Six foundations for lasting AI value

What leaders need to get right to move beyond promising pilots.

If your board asked you tomorrow whether your organisation could turn its next AI pilot into lasting business value, what evidence could you show them?

That question sits at the heart of our recent conversations with senior business and technology leaders across financial services, insurance, telecoms, leisure, hospitality, the public sector and industrial supply chain. Most now have AI in daily use. Far fewer can show that it has materially improved margin, service quality or throughput.

The difference is rarely the model. Capable AI is easy to acquire, for you and for every competitor.

What sets organisations apart is the architecture around it: who owns the outcome, what the AI knows about your business, how it reaches your systems, how it is controlled and how your people work with it.

So it is worth asking where you stand. Are you adding AI one tool and one pilot at a time, starting from scratch each time? Or are you building the foundations that make every implementation easier, safer and more valuable than the last?

Drawing on our paper From AI Activity to Enterprise Advantage [link], here are six building blocks of an enterprise intelligence architecture, and the questions every leader should be asking of each.

Six building blocks to get right

1. Does anyone own the outcome, or just the tools? (Strategy and ownership)

AI makes plausible ideas easy to generate, so investment tends to follow the most impressive demonstration rather than the biggest opportunity. Without a clear strategic roadmap and an accountable owner, teams end up pursuing many directions at once and lose sight of what the business is trying to achieve.

Every opportunity should compete against your strategy, the evidence behind it and the capacity it would consume, with "stop" and "not yet" as genuine answers. Ownership also means deciding where released capacity goes: lower cost, better service, growth or different work. Asking people to do more with the time saved is not a strategy.

The question shouldn't be, "who is using AI?" It should be, "who is accountable for the outcome it is meant to improve?"

1. Does your AI know what your best people know? (Proprietary knowledge)

The knowledge competitors find hardest to copy is often the hardest for AI to use, because much of it was never written down. Experienced people know which rule really applies, which exception matters and when a correct-looking answer should still be challenged. A model knows none of this unless you give it.

Start with one consequential workflow and ask what an experienced colleague knows that a new starter would not. Capture what can be documented, route tacit judgement to the people who hold it, and give each important source an owner accountable for its accuracy. You don't need every source ready at once: prepare the minimum viable context the workflow needs, and extend it as you learn.

1. Are you starting with the outcome, or the technology? (Outcome-led use cases)

"Use AI in our contact centre" describes an activity. "More people getting their problem resolved first time" describes an outcome, and it sets both the scope of the work and the evidence of success.

Treat that outcome as a hypothesis until there is evidence the problem matters enough to solve. Then redesign the work rather than adding an agent to each step of today's process; some steps may disappear, and others may need no AI at all. Agree the baseline and what "good enough" looks like before you choose the technology.

The question shouldn't be, "where else can we use AI?" It should be, "if this succeeds, which business measure will improve?"

1. Can your AI reach the systems that matter, and survive the next model? (Architecture and orchestration)

Much of your most valuable context sits in systems built long before AI, and they rarely need replacing first. Map which systems hold the records and rules a workflow depends on, how reliable they are and how they can be reached safely, then connect them through maintained integrations. AI can now help with that discovery, reading legacy code and data structures for specialists to check.

Orchestration matters too: how AI, people, tools and systems hand work to one another across the workflow. Build so that the next team inherits tested integrations, evaluation sets and guardrails rather than starting again. Models will keep changing, so keep what differentiates you, such as your context, workflow logic and evaluation evidence, under your control and portable.

1. Is trust designed in, or bolted on at the end? (Governance and control)

"Human in the loop" is too vague to design around. For each workflow, decide what AI may recommend, decide or act on, which systems it may reach, when a person must step in and who is accountable for the result.

AI is different from conventional software because its behaviour can change when nothing in your code has: a model update or a new connected tool can shift the results. Evaluation against the real task therefore continues for as long as the capability runs. Build security and policy requirements in from the start, extend autonomy only as evidence grows, and make sure someone owns quality and cost per successful outcome after go-live.

1. Are your people shaping the AI, or just using it? (People and operating model)

People bring the judgement, expertise and trust that turn working technology into adopted practice. The people who do the work should help decide where AI assists, what is automated and where human judgement or relationships create the value.

Where roles will change, be open about it and redesign them with the people in them; they hold knowledge the business can least afford to lose. Make sure learning carries forward too. A stopped pilot can still sharpen an evaluation method or rule out an architectural choice, but only if someone reviews it and changes what happens next.

To summarise: strategy tells you where to aim; knowledge and architecture determine what AI can do; governance and people determine whether it is trusted, adopted and repeatable. Weakness in any one block limits the value of the others.

Putting the questions to work: five steps

These six questions above are not a maturity score. Used well, they give you a structured way to see what is backed by evidence, where gaps could limit value or scale, and which conversations need to happen next.

Here’s how to get the most from them:

1. Pick one opportunity
   Choose a single consequential use case, such as the pilot you most want to scale. Answers are more honest, and more useful, when they describe something real rather than the organisation in general.

2. Answer with evidence, not intent
   For each building block, ask whether it is in place, partly in place or not yet there. Only call it “in place” if you could show someone the proof today. Good intentions and planned work count as “partly”.

3. Bring several perspectives
   Ask business, technology, data and risk colleagues to answer the questions separately, then compare. Where people answer differently, that disagreement is often the most valuable finding of all.

4. Start with the weakest block
   Because the building blocks work as a system, the weakest one usually limits the rest. Strength elsewhere will not compensate for it, so that is where effort is likely to pay back first.

5. Revisit them as you progress
   Ask the questions again as the opportunity moves from idea to pilot to production. Noting your answers each time gives you a simple record of progress, and a clearer case when it is time to decide whether to scale, strengthen or stop.

From activity to enterprise advantage

The six building blocks are explored in depth in our paper, From AI Activity to Enterprise Advantage, alongside the experiences of the senior leaders we spoke to.

Download the whitepaper here

If you’d like to explore your answers with our team, or compare notes with other enterprise leaders facing the same questions, join us at one of our upcoming workshops or AI breakfast roundtables.

You can find out more about upcoming events here.