Expert Perspective: Lessons on AI Builder Stack

Every component worked, almost nothing reached production

Three years ago I was brought into a programme that had spent eighteen months building AI applications on what looked like a solid foundation: a capable foundation model, a well-configured low-code platform, a clean vector database. Every technical component performed. Almost none of the applications reached production.

The problem was not the stack. It was everything underneath — the design discipline, the governance integration, the reuse logic. We had built each application as if it were the only application in the estate. No shared prompt libraries. No standardised output formatting. No integration design reviews. We had acquired good instruments and produced an unplayable score.

I have seen this pattern in enough programmes since that I treat it as structural, not incidental.

The stack amplifies your discipline, it does not supply it

If your teams do not build traditional software with reuse discipline, your AI platform will not fix that — it will accelerate its absence.

Digital accelerators (D6 of the 6xD framework — the dimension covering tools that compress delivery timelines) inherit the organisational conditions they deploy into. When a unified data and platform foundation is absent, AI applications built on top of it cannot draw on clean data, share components, or be governed consistently. The stack performs. The foundation beneath it is not ready.

Four failure patterns recur across programmes

Four patterns showed up in that programme — and have shown up consistently since.

The proof-of-concept trap. Teams build fast, demonstrate results, receive approval, then discover the prototype cannot move to production without a full rebuild. The demo loop becomes the delivery model. Nothing reaches users at scale.

Governance applied after the build. AI ethics frameworks and risk review processes sit outside the build pipeline. By the time a review triggers, the application is already embedded in a workflow and politically difficult to modify. Governance becomes a post-hoc approval, not a quality gate.

Integration design treated as an engineering detail. How an AI application connects to the data it needs and the people who act on its outputs gets decided by developers under time pressure, not by the people who own the business process. Brittle integrations break when surrounding systems change. Trust collapses faster than it was built.

No one owns reuse. Each team builds its own prompt library, its own vector indexing approach. The organisation pays full price for the same capability five times. The fifth build is no better than the first.

Lay the platform foundations before you scale the build

The decision in front of you is whether to invest four to six weeks establishing platform foundations before scaling your AI build programme — or to scale without them and spend the next eighteen months rebuilding applications that do not meet the quality bar.

Establish shared prompt libraries, shared output formatting conventions, and shared integration patterns before the first production application begins. Wire governance into the build pipeline as a checklist inside the sprint process, not a gate at the end of delivery. Treat integration design as a product design decision and put the business process owner in the room. Create a central AI component library with a named owner and a maintenance cadence.

One path is slower to start and faster to compound. Choose that one.