AI consulting services. Built, not advised.

Looking at how the business really works, finding the jobs that repeat the same way every time, and deciding which of them are worth handing to software. Then building the ones that are.

AI consulting services, done properly, are that deciding job rather than the building one. The building follows it and is the easier half.

The tools are cheap and there are hundreds of them. What is scarce is knowing which job in your business is worth handing over, and that answer does not come out of a product demonstration. It comes out of watching a week of the business.

What gets looked at

• The jobs that repeat the same way.

  Anything happening more than a few times a week, with no judgement call in the middle of it, is a candidate.

• The jobs waiting on somebody noticing.

  An email that has to be seen, a form that has to be checked, a stage nobody moves until a person looks.

• The information typed in twice.

  Anything a person copies from one screen to another is a handoff that nobody designed and everybody pays for.

• The jobs that stop when it gets busy.

  These are the expensive ones, because they stop in exactly the week they would have earned the most.

What comes back is a short list in the order worth acting on, with a rough cost beside each one. Not a strategy document, and not a list of everything technically possible.

A good tool aimed at the wrong job still has to be maintained by somebody. Most of the value here is in what does not get built.

The common failure is not picking a bad product. It is automating something that happens twice a month, or automating a process nobody had written down, so what gets built is a faster version of something that was already wrong.

The second common failure is starting with the most interesting job instead of the most expensive one. The interesting one makes a better demonstration. The expensive one pays for the work.

What the deciding is worth

• The builds you do not pay for.

  Every job that should be left alone is money not spent and one less thing to keep working.

• The order the work happens in.

  Doing the wrong one first costs months, because the second one waits on it.

• Something that keeps running.

  A build nobody understands stops working the first time the software around it changes.

The Australian government's guidance on using AI safely in a small business makes the same point from the other direction: start from the problem, not from the tool.

The job has to be written down, the information it needs has to exist somewhere, and somebody has to own reading what it does. Where any of the three is missing, building goes badly.

This is the part that gets skipped, and skipping it is why so many automation projects quietly stop being used. A process that only lives in one person's head cannot be handed to software, because nobody can say what it is supposed to do.

The three that have to be true

• The job is written down.

  Not in detail, but end to end, including what happens when something unusual turns up.

• The information exists somewhere.

  If a step needs a fact nobody records, that fact has to start being recorded before anything can use it.

• Somebody owns reading the output.

  Every automation needs one person who notices when it stops or when it says something wrong.

Writing the job down usually takes longer than building it, and that is normal rather than a sign something has gone wrong. It is also the part that makes the build survive the first change to the software around it.

It will not make a judgement call you would want to stand behind, it will not fix a process nobody has written down, and it will not know anything about your business that is not recorded somewhere.

The honest limit is narrower than the marketing suggests, and knowing where it sits is most of what stops a build going wrong. Anything requiring a decision that carries consequences stops and waits for a person.

Where it stops, deliberately

• Anything that is a judgement call.

  It routes the case to a person rather than guessing at it. A wrong answer sent in your name is the risk.

• A process nobody wrote down.

  It cannot infer how you work. Where the process is not settled, settling it is the work.

• Anything not recorded anywhere.

  If the business does not have the fact, neither does the software.

Where the work is genuinely repeatable and written down, it runs. Where it is not, it stops and waits, and every run is logged so you can read what it did.

Somebody sits with the work. The repeated jobs go on a list with an honest cost beside each, one comes off it first, its limits go down on paper, and it runs watched before it runs alone.

One job is set up and left running before the next starts, and the first one is the one costing the most rather than the one that demonstrates best.

1. Somebody sits with the work

A week of the business, looked at properly. Not a meeting and not a survey, because what people say they do and what actually happens are different.

2. The repeated jobs go on a list with a cost

A rough number beside each one, in hours or in lost work. That number is what decides the order rather than anybody's opinion.

3. The first one comes off the list

Usually the one that stops happening when the business gets busy, because that is the one costing money in the weeks it can least afford to.

4. Its facts and its limits go down on paper

What it is allowed to say, what it must never say, and what it does when something unexpected turns up. This is written before anything is built.

5. It runs watched before it runs alone

For a fortnight, with a person approving each message. That fortnight finds the cases nobody predicted, in time to fix them.

6. The next one starts once the first is boring

A job is finished when nobody is thinking about it any more. That is the signal to start the next one, rather than a date on a plan.

The looking is short. Writing down how a job actually happens takes longer than anybody expects, and it is the part deciding whether the building works.

The order it happens in

• The looking comes first.

  A week of the business, and nothing has been built at this stage.

• Then the writing down.

  The slow half, and the part most projects skip on their way to being unused a year later.

• Then one job gets built and watched.

  It runs with a person approving each step, long enough to meet the cases nobody predicted.

A single job takes about a week to build and a week to watch. A set of them takes a couple of months, and each one after the first is shorter because the connections already exist.

Ask what they look at before recommending anything, whether they sell the tools they recommend, who owns what comes out of it, and what they would tell you not to build.

These are worth asking of anybody, including this one, so the answers given here sit under each question.

• What do you look at first?

  A week of the business. Anybody recommending something before looking is selling a product rather than answering a question.

• Do you sell the tools you recommend?

  No. Whoever supplies your software is not the right person to ask whether you need it.

• Who owns what comes out of it?

  You do. It runs on your accounts, in your name, and it keeps running if the arrangement ends.

• What would you tell me not to build?

  Anybody who cannot answer that has not looked hard enough, or is not going to tell you.

• Am I locked into a contract?

  Not here, it runs month to month. Any arrangement you cannot leave at the end of a month is worth a second look.

All of this can also be done in house, and plenty of businesses do it well. The six steps above are the same either way.