How to run an AI implementation that actually works

Why most AI projects fail

In our experience of implementing AI in law firms and professional services businesses, the failure mode is almost always the same: a tool is chosen, deployed, and then quietly abandoned because it does not fit the way people actually work.

The tool is rarely the problem. The missing element is the work done before the tool is deployed — understanding the specific problem, mapping the existing workflow, and designing the integration so it reduces friction rather than creating it.

A successful AI implementation is not a technology project. It is a process redesign project with technology as one component.

The right starting point

Start with a problem, not a product. Before any tool is selected, the question to answer is: what specific task are we trying to change, and what does success look like?

Useful questions to ask at this stage:

1. What is the task currently? Who does it, how long does it take, and where does it create friction?
2. What would a materially better version of this task look like? Faster, more consistent, cheaper — what is the specific improvement?
3. Who are the people doing this task today, and what is their appetite for change?
4. What data does the task rely on — and is that data accessible, accurate and well-organised?
5. What does "good enough" look like for an AI-assisted version? AI does not need to be perfect to be valuable — it needs to be better than the current process in the ways that matter.

Scoping the project

Scope is where most implementations come unstuck. The instinct is to start broad — "let's do AI across the whole contract review workflow." The execution reality is always harder than the ambition.

A better approach: start with one clearly defined task within the workflow. Complete it. Prove it works. Then expand.

A well-scoped first phase has:

1. A single use case with a clear definition of what it covers and what it does not.
2. A small group of users — ideally two or three people who are motivated and will give honest feedback.
3. A defined measure of success — not "people are using it" but "contract review time has reduced from X to Y."
4. A timeline with a review point — typically four to six weeks — at which the decision to continue, adjust or stop is made explicitly.

The change management piece

This is the part of AI implementation most commonly underestimated — and most commonly responsible for failure.

Fee earners who are sceptical of AI, or who have had a bad experience with tools sold as transformative and delivered as clunky, will not adopt a new tool because someone in management says they should. They will adopt it when it makes their day easier and they can see why.

What works:

1. Involve the people who will use the tool in the selection and piloting process. Their input improves the implementation and their buy-in makes adoption more likely.
2. Be honest about what the tool does and does not do. Overselling creates the cynicism that kills adoption.
3. Give people time to learn properly — not a one-hour training session but real time with the tool on real work, with support available.
4. Identify and support the early adopters. They become the internal advocates who normalise the tool for everyone else.

Measurement and iteration

An AI implementation is not complete when the tool is deployed. It is complete when the process is working and measurable improvement is visible.

At the four to six week review point, look at:

1. Adoption rate — are the intended users actually using the tool, and regularly?
2. Time impact — has the target task taken less time? By how much?
3. Quality impact — has the output quality improved, stayed the same, or declined in any area?
4. Friction points — what is causing people to revert to the old process, and are those points fixable?
5. Next use case — based on what you have learned, what is the most natural extension of the work?

Most successful AI implementations expand gradually from a narrow first use case. The first implementation is as much about learning how to implement AI in your organisation as it is about the specific task — that knowledge compounds.