Enterprise AI: Nothing to Fear, Plenty to Do

A short note on where to start, for people who have to decide

There is a great deal of noise and anxiety about AI at the moment. My view, after thirty years in enterprise architecture and two waves of AI, is that most of the efforts are misplaced. That is not an argument for waiting. It is an argument for acting in the right order to get real value and continue when this bubble, like many others, bursts. Companies that don't will fund experiments and pay for the show.

AI projects and ROI

Managers are under double pressure: hype and agitation from the huge AI bubble, and a “double bind” of stagnating demand and rising costs from fragmented logistics, energy, and raw materials, which makes all traditional levers ineffective. AI is a new lever.

2025 Deloitte survey of 1,854 senior executives shows that 85% increased AI investment, but only 6% got a return within 12 months. In the UK, 3/4 of adopters report productivity gains, but only around 12% report revenue growth.

The problem is not the technology, but the method. Almost every programme starts from the tool and looks for a place to put it: We've decided it will be AI; go and find out what for. Nobody would say that about any other purchase.

A project that starts from the tool hits four walls.

1.       It only finds what is suitable for the tool. All shortlists are almost the same: document review, knowledge search, and a chatbot. That's a map of where data happened to pile up, and it rarely matches the map of where the money is.

2.       The data is wrong. Documents record what happened, not why. AI needs the reasoning, but nobody wrote it down.

3.       The team works at the wrong altitude. Tools can automate tasks while money lives in processes. Speeding up a step that isn't the process bottleneck moves queues.

4.       The core part is missing. The process owners reject ready AI pilots because new operational processes and incident and exception management procedures are not in place. That is not resistance but the most rational decision.

"It's early" – Yes and No

Electric motors were ready in the 1880s, but productivity statistics didn't notice that for forty years, because factories were still built around a central steam shaft. The gain came when engineers rebuilt the floor around the flow of the work and the electric motors.

So yes, until the organisation has rearranged around the new economic situation and new AI instrument, not simply equipped with AI, it is early. That integration competence for each company accumulates only by trying, and it can't be bought later. Waiting is the one option that isn't available.

The Good Start

The first step is money. Money grows in two ways: through higher sales, which require higher demand, or through genuinely paid cost reductions. A process should not be the most productive, but should match the market. Capacity the market doesn't absorb isn't an achievement; it's inventory.

The second step is people. AI agent is not an employee. An agent doesn't tire, doesn't fear consequences, is probabilistic, doesn't learn from its mistakes overnight, and makes the same resonating errors. Replacing people with agents inside an unchanged process removes the knowledge holders and trainers.

Six Questions as a Probe

Six questions, each is worth asking only if the previous one has an answer.

Question	If the answer is…
1. Where does the money move? Which revenue line grows, or which paid cost disappears?	"Efficiency" or "productivity": no channel has been found yet
2. What is the constraint? Which step actually limits throughput or the sale?	Can't name it: not ready to choose anything
3. Will the gain stay with us? If a competitor buys the same tool, what remains?	Nothing specific: real value, but not yours
4. Why does this need AI? Priced against simpler alternatives, not against manual work	Only compared to doing it by hand, the case is unproven
5. What does the new process look like, and what data does it produce? Roles, exceptions, and accountability, agreed by the owner in advance	"We'll work that out after the pilot": the pilot can't convert
6. Which tool?	Asked first: the programme is running backwards

The method works, and most candidates fall at questions one or two, but stopped pilots should still leave value behind: a process now described, a constraint located, a workflow that produces usable data.

The Correct Approach

Not an "AI strategy". The company is integrating AI into its business, not making AI. Integration is a process, so the more accurate term is an AI Integration Strategy.

It fits on a page and has three sections:

·         Where our money comes from and what constrains it

·         What we've learned from attempts so far, including the failed ones and

·         What we are currently doing, with a risk appetite, budget cap and a stop rule.

If those three sections are written honestly, the use cases will select themselves.

The architecture that follows from it, the three-layer AI-Native enterprise, is in the accompanying one-page overview and in my book, Architecture of Intellect.