Many businesses have experimented with generative AI, yet few projects reach production. This article shares four lessons from more than 40 AI projects delivered by Zentrova.
Over the past two years, almost every organisation has experimented with generative AI: an internal chatbot, a document summariser or a writing assistant. Yet according to Zentrova's Data & AI division, only around a third of pilots actually reach production and deliver measurable value. So what makes the difference?
Lesson 1: Start from the process, not the model
Successful projects always begin with a specific process that is expensive or time-consuming, such as credit application handling, support requests or recurring reports. When the goal is quantified from the outset, choosing the right model and architecture becomes far more straightforward.
Lesson 2: Data determines quality
Large language models only perform well when they are given the right context. Most of the effort in our projects goes into cleaning, permissioning and organising the enterprise knowledge base so that retrieval-augmented generation can work. A messy document store produces messy answers, however capable the model.
Lesson 3: Measure and govern continuously
- Build a standard evaluation set before go-live and rerun it after every change.
- Track the rate of incorrect answers, the cost per request and response times.
- Give users a quick way to flag results that are not accurate.
- Apply sensitive-data filtering and log every interaction.
Lesson 4: People remain at the centre
Generative AI works best as an assistant that helps employees work faster and more accurately, rather than as a full replacement. Projects with a structured user training programme see adoption rates three times higher than those that simply hand over a tool.
The question is no longer whether to use generative AI, but where to use it to create the greatest value with risks kept under control.
A suggested roadmap
We recommend a three-step approach: pick two or three high-value processes and pilot them over eight to ten weeks; build a shared data foundation and AI governance framework; then scale gradually to other departments. This keeps costs under control and builds internal confidence before any large-scale investment. A short pilot with a clear success metric is usually worth more than a long list of ideas.
Zentrova's Data & AI division now has more than 400 specialists, ready to support organisations from readiness assessment through to deployment and operations.
