How to implement AI in your business

Most guides on this subject will tell you to start with a use case. Pick something small, they say. Run a pilot. See how it goes.

That advice is not wrong. It is just incomplete in a way that causes predictable problems six months later.

Businesses that implement AI well are not necessarily doing more than businesses that struggle with it. They are doing things in a different order. The sequence turns out to matter more than most people expect - and getting it wrong is one of the main reasons that early enthusiasm turns into a collection of disconnected experiments that never quite adds up to a change in the business.

This is what we see consistently working with established UK businesses on AI implementation. The tools are rarely the problem. The order is.

Why most AI implementations stall

The typical pattern goes like this. Someone on the leadership team - or several people at once - starts using an AI tool and finds it genuinely useful. This creates momentum. The business starts looking for places to apply AI. Individual teams start experimenting. A few things work. A few things do not. Six months in, there are a handful of people using AI tools they like, a few workflows that have changed, and no clear picture of whether any of it has moved a number that matters.

The problem is not the tools. It is that the experimentation happened before anyone had a clear picture of where AI actually creates leverage in the specific business. Without that, the natural tendency is to implement AI where it is easy, or where someone enthusiastic is already interested - not where it compounds.

56% of CEOs in a recent PwC survey reported zero measurable ROI from their AI investments. That is not a technology problem. It is a sequencing problem.

Start with the business, not the technology

Before selecting a tool or defining a use case, the right question is: where does this business lose time, quality, or speed in ways that repeat consistently?

This is not a complicated audit. In most businesses, a few conversations with the people doing the operational work surfaces the same problems repeatedly. Proposals take too long to produce. Qualification calls are followed by hours of admin. Research that should take thirty minutes takes half a day because the information is scattered. Reports get produced manually every week using the same template.

These are not glamorous AI opportunities. They are also not random. They share a characteristic: they are high-frequency, repetitive, and the output quality depends heavily on who is doing them and how much time they have. These are exactly the conditions where AI creates consistent, measurable leverage.

One professional services firm we worked with identified pitch preparation as the clearest opportunity - a process that consumed significant senior time before every new business meeting. By implementing AI in that single area properly, they reduced the time spent by 42%. That became the foundation for everything that followed. It was not the most exciting application of AI. It was the right one to start with.

The businesses that implement AI well tend to start by mapping this honestly - not to find every possible application, but to find the three or four places where the case is clearest and the impact is most direct.

The sequence that actually works

Once you have a clear picture of where the leverage is, the order of implementation matters considerably.

The instinct is to start with something low-risk. The problem is that low-risk usually means low-impact. If the highest-leverage opportunity in the business is in how you produce client deliverables, that is where implementation should begin - even if it requires more change management than automating something peripheral. Start where the ROI is clearest, not where the change is smallest.

The next mistake is running pilots across several functions at once. The result is that nothing gets implemented deeply enough to actually change behaviour. One area done properly - where the team has genuinely adopted a new way of working and the output has measurably improved - teaches you more than five shallow experiments running in parallel.

Then there is the question of when to expand. Most businesses move too fast here. A tool gets introduced, people start using it, and the business takes that as implementation done. The real test is different: has the underlying process changed? Is the new approach now the default, rather than an option that some people use some of the time? Yes, you need early adoption to get there. But adoption and implementation are not the same thing. Expanding before something is properly embedded means the business ends up with a wide surface area of partial change rather than meaningful depth anywhere.

By the time one area is properly embedded, the business has a much clearer picture of what makes implementation succeed in its specific context - what the team needs, where the friction is, how to measure the result. That learning makes every subsequent implementation faster and more predictable. The sequence compounds. The experiments do not.

What “implementation” actually means

There is a difference between someone using AI well and AI being implemented in the business.

Implementation means the process has changed - not just for one person, but as the default way work gets done. There is a clear owner. The approach is documented well enough that a new person could follow it. There is a way to measure whether it is working. Without those three things, what you have is adoption by some individuals, which is useful but fragile and does not compound.

This is also why AI implementation is harder than most people expect going in. The tools themselves are often the easy part. The work of changing how people work - consistently, not just for the early adopters - is where the effort actually lives.

How to know if it is working

The measure of AI implementation is not tool usage. It is whether something in the business has changed.

At the six-week mark: has the time spent on this process reduced? Has the quality become more consistent? Is the team doing it this way by default, or only when they remember to? If the answer to the first two is yes and the third is not yet, the implementation is not finished.

At the three-month mark: has this freed up capacity that has been redirected to something else? If AI has made a process faster but the time saved has dissolved into the general pace of work, the implementation has value but has not yet changed the business.

The businesses we work with typically see the first measurable result within 8 to 12 weeks when implementation is sequenced correctly. If nothing has changed at that point, the sequence is probably wrong. Not the tools.

Where to start

The most useful first move is not picking a tool. It is getting a clear picture of where the leverage actually is in your business - and that means looking at your operations, not your tool shortlist.

A good starting point is our AI Readiness Assessment - it takes about five minutes and gives you a clear read on where your business is starting from and what the most productive focus would be.

If you want to go deeper, The CEO Guide to Implementing AI covers the full process in detail - including the common mistakes that cause implementations to stall at each stage.

And if you want to talk through where your business specifically is and whether it makes sense to work together, the About page explains how we work and who we typically work with.

Riaz Kanani is the founder of Connected Paths, an AI implementation consultancy working with CEOs of established UK businesses.