Pick the job before you pick the tool

The version of this that goes wrong starts with the tool. You read about an agent, you open the thing, and within an hour you have something scheduled. It wakes up and it reads a document. It reports that the document has not changed, and tomorrow it does exactly that again. No output moved. Nothing got decided.

That is what starting from the tool gives you. Motion, and a bill.

A job is not a topic, not a wish, not "AI for marketing", not a tab you keep open because the demo looked clever.

The rules

A job has an input, an output, an owner and a frequency. Name all four or you do not have a job yet. The input is a thing that exists somewhere you can point at. The output lands somewhere a human already looks. The owner is a person, not a team. The frequency is a real interval, not "ongoing".

Write it as a sentence a temp could follow on day one. If the sentence needs you standing behind it explaining, the agent will need you standing behind it too. Every gap you leave gets filled with a guess. The guess will be confident and wrong.

Pick the job you do badly, not the one you do well. The work you are good at is already fast. The wins sit in the work that gets skipped, deferred, or done at nine at night by whoever has not gone home. Boring, repetitive, low-stakes, unloved.

Time it once, by hand, before you automate anything. You need a before number or you will never know if this worked. A stopwatch and one honest run. Write the minutes down.

Refuse the tool comparison. No stack diagram, no fourteen pricing tabs, no thread about which model is ahead this week. The tool is the last decision you make, not the first.

Do it now

1. List five things in your business that happen every week and that nobody enjoys. Write them down fast and do not filter.
2. Cross out anything a customer sees directly. Those come later.
3. Pick the one with the clearest input, not the most valuable one. The clearest.
4. Fill this in, exactly this shape, in a plain text file:

Job:      <verb> <the thing> from <where it lives> and put <the output> in <where it lands>
Owner:    <one person>
Runs:     <how often, as a real interval>
Costs now: <minutes per week>
Evidence: <how I got that minutes number>

5. Do the job once yourself, timed, and fill in the two bottom lines from that run rather than from memory.
6. Open a chatbot and paste this, with your job line in it:

Here is a job I want to hand to an agent.
<paste your job line>

List every input this agent would need that is not stated in that line.
List every decision it must make that I have not specified.
Do not suggest tools, do not write a prompt, and do not improve my wording.

7. Every gap it returns goes back into the job line, or gets ruled out of scope in writing.

A worked example

Say you run a four-person lettings agency. Every Monday somebody checks which rents landed in the bank and chases the ones that did not. Nobody likes it. It slips whenever the week gets busy.

Job:       Match this week's bank export against the tenancy list
           and put a list of missing rents in the shared Rents folder
Owner:     Priya
Runs:      Every Monday at 9am
Costs now: 15 minutes per week
Evidence:  The match alone, timed by hand on one Monday

Notice what the line leaves out. No tool, no model, no app. Just the work, measured.

The failure you will hit

You will pick something too big. It will sound sensible when you say it out loud: handle customer support, sort out reporting, get on top of the numbers. Then you will spend a fortnight building nothing and conclude AI is overrated. The cut is simple: if the job takes a person more than twenty minutes a week, you have picked a bundle of jobs. Split it and take the dullest piece.

Under that sits a quieter mistake, which is picking a job you cannot check. If you have no way to tell whether the output is right, you have picked a job you will have to trust blind. Put it back.

Not for you if

If you will not time yourself doing the work once, stop here. You will have no before number and no way to prove the agent earned its keep. You will have no argument either, when somebody asks you why you bothered.

Your line comes first. Step two builds the agent that reads it, and the Vault is the paid library. The price is on the page.