Why do so many AI implementations stall?

Most stall between purchase and practice. A business buys licences, runs a launch session, and waits for the productivity to arrive. A few enthusiasts use the tool well. Everyone else tries it on a real piece of work, finds it needs too much setting up, and goes back to the old way under deadline pressure. Six months later, the licences renew and the week looks the same.

The pattern is predictable because nothing about the actual work changed. Implementation, done properly, is the act of changing it. Why AI rollouts fail covers the causes in more depth.

What gets built during implementation?

A working version of one of your processes, running in the tools your team already has. In practice that usually includes:

  • the connections to where the inputs live, such as a shared drive, a practice or case management system, or an inbox;
  • your templates, prepared so they can be filled reliably rather than copied and edited;
  • the drafting rules, written in your house style and terminology;
  • a review step that routes each output to a named person and holds it until they approve;
  • a record of what was produced, from which sources, and who signed it off.

What happens at each stage?

Scoping: map it

We sit with the people who do the job and record every input, template, handoff and check. By the end of scoping we know where the hours go, what the system must do, what the fixed fee is and how long the work should take. Access to systems is agreed at this point.

Build: on your own systems

The system is built on your own documents, templates and language, inside the tools you already run. It is tested on live work from your current workload, not on a demonstration file, so problems appear while we are still there to fix them.

Parallel run, then go-live

The new system runs alongside your current process on the same work, so its output can be checked against what your team would have produced. The team is trained, and it goes live only once they sign it off. The old route is then retired rather than left running beside the new one.

After go-live

Support continues under the subscription, covering the edge cases that only turn up in real use, with monitoring, a monthly performance report and a quarterly review.

What does the business need to provide?

Less than most teams expect. A person who owns the process and can make decisions about it. A few hours from the people who do the job during scoping. Examples of the documents the process produces, good and bad. Access to the systems the process reads from. And someone to review output during testing. You do not need clean data, a written procedure or an IT project running alongside.

How is implementation risk handled?

The main risks are overrun, a build that does not fit the real work, and a process the team quietly abandons. Each is handled by the structure rather than by promises. The fee is fixed and agreed in writing before work starts, and nothing goes live until your team signs it off. Testing on live work and the parallel run catch the mismatch early. And training plus continuing support under the subscription deal with abandonment, which is covered further on the adoption support page.

Data protection is treated as part of implementation, not an afterthought. How client information is processed, and on what basis under UK GDPR and the Data Protection Act 2018, is agreed before anything is connected. Data protection and AI on client work sets out the questions to settle.

What does "live" mean at the end?

Live means the team is doing the job the new way as a matter of routine, on real work, with review and sign-off in place. It does not mean a pilot, a proof of concept or a demonstration that works on sample files. That definition is what your team signs off against at the end of the parallel run. If you want to know which process to implement first, the audit estimates it from your own figures.