Decide what “good” looks like, before you ask AI
Your Essential Report from the Frontlines of Applied AI
Your boss or a senior colleague can tell you whether a piece of work is any good, because they’ve done that work themselves. AI often is unable to, and it won’t warn you when it’s out of its depth.
So this week is about the standard you hold it to. There’s a prompt below that pulls that standard out of you before you start, and a case study from evidence reviews, where the rule for stopping is agreed before any screening happens.
Plus the week’s news in applied AI — a classroom, a complaints line, a parliament and a warehouse.
You can only judge an AI’s answer if you’ve already thought through what a good answer should contain.
AI is almost always confident when it answers. Clean structure, assertive headings, an even tone — it reads as authority. We call that ‘format-based acceptance’: you are more likely to believe something is high quality if it’s tidy and well structured. But AI does not possess your unique experience, and has no real opinion on what is good.
So rely on your experience, and take the time to write the standard down before you ask. Three or four things a good answer contains, and the one thing it must not get wrong.
When the work is past what you could have done yourself, you can still check that the sources exist, the arithmetic holds, and it answers what you asked.
That list is also most of the brief. Give it to the AI before it drafts, and you have both the instruction and the thing you’ll check against.
- Education
Trinidad and Tobago’s education ministry gave Form One and Two students an AI learning assistant
The platform carries e-books, video simulations and an AI avatar that takes pupils through problems with step-by-step explanations, across mathematics, English language arts, social sciences, integrated science, ICT and Spanish. Education minister Dr Michael Dowlath demonstrated it at the launch.
Trinidad and Tobago Television, 2 September 2026
- Insurance
The Korea Life Insurance Association began handling consumer complaints with AI this month
It transcribes calls between consumers and counsellors as they happen, sorts the complaint by type, and puts supporting material in front of the counsellor mid-call. Three systems went live this month; the association plans 17 by next year.
Seoul Economic Daily, 7 September 2026
- Public sector
The Swiss parliament is giving members an AI assistant for the autumn session
PIA goes to council members, their staff and parliamentary services, and is aimed at working through committee documents rather than internet research. Swisscom runs it with the data kept in Switzerland, on a budget of up to CHF 150,000 for a pilot of roughly a year.
Radio Central, 4 September 2026
- Logistics
CJ Logistics put two humanoid robots to work packing boxes in a Korean warehouse
They load cushioning material into boxes at the Olive Young centre in Yongin — what CJ Logistics says is its first use of humanoid robots in a live logistics process, after field tests at its Gunpo site last year. It plans to extend them to picking, sorting, inspection and packaging.
The Korea Times, 3 September 2026
Most briefs name the task and skip the standard, so the AI guesses and you are left with nothing to mark the draft against. Run this first and keep the answer next to you while you read what comes back. It is written for a decision and works the same way on a document, a plan or a hire. Keep the topic line general — it does not need anything confidential in it.
"I'm making a decision about [topic]. What would a successful outcome look like for this decision — not the best possible outcome, but a clearly good one? What criteria would a thoughtful observer use to judge whether this decision was made well, independent of how it turns out? What am I likely to regret not having achieved if the decision goes poorly?"
Prompt 4.4 from AI+ Decision Making, word for word.
Agree what counts as good enough before the AI starts, and write it into the protocol.
Cochrane’s reviewers use machine-learning screening to rank thousands of studies so the likely-relevant ones surface first. What keeps it defensible is a stopping rule agreed in the protocol before screening starts: how far down the ranked list the team must go, and what the classifier’s performance has to be at that point. The tool’s own estimate of what is left does not get a vote.
In 2025 Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence published a joint position on AI in evidence synthesis: the author is ‘accountable for the content, methods, and findings’ of the review, including the decision to use AI at all.
Set the threshold while you are still planning. A draft that reads well is hard to argue with once it is in front of you.
