AI in the enterprise estate: an architect's view of what is load-bearing and what is décor

Every technology wave produces the same architectural question, and the current AI wave has merely raised its price: of everything we could build, what would actually carry weight in the business – and what is decoration that happens to demonstrate well?

The distinction matters because the two justify entirely different investment logic. A load-bearing capability is one the business would notice failing within a day – it sits in a revenue, cost, or risk pathway, its outputs feed decisions or transactions, and its failure modes therefore need engineering, not hoping. Décor is capability that demonstrates well, photographs well in a board pack, and could be switched off for a month before anyone outside the project noticed. Décor is not worthless – it teaches, recruits, and signals – but it should be funded as education, at education prices, and never allowed to wear the business case of infrastructure.

The uncomfortable pattern across the estates we examine is that AI portfolios are budgeted as though load-bearing and composed largely of décor. The correction is not enthusiasm or scepticism; it is architecture. Four questions sort most initiatives in under an hour each.

One: what decision or transaction does the output feed, and what happens when the output is wrong? Load-bearing AI has a named consumer of its output and a designed answer to error – thresholds, human checkpoints, reversibility, liability. If the honest answer to “what happens when it is wrong” is “someone notices and shrugs”, the initiative is décor, whatever its budget line says. This single question reclassifies more AI portfolios than any other.

Two: does the data foundation exist, or is it being assumed? Most enterprise AI disappointment is data disappointment wearing a fashionable coat. Models are rented; the data that makes them useful – defined, owned, lineage-traced, lawful to use for the purpose – must be built, and in most estates has not been. An AI investment case that does not price the data work is not an investment case; it is a demonstration budget with ambitions. In our assurance work this is the single most common load-bearing assumption sitting unexamined in an “obvious” step.

Three: is the economics of scale actually known? Pilot economics and production economics diverge sharply in this domain: inference costs scale with use, accuracy requirements tighten with consequence, and the human oversight that made the pilot safe becomes a standing operational cost. A capability can be genuinely valuable at pilot scale and genuinely uneconomic at production scale. The time to compute this is before the dependency forms, because the alternative is discovering it from the platform bill.

Four: what does the capability do to the estate around it? AI components are not additive; they are architectural. They create new data flows with new governance exposure, new supplier dependencies with unusual concentration risk, and new integration seams – and seams, as ever, are where estates rot. An initiative assessed in isolation and an initiative assessed as a change to the estate frequently receive different verdicts, and only the second verdict is real.

None of this is an argument for timidity. It is an argument for the ordinary discipline that every previous wave eventually forced on its adopters – usually after the money was spent. The organisations that will be glad of their AI investments in five years are, in our judgement, distinguishable now: they can name the decision each capability serves, they have priced the data foundation honestly, they know their production economics, and they treat each addition as architecture rather than installation. The others are funding décor at infrastructure prices, and the invoice for the difference always arrives.

An hour with each initiative and the four questions above will tell you which portfolio you have. We are happy to be the ones asking.