How long does a first AI job take? The supplier says two weeks. You run operations, and the job is drafting proposals and quotes.

Week six arrives and the step still isn’t live. The team hasn’t agreed what a good draft contains. The price list exists in three versions. The director who signs proposals off hasn’t had an hour for the trial.

Imagine that’s four extra weeks of licence fees with nothing live, and a stalled first job makes the second harder to argue for. I’ll give you a model I call the two clocks, so you can date every block before you start.

The tool’s clock and the job’s clock

Every first AI job runs on two clocks at once, and a supplier’s plan usually shows only one of them.

The tool’s clock belongs to the supplier. Connecting a tool to a mailbox or a folder might take them a morning. Writing the instructions and running test material through them is normally a few days of adjusting. That work is bounded, predictable and largely someone else’s problem.

The job’s clock belongs to you. Deciding exactly which job to run, agreeing what a good output contains, sorting the documents the tool will read and checking its early output all pass through your own people. They do it around the proposals, orders and customer calls they already have in the diary, and none of it can be bought quicker.

So when a supplier says two weeks and the whole thing takes two months, nobody has necessarily done anything wrong. The two weeks was the tool’s clock. The two months was the job’s. Once you see both, you can plan the one that decides your date.

Week one: write the job and the standard

The first block needs nothing installed. It starts with writing the job down in five lines: what triggers it, what goes in, what comes out, who checks it and what happens next. That page takes an hour, and every later week leans on it. For proposals, an enquiry arrives and goes in with past proposals and the price list. A first draft comes out, the sales lead checks it and the director signs it off.

The same week holds the one meeting that decides most of the timeline. The person who writes proposals, the person who’ll check the drafts and the director who carries the cost of a bad price sit down for an hour. Together they agree what good looks like in terms a reviewer can apply in seconds.

While that’s happening, ask whoever writes proposals today to time the next 10 and count how many come back for changes. Those two numbers are your before position. This is the only week you can still collect them honestly, because once the tool is running, the old way has gone.

Preparation is the block with the widest range

The second block is the one estimates miss, and it has the widest range of any. Before a tool can read your material, someone has to decide which version of each document is current, make the terms consistent and settle which folder holds the truth. Preparing documents can’t be handed to the supplier, because only your people know which price list is real.

How long it takes depends on one thing you can check in an afternoon: whether the job’s source material already lives in one agreed place. If it does, preparation takes days. If the answer to “where’s the current price list?” includes the word “usually”, plan for weeks.

Treat those weeks as time well spent. A tool pointed at contradictory documents produces confident, contradictory output, and you pay for that later in checking time. For proposals, that might mean one agreed price list, one set of standard terms and a folder of 20 good past proposals the tool is allowed to read.

The trial takes an afternoon, the decision longer

With the material ready, the trial itself moves quickly. Run 20 real enquiries from last month through the tool, including the awkward ones, and sort the drafts into three piles. Usable unchanged, needed a small fix, and wrong in a way that would have caused trouble with a customer.

The afternoon is the easy part. The slow part is the decision that follows, because the results land with people whose diaries are full of customer work. A verdict that could be reached on Friday waits for the meeting after next, and a fortnight disappears without anyone touching the work.

If you want the timeline held, book the decision meeting before you run the trial. Write the walk-away conditions down at the same time. A trial with a date and a threshold attached turns into a decision. A trial without them turns into a discussion.

Plan three weeks of corrections before it settles

Go live and the job enters its noisiest stretch. Plan on the first three weeks of a live step bringing corrections, questions and the occasional enquiry nobody predicted. This is the period people mistake for failure. It’s the step meeting real input for the first time, and the correction notes it throws off are the most useful information the whole project produces.

Protect two things through those weeks. First, the reviewer’s time, which is real work and belongs in their diary rather than around its edges. Second, the habit of writing corrections down instead of fixing them quietly, because a pattern you can read is a pattern you can fix.

Corrections that repeat point at your instructions or your source material. Corrections that fade to almost nothing after a fortnight mean the step is settling on schedule.

Where your job sits on the clock

Now put your own job against the model. Five questions tell you where you are and what’s likely to stretch.

Is the job written down in five lines? Has the standard been agreed by the three people? Does the source material live in one agreed place? Is the decision meeting in the diary with a threshold? Is the reviewer’s checking time booked for the first three weeks? Every “no” is a block you haven’t started, and a place the job’s clock will run long.

Two of those answers carry a warning. When preparation keeps stretching, the business is finding out that its documents disagree with each other. That problem predates the tool, and the AI job has simply put a date on it. When the delay is waiting for decisions, the job has no real owner. A step that can’t get a verdict in week four won’t get its instructions changed in month six, and that’s how a working tool fades out of use.

Both delays are cheaper to fix in the calendar than after launch. Put a name against each one on the plan.

When the job counts as finished

The step counts as settled when corrections have faded and stayed low through the back half of those three weeks. It also needs someone other than the person who set it up to run it start to finish without help. Then comes the last block: a month of watching in which you change nothing and stay out of the way.

Say the plan is a week on paper, three weeks of preparation, a week for the trial and the decision, and three weeks of settling. That’s eight weeks to a step your team runs without thinking, then a month of watching. With your material already in one place, the same sum might come to five.

Only after that month should you pick the second job. By then the timeline repays you. The written job, the review step, the standard and the supplier contact all carry across, so the next job should take noticeably less than this one did.

Put the five blocks in the diary

This week, sit down for an hour with whoever will own the job and a blank calendar. That’s probably you and the director who signs proposals off.

Mark five blocks: a week on paper, your honest estimate for preparation, and an afternoon of trial with the decision meeting already booked. Then three weeks of settling with the reviewer’s time protected, and a month of watching. Put a name against each block.

Then answer the five questions above and circle every “no”. Those are the blocks most likely to run long, and they’re where your first week of effort should go.