The HOW is now available to anyone. Knowing the WHAT is what keeps you valuable.
Your Essential Report from the Frontlines of Applied AI
Welcome to the first weekly edition of AI+You.
What you'll get every week: one idea worth learning, significant news in applied AI, a prompt you can use today, and one organisation doing it well — all derived from our research and our books.
It is prepared for the working professional using AI in their job every day, and our goal is simple: to become a trusted resource you rely on to stay competitive in a lightning-fast field.
The timing is deliberate. The AI+ collection has just launched — more than 300 titles across 166 professions, written by practitioners and refreshed every six months. AI+You covers what changes between editions.
Which brings us to this week's idea.
The HOW is now available to anyone. Knowing the WHAT is what keeps you valuable.
You can now ask AI to do almost anything — write copy, draft contracts, design creatives, do research. And it can do it at a quality and skill level that will impress you.
But it has no idea what's worth doing. Which leaves the two questions it can't answer for you: WHAT should you ask it to do? And WHAT should the result look like?
The answer to WHAT comes only from your judgement and experience — every project you've delivered, every client you've read correctly, every call that turned out to be the right one.
Since AI now knows how to do so many things, deciding WHAT to ask it to do is now the most important part of our jobs.
- Healthcare
Endoscopists used AI to spot polyps during colonoscopy — and got worse at spotting them unaided
Across four Polish centres, detection on standard colonoscopies fell from 28% to 22% once AI had become routine. Observational, not a randomised trial.
The Lancet Gastroenterology & Hepatology
- Healthcare
Swedish breast screening used AI as a second reader and found 29% more cancers
A randomised trial of 105,934 women. No rise in false positives, and radiologists' reading workload fell by 44%.
The Lancet
- Public sector
UK civil servants used AI to sort 50,000 consultation responses into policy themes
Two hours and £240 for work that normally takes weeks; experts then spent 22 hours checking it. The department reckons it could save 75,000 days a year across government.
UK Incubator for AI, DSIT
- Banking
HSBC uses AI to screen 980 million transactions a month for money laundering
It finds two to four times more financial crime than the rules-based system it replaced, with 60% fewer false alarms — so investigators spend their time on real risk.
HSBC
Before you paste it. Most proposals arrive commercial-in-confidence. Check you are permitted to put this one into the tool you are about to use.
I am evaluating [product / service / vendor proposal] for our library. I will paste the vendor's proposal/brochure/trial documentation below. Produce a structured summary with: - What is being offered (in plain language). - What is explicitly included and what is not. - Any claims in the material that are vague, hedged, or aspirational rather than concrete. - Pricing, licensing, and renewal terms, with anything unclear flagged. - What this would mean for our workflows (best-guess only, based on the text). - 3–5 questions to put to the vendor before making a decision. Work only from the material I paste. If something is not in the material, say so rather than generalising from industry norms. Material: [paste]
Prompt 4.7 from AI+ Librarians, word for word.
Set the quality bar, measure where AI reaches it, and put human review where the evidence says it is needed.
The Library of Congress tested GPT and open-source models on drafting catalogue records, measuring accuracy field by field. Title and author reached 99 per cent. Subject and genre fell below 50.
Its threshold for automating a field is 95 per cent. Nothing cleared it. So the model drafts and a cataloguer corrects every record before publication, with the measurement showing where to look hardest. The review is not caution — it is where the record is made correct.
Most work has the same structure. Parts of a task are mechanical and AI is reliably good at them; others carry the judgement and it is not.