Guides ยท 22 November 2026
How to test AI on your own work
This site gives every job a score. So here is the honest limit of a score: it tells you how exposed a kind of work is on average. It cannot tell you how well AI does your work, to your standard. You can measure that yourself in about ten minutes, and what you learn is the part of your job that is hardest to replace.
The short answer. Next time you correct an AI result, do not just fix it. Say what you like and dislike, and have the AI turn each comment into a yes or no check. Save the list. Use it to grade the next result, the next prompt and the next AI tool.
Two ways to look at the same future
In November 2026 the publication Every held a conference on what work looks like after automation. In the closing talk, its chief executive Dan Shipper set out two ways to predict what happens to jobs. One draws a line through a large pile of data. The other looks closely at one real person doing real work.
Source: What's Left for Humans After Automation, Every, 11 November 2026, by Grant Harvey and colleagues. The talks and interviews described below were reported there. The figures in this section are theirs and we have not checked them. The method, the prompts and the link to our own data are ours.
The big-picture view looks alarming. Every reports that on one OpenAI test of professional tasks, AI went from beating human experts about one time in eight in 2024 to about five times in six in 2026. Our own scores are the same kind of view: three studies, 879 jobs, one number each.
Shipper's objection is that this view compares AI with people who are not using AI. It holds the human still. Up close, people get better at their work as the tools get better, and the part that stays scarce is deciding which of many possible results is the good one. His line, as Every quotes it: "Generation is really cheap, but ranking is hard."
Our own data points the same way. The jobs we score as most exposed have not shrunk. The number of people in them rose 6% between 2021 and 2024. The study is here. Exposure tells you which tasks are changing. It does not tell you who is good at judging the result.
Your corrections are the test
At the same conference, Michael Taylor, who leads AI testing at Every, showed a way to measure AI against your own standard without knowing anything about testing. His claim is that you only need an opinion.
You already have one. Every time you read an AI draft and think "too long", "wrong tone" or "it missed the main thing", that is a standard. It disappears when you close the chat. The method is to catch it.
- Pick one task you do often. A weekly update, a quote, a lesson plan, a handover note.
- Get an AI result for it, the way you normally would.
- Say what you think. What is good, what is wrong, what you would never write. Talking works better than typing, so use voice input if you have it.
- Have the AI turn each comment into a check. A check is a question about the result with a yes or no answer. "Is the bad news in the first three lines?" is a check. "Is it good?" is not.
- Save the list with the task name on top.
- Grade the next one. Paste the list and a new result, and ask which checks failed. Do this when you change your prompt, and when you try a different AI tool.
Paste this to start:
I do this task often: [name the task]. Below is an AI result for it. After it, I will say what I like and do not like. Turn each thing I say into a check: a question about the result that anyone could answer yes or no by looking at it. Use only what I say. Do not add standards of your own. For each check, tell me if it would be wrong for a normal variation of this task (for example, true this week but not next week). Widen those or leave them out. Give me the final list in one block, with the task name at the top, so I can save it. RESULT: [paste] WHAT I THINK: [say it however it comes]
The trap: a check can teach the wrong lesson
Taylor's warning is that a check can be passed while the work still misses. His example was a drawing of a cat: after he said it did not look like a cat, the AI made itself a rule that the cat must face forward, and then drew a forward-facing cat when he asked for one curled up asleep.
We hit the same thing in our own test of AI writing styles. We improved a style guide using notes on four practice pieces. On those four it then won 8 judgments to 0. On four new pieces it lost 2 to 6. The guide had learned the practice pieces.
Three habits keep your checks honest:
- Read each check and ask "would this be wrong next time?" "Mentions Friday's deadline" should become "states the next deadline".
- Keep the list short. Twelve checks or fewer. Long lists reward results that tick boxes.
- Trust the work over the score. If a result passes everything and still disappoints you, the checks are wrong, not you.
Where the saved hours go
The fear behind most visits to this site is that time saved by AI becomes jobs cut. Every interviewed Hugh Wimberly, an engineer who builds AI helpers inside client companies, and he described what he has seen across six clients.
At one firm, setting up a new customer took up to 100 hours of mostly routine clicking. Automation cut that to about 30. The team was not cut by the same share. Cheaper set-up meant the firm could serve customers who could not afford it before, and it wanted more staff. At an accounting firm, the time saved is going into reading clients' records for opportunities the client has not asked about, which moves the accountant closer to an adviser.
He was plain about the limits, and so should we be. He said some jobs did go, including overseas roles that were one repeated click. His clients chose to do this with their teams, which is not true everywhere. This is one person's account, not a study.
It does suggest something you can try this week. Take what you already hold about the customer, manager or team you know best, check what you are allowed to share with an AI tool, and ask:
Here is what I have on [who]: [paste notes, reports or emails with names removed]. What is one problem or opportunity they are likely to face soon that they have not asked me about? Show me the evidence in what I pasted. If the evidence is thin, say so.
Take the best answer to your next meeting. The routine part of the task is what AI reaches first. Knowing the people you serve well enough to see what they need next is the part that grows.
What to take from this
- A score, ours included, describes the average job. It is a reason to look closer, not a verdict.
- Your standard for good work is worth writing down. It is what lets you direct AI, compare tools and show an employer what you add.
- Checks need looking after. Test them on new work and drop the ones that stop fitting.
The people quoted here spoke at or were interviewed for Every's conference, and their views are their own. We have not spoken to them.
Take the skill with you
One file that does the steps above. It turns your comments into checks, flags the ones that would be wrong next time, and grades any later result against your list.
Download the skill (free)It is one plain text file. Open it, copy everything, and paste it as your first message in ChatGPT, Claude, Copilot or Gemini. Then follow what it asks. In Claude you can also add it as a skill. No sign-up is needed for this one. We checked it against our ten skill checks. Nobody outside our team has used it yet.More guides
Found a mistake? Use the feedback box at the bottom right. Who writes this site.