Human review means checking an AI output against its source evidence before the firm relies on it. The person doing the check looks for unsupported claims and fixes or rejects them. An approval button can record who accepted a draft, but it cannot show what that person checked.
Marion Clegg asks her adviser what raising her drawdown income to £2,000 a month would do to her pot. The AI's meeting notes say she agreed to the increase. The adviser approves the notes after a quick read, and a letter based on them repeats the claim. Marion complains. When the firm plays the relevant part of the recording, it hears a question, not an agreement.
The recording was the source for the notes. Checking the claim against it before using the notes would have caught the error. A suitability report needs a different check against the fact-find, statements and other evidence it draws on. The person reviewing it must be able to follow its claims back to those sources. Reading a fluent draft or clicking approve does not do that work.
Not to be confused with
- Review, in this entry, means checking the AI's work. It is not the annual review with the client.
- A file check may happen later or on a sample. Human review happens before the firm relies on a particular AI output.
- Clicking approve may record who accepted the work. It does not show that anyone checked the source.
Overheard
Compliance officerMarion's notes say she agreed, but the recording has only her question. We approved the notes without checking that claim.
AdviserThen I'll check the notes against the recording before I approve, starting with anything a client agreed to.
Where it sits
Human review
- guards againstHallucinationWrong-subject claimTranscription error
- works withAI-enabledHarnessAgentPermissionSource
Grey words are still being written.
Sources
Last reviewed
- 1Complacency and Bias in Human Use of Automation: An Attentional IntegrationParasuraman and Manzey, Human Factors ·
- 2Automation bias: a systematic review of frequency, effect mediators, and mitigatorsGoddard, Roudsari and Wyatt, Journal of the American Medical Informatics Association ·
- Status
- Reviewed
- Named reviewer
- None recorded
