The installation manual for a new machine used to take Teodora Ilić eleven days. It now takes eight. The firm in Wigan that makes the dishwashers had been told it would take two, and the gap between eight and two is the whole subject of this page.
Four of the eleven days were drafting and seven were checking. The drafting is now one day. Nothing else moved, because a job is a bundle of tasks, and a model moves some tasks rather than the job.
Sorting one job into three piles
Take Teodora's actual week and sort it by a single question: how expensive is it to check a draft of this. The first draft of a procedure from the engineers' notes is cheap to check, because she knows what a correct procedure reads like and can see a wrong one in a minute. So is the release note from a changelog, so is a paragraph rewritten three reading levels down, and so are the four translations, each read by somebody on the sales side who speaks the language. All of those moved, and moved a lot.
The tasks that move are the ones where a draft is cheap to check. That sentence does the work of every list of AI-proof professions ever published, and it sorts tasks rather than jobs, which is why it keeps being right.
The second pile moves only with the right page in front of it. Which gasket fits a unit built in 2019 is answerable, and answerable wrongly and fluently, unless the fitment table is retrieved and quoted. The parts list against the current drawing is the same shape of problem. These are the tasks where the work went into plumbing rather than into prompting.
The third pile did not move at all. Standing in front of the machine to confirm that a procedure is physically possible for a person holding a torch. Deciding what the procedure should be. The one email a month from an engineer in Leeds saying the manual is wrong, which is where the job has always actually been. And signing it off. The tasks that stay are the ones where the checking is the work, and no draft helps with a task that was already a check.
Where the seven days went
They did not go anywhere. That is the finding. The drafting fell from four days to one and the checking stayed at seven, so the queue moved to the checking, and now sits there permanently. The bottleneck moves to whoever reviews, and nobody staffs for that, because a review is invisible in a plan in a way that a draft is not.
It gets worse before it gets better, for a reason worth naming. Time saved on drafting arrives as more drafts unless somebody decides otherwise. Teodora's colleagues can now produce a first pass at a manual themselves, so three arrive in a week where one used to, and every one of them lands in front of the same person and the same seven days. The saving was real and it was spent before anybody chose how to spend it.
The eight days, said out loud
The stall here is that the saving never gets claimed. Everybody says the drafting is faster and nobody says the number, so the old figure of eleven days quietly becomes the new figure of two in the plan, and the person doing the checking absorbs the difference. It does not feel like a stall because the work genuinely improved and nobody is being dishonest.
The move that gets past it is small, unpleasant and specific: write down the before and after for each task, in days, and say the total. Eight days, of which seven are a person reading. That sentence is worth more than any argument about what the technology can and cannot do, because it is checkable and it is about this job. What the person is still for is a question that only has a real answer once it is asked task by task.
The order
Automating is the third step and it is where everyone starts
The first two steps produce no artefact, take an afternoon between them, and decide whether the third one is worth building.
- Sort your own tasksBy how expensive a draft is to check. Nobody can do this for you
- Find the constraintsWhat may leave the building, and whether a reader is told
- where everyone startsAutomate one taskThe one with the cheapest check, and exactly one
- Watch where the queue goesIt moves to whoever reviews, and that is a staffing decision
- Then read your own fieldUseful after your own sort, misleading before it
The steps are numbered because the order is load bearing. Taken from the third, the output is a pipeline producing drafts nobody has the hours to read.
The output is yours whoever drafted it
A wrong torque figure in a service manual is a wrong torque figure. You are accountable for output you did not write, and the sentence explaining that a model produced it changes nothing about the machine, the engineer or the firm. That is the same standard as work handed to a junior colleague, which is a familiar arrangement and not a new one.
Two constraints on top of it are not a matter of judgement. What may leave the building is somebody's decision and usually a written one, and pasting a customer's site drawings into a consumer chat window is a data question before it is a quality question. And whether a reader is told is a live argument in most fields as of 2026, settled differently in each of them. The constraints that are not optional is worth reading before the automation rather than after the first incident.
The order, and the tempting reversal
Sort the tasks first, find the constraints second, automate one task third, and watch where the queue goes fourth. Wiring models into work that repeats is genuinely the third step and it is where almost everybody starts, because it is the one with a build in it.
Started there, the reliable outcome is a pipeline that produces drafts nobody has capacity to check, which is the seven days problem with a schedule attached to it. What the change looks like in each discipline is a reference to dip into once your own sort is done, and read first it produces confident opinions about other people's jobs and none about your own.
What to open in 2026
The assistant inside the document editor, meaning Microsoft 365 Copilot or Gemini in Google Workspace as of 2026, replaced copying a paragraph into a chat window and pasting the answer back. Neither has a free path: both are a licence per person per month, and that number is what the business case actually turns on rather than any claim about capability.
NotebookLM answers from documents you give it and points at the passage it used, with a free tier as of 2026, and it replaced asking a general assistant about your own handbook and receiving a fluent answer from somewhere else entirely. It is the tool for the second pile: the fitment table in, the gasket question answered, the paragraph cited so the answer can be checked in seconds.
DeepL does the four translations, free up to a character limit as of 2026, and it replaced a general assistant asked to translate, which will not hold a glossary. A manual has fixed terms that must be identical on every page in every language, and a glossary is the feature that difference is made of.
Which leaves the question this page keeps circling. Everything above assumes Teodora is still the person who signs the manual off, and the sorting exercise says nothing about what happens when the third pile is small enough that the firm stops needing somebody to hold it. That is a question about the job rather than the tasks, and it is answered by what a person can still be trusted with rather than by what a model can do, which is choosing work, getting it, and moving up. The version where nobody is checking at all is a model given tools and allowed to act.