Where to start preparing a company for AI

Almost everyone has worked out by now that artificial intelligence doesn't sit well on top of a disorganised operation. What's left unsaid is what you actually do with that conclusion on Monday morning.

The most common reading is the worst one: tidy everything up first. Standardise all the data, connect all the systems, write down all the procedures — and only then think about AI. It's a two-year project that tends to die around month six, when someone asks what's actually been gained from it so far.

There's a shorter route.

Choosing the first case

Instead of preparing the company, prepare just one thing. And choose it carefully, because the choice does half the work.

A good first case has four features at once:

• It happens often. Dozens of times a week, not twice a month. Only then is there enough material to learn from and a gain worth noticing.
• It bothers someone. If nobody complains about that task, it isn't worth starting there — there'll be nobody championing it internally.
• The data already lives in one place. If the information you need is spread across three systems and a notebook, pick a different case to start with.
• It can be measured. It has to be possible to say, after two months, whether things got better or worse.

Typical candidates: sorting requests that come in through the site's form, drafting the first version of proposals that always follow the same template, answering clients' repeated questions, categorising expenses.

Tidy up only what that case needs

With the case chosen, the preparation work stops being abstract. You look at the information that task consumes and ask three things.

Where is this piece of data born? If there are two answers, pick one and treat the other as a copy. Is it always written the same way? If half the records use "Ltd." and the other half "Limited", someone is going to have to standardise it. Does it arrive on its own? If it depends on a weekly export done by hand, the system will be working with last Thursday's information.

It's a job of days or weeks when it's limited to one case. It's a job of years when you try to do it for the whole company.

Straighten out the process before automating it

There's one question worth asking before writing a line of code: if five people did this task, would they do it the same way?

If the answer is no, what's behind it isn't one process — it's five. And a system that learns from five different processes is going to produce answers that look arbitrary, because they are.

Writing down the rules that today only exist in the head of whoever does the task is often the single most useful part of the whole project. Some companies solve half the problem right here and conclude they don't need the second half after all. That's a perfectly acceptable result, and a cheap one.

Say what you expect, in numbers

"Gaining efficiency" isn't a goal, it's a hope. Before starting, write one sentence with a number in it: halve the time between a request arriving and someone answering it; go from three days drafting proposals to the next morning; handle seven out of every ten repeated questions with no human involved.

Measure your starting point before changing anything. It's the only way that, two months from now, the conversation isn't about impressions.

And agree upfront what happens if it fails. A first case that goes wrong and gets switched off is a result; one that goes wrong and nobody switches off is a problem.

The second case is cheaper

Once the first case is up and running, notice what's left behind: the connections between systems are already built, the data in that area is already tidy, and the team already knows what to expect. The second case gets the benefit of all of it.

That's why this route is faster than tidying up the whole house first, even though it looks slower at the start. The tidying still happens, but it pays for itself at every step — instead of being a blind investment stretched over two years.

This is how we work in AI consulting and delivery and in process automation.

Already have a case in mind?

Tell us what the task is and how many times a week it happens. We'll tell you if it's a good first case, what needs tidying up first, and what it costs.

Request a quote