Your First Ninety Days With AI

Most businesses that get nothing from AI did not choose badly. They started without a plan, in the middle, on whatever seemed most interesting.

This is the sequence that works, laid out by month, for a company between five and fifty million with no dedicated technical staff.

Weeks one to three: find the process

Not a strategy. One process, chosen against four conditions.

It happens every day, so payback arrives in weeks and problems surface immediately. It follows the same shape each time, so the rules can be written down. Nobody enjoys it, so you are not fighting the person who owns it. And it currently waits on somebody being available, which is where the capacity actually is.

Lead responses. Document chasing. Report assembly. Routing. Quote preparation. None of it interesting, all of it where the hours are.

If two experienced people describe the process differently, stop and settle that first. An hour in a room, and frequently that hour delivers most of the value the automation was going to.

Week four: write down the number

Agree what this should move and record where it stands today.

Elapsed time from trigger to done. Hours spent. Error rate. Enquiries answered within an hour. Whatever it is, capture it before anything changes.

This is the step most often skipped and the reason so many projects are later described as producing no measurable return. Without a baseline there is nothing to compare against, so the result is a matter of opinion and opinions do not get budget.

Weeks five to eight: build the narrow version

Something in real use inside a month, however narrow. Not a pilot, not a demonstration. A person doing actual work with it.

Three things have to be true. It triggers on an event rather than a person. It writes into the system where the work already happens, so nothing has to be carried by hand. And when it fails it retries, then tells a named person with the reason attached.

That last one determines whether you are allowed to automate anything else. A system that fails silently gets discovered by a client, and the conclusion inside the business is not that one step needs fixing.

Weeks nine to twelve: watch it and fix what you find

Leave it alone and observe. What did it get wrong, what did people work around, what exception did nobody mention during the mapping.

The exceptions are the real work. They live in individual heads and only surface when a system meets them, which is why building first and refining second is faster than trying to specify everything in advance.

At the end of the period, compare against the number you recorded in week four.

What the research suggests about where to point it

Stanford’s AI Index for 2024 reviewed the evidence on assisted work and found it both speeds up completion and improves output quality, with the largest gains among people who were previously less skilled at the task.

That is a useful filter for choosing. The strongest cases are where capability is currently limited by one person’s availability, not where an already efficient process could run slightly faster. In a business where one experienced individual is the bottleneck on everything, raising the floor matters more than raising the ceiling.

Who has to be involved

Three people, and one of them is not optional.

The person who does the work today, because they hold the exceptions and no map is correct without them. The person who owns the outcome, who decides when something is good enough to run rather than perfect. And somebody with access to the systems, because half of first attempts stall not on capability but on waiting three weeks for a login.

Sort the access in week one. It is the least interesting item on the list and the one most likely to cost you a month.

What it should cost

A first automation of the kind described here is a small project, not a transformation programme. If a proposal for one process comes back looking like a platform migration, the scope is wrong.

The useful test is whether the payback is legible. If a process consumes fifteen hours a week and the build removes ten of them, the arithmetic is obvious to everybody in the room and nobody has to be persuaded. If the return depends on a projection three years out, that is not a first project.

Keep the first one small enough that being wrong about it costs you a quarter rather than a year.

What not to do in the first ninety days

• Do not start with the most painful process. It is usually the most complex, and complexity is where first attempts die.

• Do not run a committee. One process, one owner, one number.

• Do not buy capability and look for a use case. That produces activity and no result.

• Do not automate something running monthly. The build cost outlives the process.

• Do not skip the failure path because the happy path is more interesting to build.

What month four looks like

One process running, a number that moved, and a team that has watched something work.

That last part is worth more than the hours saved. The second automation in a business is always faster than the first, because by then you know where your own systems are awkward and people have stopped assuming it will not work.

Businesses that try to do five things at once in the first quarter generally finish none. Businesses that finish one usually do four in the following year.

Sources

• Stanford Institute for Human-Centered Artificial Intelligence, Artificial Intelligence Index Report 2024, 15 April 2024.

AI Optimize runs exactly this sequence, and will tell you in week three when the honest answer is that the process should be fixed rather than automated. That work sits under Workflow Automation.