The AI rehiring stat that changed meaning

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

Try Wispr Flow free

Fortune’s headline this week: ‘Nearly 20% of US workers replaced by AI are being asked back.’ The survey behind it says something else. 19% of the employed adults it polled in the US said they’d been replaced by AI and then asked back. A share of everyone asked had become a share of everyone replaced. Last week I asked you to trace one number. So I traced this one.

The trace. The survey belongs to SHL, which sells talent assessments: an online panel of 1,029 employed adults in the US and 1,007 in the UK, run in June and July. The figure the headline implies, the share of AI-displaced workers who were asked back, isn’t in SHL’s release.

The pay figure moved the same way. SHL says 51% ‘would want’ a substantial rise before returning. In Fortune’s headline, those workers are now demanding higher pay.

The better evidence. One panel’s self-reports are a signal. Employers’ answers are firmer. Robert Half found that 32% of US hiring managers who eliminated positions after adopting AI later reinstated the same or similar roles. The reasons they gave: knowledge or context AI couldn’t replace (40%), relationships it couldn’t replicate (39%), rising demand (39%), and more oversight and quality control than expected (38%). Gartner calls it the ‘cut fast, rehire later’ trap.

My read. This part is interpretation, not Robert Half’s finding. Those answers don’t tell you how well the tools performed. They suggest the firms cut before they knew what the roles contained. So before you hand a task to AI, list what’s in it, including the parts that sit in no procedure. The prompt below does that in ten minutes.

A fictional lettings agency runs the whole test. You get the 80-case trial, where the assistant missed two urgent repairs and caught three things the human forgot, the decision record the agency kept, and a checklist to copy. Read the companion.

Prediction: by 30 June 2027, at least three companies on the 2026 Fortune Global 500 will say on the record that they have rehired for, or re-created, roles they cut citing AI. The companion has the counting rules. SHL · release · Fortune · headline · Robert Half · 32% · Gartner · the trap

Run this before handing a task to AI, a contractor or a new hire. It interviews you. Allow ten minutes. I’m thinking of handing a task over. Before I do, help me find the work nobody wrote down. Don’t tell me yet whether AI can do it. TASK: [what it is, who does it now, how often] WRITTEN VERSION: [paste the procedure, checklist or job description, or ‘none’]

1. Interview me one question at a time about the last three times this task was actually done. Ask what happened that isn’t in the written version: what was checked, who was asked, what was decided, what went wrong. 2. Stop after ten questions, or sooner if my answers stop adding anything. 3. List every step, written and unwritten. Tag each unwritten one. KNOWLEDGE: only this person knows it. RELATIONSHIP: it depends on who’s asking. JUDGEMENT: a call, not a rule. PEAK: it only shows up when it’s busy. 4. Sort all the steps into three piles. HAND OVER. HAND OVER WITH A CHECK: name the check and who does it. KEEP. If you’re unsure, it goes in KEEP. Mark every pile PROVISIONAL: nothing switches until it has been tested on real cases. 5. For each KEEP step, say what would have to be written down or tested before it could move. 6. Finish with a one-page brief of everything I told you. Facts only, no verdicts. I’ll take it to a fresh chat.

Why it works: ask someone for their procedure and you get the tidy version. Ask about last Tuesday and you get what happened. The four tags are my own framework, prompted by the reasons Robert Half’s managers gave. One limit: it only finds what you remember, so do the interview with the person who does the task.

Employers cited AI for 1,265 job cuts in October, the lowest month this year, says Challenger. ‘Market and economic conditions’ is now 2026’s leading reason. The same day, Gartner reported that only 21% of 275 customer-service leaders said their organisations had reduced human- agent headcount because of AI. Quieter than the headlines. Challenger · October report · Gartner · survey

Claude Fable 5.2 and Mythos 5.2 arrived on Monday, 20% cheaper and up to 30% faster. In Anthropic’s own benchmark results, Fable 5.2 scores 41.3% on the Remote Labor Index against 17.9% for Fable 5.1. The line to read twice: Anthropic says both models voice suspicion that they’re being evaluated more often than any recent model, which makes real-world behaviour harder to judge. Anthropic · announcement

Muse Ember, out on Monday, costs $0.10 per million input tokens and $0.50 per million out, for prompts of up to 200,000 tokens. It’s built for agents’ repetitive jobs: tool calls, extraction, document analysis. Routine work is getting very cheap to hand over. Your attention belongs on the rest. Meta · announcement · PYMNTS · report

Scribe Scribe records someone clicking through a browser task and turns it into a step-by-step guide with screenshots. The free plan covers browser work; desktop apps need a paid one.

I’d use it backwards. Record the person who does the task now, on three real cases. The guide is your written version. Then watch for what it couldn’t capture: the pause, the second tab, the message to a colleague, the ‘I know this customer’.

The catches. It sees clicks, not reasons. Reviewers say to expect to edit a fifth or more of the text. It records a real screen, so check it before sharing. And I haven’t run it on a live handover: this rests on documentation and reviews.

Verdict: worth the free plan for one task you’re about to hand over. Scribe

When the facts reached your assistant in pieces, don’t trust the conclusion at the end of that thread. Ask for a facts-only summary, check it, and put the question again in a new chat.

The evidence is a preprint posted on 6 November. Researchers ran 13 chatbots through ten-turn conversations with simulated users describing mental-health problems. With every detail in one message, the chatbots missed 5.6% of emergencies. With details arriving over ten turns, they missed 20.3%, and an early view rarely shifted as evidence built up. Re-presenting the facts disclosed in those conversations as a single message brought misses down to 12.1%.

The limits: one preprint, simulated patients, one kind of decision. Applying it elsewhere is my inference. And the gap between 12.1% and 5.6% isn’t the cost of summarising: those two tests didn’t start from the same information.

Pro move: have a second model write the summary. It has no earlier answer to defend. arXiv · path-dependent triage

Pick one task you plan to give to AI, or already have. Run the prompt with whoever does it today.

The piles are a first sort, not a go-ahead. Before anything switches, write down what result would stop you. Then run ten real cases through the AI alongside the person who does the task now, with their decisions standing, plus two or three rare cases where a miss would be costly. HAND OVER moves once that test is clean. WITH A CHECK moves once the checker is named too. KEEP stays with a person.

Which task, and what was the unwritten step? For example: ‘Supplier invoices: matching the PO to the delivery note, which lived in one person’s head.’ Leave out names. I read every reply.

Know someone about to ‘replace’ a role with a chatbot? Forward them this first.

Built for the most ambitious companies in GTM, Attio is an end to end revenue platform that allows teams to run multiple plays across different revenue motions and lifecycles.

Teams at the frontier are already building on it: Granola, turbopuffer, Modal, and many many more.

Start building on Attio.