, extracting dates from a contract and deciding whether a bidder is eligible carry different consequences. Putting them under a single productivity target hides the difference.
Choose a bounded task with a known input and a result that someone can verify. Describe what the tool may produce, what it may access and what it must never decide. A draft should remain a draft until the responsible employee approves it. This is particularly important when the result affects payment, employment or a legal obligation.
Exposure does not measure a company’s savings.
The ILO's May 2025 study assesses occupational exposure to generative AI using task-level analysis. It describes potential changes to work, not a forecast of redundancies at a particular employer. [1]
For an employer, the practical question is what happens to an entire process. If an assistant drafts a report faster but a manager spends longer correcting it, the apparent saving can disappear. Measure preparation, checking, correction and follow-up together. Include the time needed to find the original evidence when a claim is challenged. Avoid translating a global exposure estimate into a local headcount target.
Give the reviewer a workable job.
Consider an illustrative pilot that summarises maintenance reports for a site manager. The reviewer needs access to the original reports, a clear way to flag omissions and authority to reject the summary. A fluent paragraph can conceal a missed safety issue. Checking only a sample of low-risk fields would not prove that critical warnings are preserved.
Define error categories before the pilot starts. A wrong date, an invented action and a missing warning should not receive the same weight. Set an acceptance threshold for the intended use, record failures and retain a manual route. In multilingual operations, test the languages people actually use, including the terms used on site.
Set boundaries for information and decisions.
NIST's 2024 generative AI profile identifies risks including confidently incorrect content and data privacy, and provides voluntary risk-management guidance. It is a useful technical reference, not a substitute for the rules applying to a South African or Congolese employer. [2]
Agree which information may enter the selected service, who can access the outputs and how long records are retained. Use synthetic examples for early tests where possible. Keep applicant records, employee grievances and confidential contracts out of unapproved tools. A reviewer cannot repair an unauthorised disclosure after it has occurred.
Train people for the changed process.
A pilot creates new work: checking sources, identifying omissions, reporting failures and deciding when to stop using the tool. Give these tasks to named people and allocate time for them. Explain to staff what the pilot is measuring and how results will be used. Involve the people doing the work when redesigning the process.
Before expanding, compare the pilot with the previous process on accuracy, total time and unresolved errors. Keep a record of the version tested and repeat the checks when the tool or task changes. Expand only where the evidence supports a useful result. If checking costs more than the task saves, a simpler workflow may be the better investment.
A question for your team
Who can recognise a convincing but wrong output, and do they have time to check it?
Sources & notes
Sources checked on 29 October 2026. This article presents our analysis of the cited research and its implications for businesses. Numbered references link to the original sources.
