Do I actually need an AI plan, or is this hype?

You need a plan for one or two specific processes, not a company-wide AI strategy. The hype is in the framing, not the effect. The measurable change in businesses that have done this is in how long things take and how many people have to touch them, and that shows up in price and response time where your customers can see it.

Is this just about cutting headcount?

Usually not, and businesses that treat it that way tend to get the worst results, because they remove the people who knew why the process existed. The realistic pattern is the same number of people doing more valuable work, with the re-keying, chasing and formatting handled elsewhere. What does change is the mix of roles you hire for.

Where should a non-technical business start?

Pick the process your customers complain about, and count the days it takes and the number of people who touch it. Then ask which of those steps produce a decision and which only move information. The ones that only move information are your plan. You do not need to understand the technology to do that analysis.

How do we do this without losing control of quality or compliance?

By deciding in advance what a human must approve before anything counts as finished, and making that check real rather than assumed. In our own delivery AI proposes and a person decides: every change is a proposal, automated checks run first, and a human approves it. Nothing is applied automatically.

How long before we see anything?

Rewriting a single process takes weeks, not quarters, and you will see the difference in that process immediately because the waiting steps are gone. The compounding benefit, where you iterate more often because each cycle is cheaper, takes a couple of quarters to become obvious to anyone outside the business.

Your business has legacy processes, your competitors don't

Q: We bought AI licences already and nothing much happened. Why?

Because a licence changes what a person can do at their desk, not how work moves through the business. If every message still has to be opened and read by a person before anyone knows whether it matters, you have made the reply faster inside a process whose cost was always the sorting and the waiting. The gain is in redesigning the path, not the step.

Every business being founded in your industry this year is designing its operating model around AI from day one. They are not better at your job. They just never built the handoffs you are still paying for.

What is actually happening in your industry?

Not what most of the coverage suggests. The interesting change is not that AI has become good at the work your people do. It is that every business being founded in your sector this year is designing its operating model around AI from the first week, and therefore never builds the machinery you are still paying to run.

They are not smarter than you. They are just younger, and in this particular decade that turns out to be worth a great deal.

The advantage is an absence, not a capability

Take the smallest process you have: the contact form on your website.

A message arrives. Somebody opens it, reads it to the end, and works out that it is not a customer at all. It is someone selling search rankings, or lead lists, or a redesign of the website they just used to contact you. They delete it and open the next one. Somewhere in that pile is a real enquiry, waiting its turn behind the pitches.

You do not have to take my word for this. Open the inbox behind your own contact form and count how many of the last twenty messages were somebody selling to you.

None of that reading produces a decision until the final step. It is sorting, done by a person, at the speed of a person, and the real customer pays for it by waiting.

A business being set up this year would not build it that way. Each message is read and sorted as it arrives. A real enquiry is flagged to a person straight away. A pitch is filed in a folder that somebody skims once a week, so nothing is deleted without a person being able to see it. The person’s attention goes only where a decision is needed.

That is one small process, and you could fix it in an afternoon. The point is that an established business is made of hundreds like it: steps that exist because, when the process was designed, a person reading everything was the only way to sort it. A business designed this year never built those steps. That is the entire advantage, and it is worth more than any individual clever thing they do with the technology.

The same inbox, sorted two ways

Sorted by hand, every message costs a person the same attention: opened, read to the end, recognised as a sales pitch, deleted, while the real enquiry waits in the same pile. Sorted on arrival, each message is read and classified as it comes in: a real enquiry is flagged to a person straight away and a pitch is filed for a weekly skim. Nothing is deleted automatically.

Why the gap widens on its own

If that were just about speed it would be survivable. Plenty of businesses win on relationships while being slower than their rivals.

The problem is what speed does to learning. If your competitor can try an idea, see whether it worked and adjust in a fraction of the time it takes you, they do not simply ship faster. They find out whether they were right far more often. A year later they have been corrected many times for every once that you have, and the difference between the two businesses no longer looks like a difference in speed. It looks like a difference in judgement.

That is the mechanism nobody puts in the board pack, and it is the only argument here that genuinely matters.

Why the licences did not fix it

The standard response is to buy everyone an AI subscription and put a training session in the diary. A year later the honest assessment is that some people find it useful for writing emails.

That is not a failure of the technology. It is what happens when you make one step of a process faster without touching the process. If every message still has to be opened and read by a person before anyone knows whether it matters, then drafting the reply faster has changed almost nothing about how long the real customer sat in the pile.

Licences against a plan

Buying a licence for everyone and running a training day leaves the process untouched, so the same work is simply typed faster. A plan has four parts: measure where the hours actually go, rewrite one process around the tool, write your standards down where a machine can follow them, and decide what a human must approve before anything counts as finished.

What a plan actually contains

Four parts. None of them require you to be technical, and the first one is the one everybody skips.

Measure where the hours go. Take one process and count two things: how many days it takes end to end, and how many people touch it. Then split the steps into those that produce a decision and those that only move information. You will usually find that the elapsed time is almost entirely the second kind. This is unglamorous and it is the whole basis of the plan.

Rewrite one process, properly. Not a pilot bolted onto the existing path. Redesign that one journey as though you were founding the business this morning, and be willing to delete steps rather than automate them. Automating a handoff preserves the handoff. Removing it is the point.

Write your standards down where a machine can read them. This is the part that surprises people. Most of what makes your business trustworthy is unwritten: what a good response looks like, what you never promise a customer, which decisions must involve a named person. As long as that lives in people’s heads, AI cannot follow it and neither can a new employee. Writing it down is a governance exercise you have probably been avoiding for years, and it happens to be the thing that makes automation safe.

Take a purchase approval. In plenty of businesses the rule for who signs off a €6,000 purchase does exist, but it lives in the finance director’s head and a half-remembered email. So the person who wants to buy something asks around, guesses, and waits.

Write the rule down, with its thresholds, the evidence each one needs and the named person who decides, and an assistant can read any request against it and tell that person exactly which gates apply, in what order, and which ones they can ignore. It does not approve anything. It removes the asking around.

One request, checked against a written standard

Someone types into an agent's chat: I want to buy new CRM software for €6,000, who has to approve it? The AI checks the amount against each rule of a written purchasing standard, shown here as an example. Under €1,000, a team lead approves: no, €6,000 is over that. From €1,000 to €10,000, two quotes and then the budget holder: that one matches. Over €10,000, four gates ending with the CEO: no, €6,000 is not over that. So the AI answers that two approvals apply, and that once two quotes are added it will send the request to the budget holder. Quotes are confirmed by procurement, the budget is decided by the budget holder, and the finance director and CEO are not needed at this amount. The AI routes the request. It never approves it.

Decide what a human must approve. Explicitly, per process, before you start. In our own delivery nothing is ever applied automatically: AI proposes, automated checks run, and a person approves. That single rule is what makes the rest of it defensible to a board, an auditor or a customer, and it costs you very little of the speed.

What I would not do

I would not run a company-wide transformation programme. You will spend two quarters producing a plan and the market will have moved underneath it.

I would not start with the process that is most broken, because broken processes are usually broken for political reasons and the technology will not help with those. Start with one that works and is slow.

And I would not let anyone tell you the answer is fewer people. The businesses making real progress have roughly the same headcount doing noticeably more valuable work, having stopped paying qualified staff to re-type things. The ones that cut first tend to remove the people who understood why the process existed, which is an expensive way to learn what the process was for.

Where to start this week

Pick the process your customers complain about most.
Count the days it takes, and the number of people who touch it.
Mark each step as a decision or a handoff.
Write down the standards that process is held to, including the ones currently living in one person’s head.
Decide who approves the output before anything counts as done.

That is a plan. It fits on a page, you can do the first three steps yourself, and it will tell you more about your exposure than any vendor assessment.

If you would like someone independent to go through it with you, book thirty minutes. I sell no software, so there is nothing at the end of it for me to recommend.