Where Should Your Business Start With AI? How to Choose the First Workflow

Most conversations about AI start with a tool. Someone has seen a demonstration, or a competitor mentioned a platform at a conference, and the question arrives already shaped: should we be using this?

It is the wrong first question. Not because the tool does not matter, but because the answer depends entirely on something nobody has established yet, which is what the work actually looks like.

The firms that get a useful result from a first AI project tend to have done the same unglamorous thing first. They chose the workflow before they chose the technology. Here is how that choice gets made.

1. Start where the work stalls, not where the technology looks impressive

Every demonstration is impressive, because every demonstration is built on a clean input. Real work stalls in specific, boring places that no demonstration ever shows you.

In a law firm it is often the gap between a client sending documents and anyone reviewing them. In an accounting practice it is the third request for the same missing statement, or the week each month when recurring work piles up behind one reviewer.

Three questions find more candidates than any list of use cases:

Where does work sit waiting for someone? Where does the same task get done again every week or every month? Where is a senior person doing something a junior person could do if the information were organised?

Answer those honestly and you will have four or five candidates. None of them will have come from a vendor.

2. Write the workflow down as steps before you judge it

Most workflows have never been written down. They live in the head of whoever does them, and they survive because that person is reliable.

Before deciding whether AI can help, write the thing out properly. What triggers it. What information it needs and where that information comes from. Who touches it and in what order. What gets produced. Who approves it. Where it goes next.

This is tedious and it is the step that decides the project. A workflow you cannot describe in steps cannot be improved in parts, and improving it in parts is the only version of this that works.

It also has a useful side effect. Writing the steps down frequently reveals that the real problem is not AI-shaped at all. Sometimes the bottleneck is one approval that sits with a person who is in court three days a week. No technology fixes that, and finding out now is considerably cheaper than finding out later.

3. Separate the steps that need judgment from the steps that only need time

Inside any workflow, some steps require professional judgment and some require somebody to spend an hour moving information from one place to another.

Deciding whether an indemnity clause is acceptable is judgment. Locating every clause of that type across sixty documents is time. Deciding how an unusual transaction should be treated is judgment. Chasing the client for the statement that shows the transaction is time.

AI is genuinely good at the second category and unreliable at the first. So the workflows worth changing first are the ones where the time-consuming steps are large, repetitive, and separable, and where the judgment stays with the person whose name is on the advice.

If a workflow is mostly judgment, leave it alone for now. You are not avoiding ambition. You are avoiding a project that will need a person checking every output, which is the same cost you started with.

4. Check that the information the work depends on actually exists

A workflow only becomes a candidate if the information it consumes is available in a form that something can read.

If the answer lives in an email thread nobody can locate, or in a partner's memory, or in a PDF that is a photograph of a page, then your first project is not an AI project. It is an information project, and it has to happen first.

This is the most common reason a promising candidate fails. The workflow was suitable. The inputs were not. It is worth thirty minutes checking before it is worth a budget.

5. Decide what better means, and how you will know

Before anything is built, name the measure and write down what it is today.

Turnaround time from request to delivery. The number of times a piece of work goes back for correction. Hours a senior person spends reviewing routine output. How much of the month's recurring work is finished by the tenth.

Pick one. Record the current number. Six months later somebody will ask whether this actually did anything, and you will be able to answer with something better than an impression.

If you cannot state the current number at all, that is a useful finding in its own right, and far cheaper to discover now than after the invoice.

6. Choose something small enough to finish

The instinct is to pick the most valuable workflow. The better instinct is to pick the most finishable one.

A contained workflow, with a clear owner, a real measure, and a small number of people involved, will teach your firm more than an ambitious project that stalls in month four and quietly stops being mentioned.

The first project's job is not to transform the practice. Its job is to produce one working thing, and a group of people who now understand accurately what this technology does and does not do. The second project is where the real value usually sits, and you only get a second project if the first one finished.

The pattern underneath all six

Every step above is a question about the business rather than a question about the technology. Which is the point: choosing the first AI workflow is a management decision wearing a technical costume.

The technology is capable now, and it is getting more capable without your help. The scarce skill is deciding what to point it at, in what order, with which parts left to a person.

That decision is not made by a demonstration. It is made by somebody willing to write down how the work really happens.

If you are trying to work out where your own firm should start, that is exactly the conversation I enjoy.