In partnership with
The best influencer marketing advice never comes from a report
Think about the last thing that got you fired up about your job this year. Odds are it didn't come from a report. It came from someone who does the job, saying what happened when they tried it.
So we've filled a day with people like that.
Maya Shaff, ŌURA. Leah Walker, Adobe. Georgia Humphries, Stanley 1913. Enara Roy, Halfday. Tyler Vaught, Edelman. Josh Rangel, Ogilvy. Sarah Whittle, ex-Crocs and Duolingo.
And that's just the first wave of speakers.
Return on Influence Festival '26 is a free, virtual day dedicated to influencer marketing. Like any good festival, there are stages. But no mud, no queues, and absolutely no porta-potties.
Sign up today, and the rest of the lineup lands in your inbox first.
iPrompt
DEEP DIVE · iPROMPT #156 COMPANION
7 OCTOBER 2026
Human review of AI decisions
Check the evidence, not the output
By the end you’ll have checked one AI-assisted decision and kept a record of how. A review can confirm the AI’s answer. What makes it a review is opening the evidence first.
7 MIN READ · WORKED EXAMPLE AND CHECKLIST
HUMAN REVIEW / WORKED EXAMPLE
One decision, checked end to end
Linden & Co is fictional, and so are its figures. It’s a 12-person accounting firm choosing payroll software. Before reading any proposals, the partners set four requirements: they can leave by month 12; it connects to their accounting system without a paid add-on; the supplier holds ISO 27001, which their client contracts require; and, among those that pass, the lowest two-year cost. An AI assistant read three proposals and recommended Supplier C for its flexible exit terms. Five claims carried the decision.
Claim | Came from | Check | Result |
C lets us leave after 12 months | Quoted from C’s proposal | Read the whole clause | Wrong. Notice can only be given after month 12 and runs six months. Earliest exit: month 18 |
B lets us leave after 12 months | Quoted from B’s proposal | Read the whole clause | Held. Three months’ notice, which can be given from month 9 |
A and B connect to our accounting system | General knowledge | Check each integration list | Held for both. C needs a paid connector |
B costs less than A over two years | Inference from the price sheets | Recalculate it | Held while headcount stays under 15 |
B holds ISO 27001 | General knowledge; not in B’s proposal | Ask B for the certificate | Not sure. Requested, not yet received |
The outcome. C fails the exit requirement, so the reason for the recommendation is gone. On the evidence, B clears the exit, integration and cost requirements; certification is still open. Linden chose B subject to the certificate and won’t sign until it arrives. If it doesn’t, A is next, and A’s exit clause gets the same full read.
The AI flagged none of this as uncertain. And the claim that failed was a quote, not a guess: a quote can be accurate and still leave out the condition that matters.
The record. ‘Payroll supplier, 7 October. Five claims listed; four settled against proposals, price sheets and integration lists. C’s exit misread (earliest month 18). Open: B’s ISO 27001 certificate, requested. B chosen, subject to the certificate. Decided by the managing partner.’
HUMAN REVIEW / THE STANDARD
What California counts as review
From 1 July 2027, SB 947 bars California employers from relying solely on an automated decision system to discipline or fire a worker. Where an employer primarily relies on one, it must have a person corroborate the decision using the data behind the output or other relevant supporting information, such as supervisory evaluations, personnel files, work product, peer reviews or witness interviews. If the output can’t be corroborated, or the reviewer finds it inaccurate, incomplete or misleading, it can’t be used. The worker must be told in writing, and can ask for a description of their own data the system used.
It covers workplace decisions in one state, and this isn’t legal advice. The steps below borrow its core: a check against evidence the system didn’t produce, before the decision stands.
HUMAN REVIEW / THE SETUP
Five steps for one decision
1 Choose. Pick one recurring decision where AI does part of the work and a mistake is costly or hard to undo. Write down what the decision must satisfy before you read the AI’s answer.
2 List. Run the newsletter’s corroboration sheet on the output. It names the claims the decision rests on, usually three to five, and where each came from.
3 Check. Open the evidence for every one of those claims, quoted ones included. Read the whole clause or record, not just the line the AI picked.
4 Record. One line: what was checked, against what, what changed, what’s still open and who decided. Linden’s record above is the model.
5 Spot-check the tool. Once a month, run it on a case whose facts you’ve verified independently, such as a contract you’ve read in full; a past decision isn’t an answer key just because you made it. First define a material failure: a decision-critical claim wrong, missing or stated with unwarranted confidence. Log the tool, model version, date and result. One case is a spot check, not proof of reliability. If it fails, check that kind of decision more strictly until a later spot check passes.
What counts as evidence
Evidence | Counts? | Why |
The AI’s own explanation or summary | No | It’s the output again, in different words. |
A second AI model agreeing | Not on its own | A useful flag. Pew found GPT-5.1 and Claude Opus 4.6 erred in different directions; two models can also share a blind spot. |
The original document, record or figures | Yes | Independent of the output, if you read the whole passage. |
A person who knows the case | Yes | California’s list includes peer reviews and witness interviews. |
Last month’s spot check | Supports, doesn’t replace | It tells you how the tool did on that case, not this one. |
HUMAN REVIEW / VENDORS AND LIMITS
Questions for the tool’s vendor
If the AI part of a decision comes from a product, ask these first.
Ask the vendor | Why it matters |
Can I see the data behind each output? | You can’t corroborate an output from the output alone. |
What was it tested against, and how far off was it? | Pew’s digital twins were off by 12.4 percentage points on average, with nothing in their answers to show it. |
Does it ever say ‘not sure’, and how often? | Offered ‘not sure’, Pew’s real panellists chose it about four times as often as the model. |
What do your terms say if an output is wrong? | Four AI labs wouldn’t promise New York City Council that their agents would always follow safeguards. Read who carries the risk. |
Can I export each output with who approved it? | It makes step four’s record easy to keep. |
The strongest objection
‘If I have to check everything, the AI saves me nothing.’ Right about everything, wrong about the fix. You check the claims the decision rests on, not the whole output: for Linden, five claims and about an hour with the proposals. If checking takes as long as deciding yourself, that’s worth knowing too.
THIS WEEK’S FORECAST
By 1 July 2027, no more than one of the six signatories to the White House Accord on Super Intelligence (Anthropic, Google, Meta, Nvidia, OpenAI and Elon Musk’s SpaceXAI) will have published an external auditor’s findings under it. It counts if a named external auditor’s findings are published by the company or the auditor and refer to the accord. A company’s own summary doesn’t count. Void if the accord is withdrawn before the deadline; a miss if two or more signatories publish.
YOUR MOVE
Check one decision with the five steps. Reply with one finding: the decision, the claims that mattered and what changed. Leave out names and confidential details.
The goal is a decision you can explain, not one you signed.
New to iPrompt? Every Wednesday, the week’s AI news turned into one thing you can do. Subscribe free
iPROMPT / 156 /