How much foundation before you move?

Everyone says data is the foundation. Everyone says AI is the future. Two surveys published for 2026 measure what happens in between — and the gap between individual benefit and organisational impact has not moved in a year.

A CPO Outlook 2026 discussion piece · EBG | Network

Data is the foundation. AI is the future. Two statements almost nobody argues with — and between them a question with no settled answer: how much foundation before you move, and how do you find value along the way?

That question is not new. What is new is that the distance between what individuals get out of AI and what organisations get out of it is now measured year on year — and over the past year it has not closed.

Eight in ten. Thirty-seven per cent. Six per cent.

McKinsey’s Global Survey on the state of AI, published in August 2026 and fielded to 1,719 participants between 4 May and 8 June 2026, puts three numbers side by side.

Adoption itself is not the problem. Eighty-nine per cent report regular use of AI in at least one business function, and 44% now say AI is scaling across the enterprise, up from 38% a year earlier. Large organisations are moving fastest: among those above $1 billion in revenue, 54% report scaling across the enterprise, and the share scaling AI agents rose from 27% to 40% while smaller organisations stayed flat.

So more is being deployed, more people are feeling the benefit, and the enterprise-level financial impact has not moved. That is the sequencing question stated as a measurement.

What happens between pilot and production

Deloitte’s State of AI in the Enterprise — 3,235 director- to C-suite-level respondents across six industries and 24 countries, fielded in late 2025 — describes the mechanism, and names it the proof-of-concept trap.

A pilot runs with a small team, cleansed data and an isolated environment. Production requires infrastructure investment, integration with existing systems, security reviews, compliance checks, monitoring and ongoing maintenance. In Deloitte’s words, use cases estimated to take three months can stretch to 18 months when integration complexities emerge, and failures that were learning opportunities in a pilot become business risks in production.

The numbers around that are worth holding together. Only 25% of respondents say their organisation has moved 40% or more of its AI experiments into production to date — while 54% expect to reach that level within the next three to six months. Anyone who has watched a transformation roadmap for a few years will recognise the shape of that second figure.

What the few who get there appear to do differently

Both reports converge on the same answer, from different directions, and it is not a technology answer.

McKinsey describes its high performers as organisations that pursue growth and innovation alongside efficiency, that back deployments with leadership commitment and operational rigour, and — the phrase that matters here — that fundamentally redesign workflows enabled by AI rather than inserting AI into existing ones.

Deloitte measures how rare that is. One third of companies are using AI to deeply transform products, processes and business models; another 30% are redesigning key processes around it; the remaining 37% are using it at surface level, with little or no change to existing processes. And 84% have not redesigned jobs around AI capabilities at all, even though insufficient worker skills is named as the biggest barrier to integration and fewer than half are making significant adjustments to talent strategy.

Which suggests the foundation question may be wider than data. A perfectly clean master data set feeding an unchanged process, run by an unchanged role, produces a faster version of the same output — and no line on the P&L.

Scope, not just readiness

One useful exhibit on this predates the 2026 surveys. MIT Project NANDA’s GenAI Divide, published in July 2025 and still the most procurement-specific mapping anyone has published, sorted initiatives not by how ready the data was but by two other things: how complex the execution is, and how broad the scope. Its examples are procurement examples.

If that pattern still holds, it reframes the question again. Full procurement orchestration failed in that sample not because the data was not ready, but because it was attempted whole.

Two findings that belong to procurement specifically

The first is encouraging. When McKinsey asked where cost reductions from AI had actually landed over the past twelve months, supply chain management came top, ahead of service operations and manufacturing. Whatever is or is not happening at group level, this is one of the functions where the benefit is showing up.

The second is stranger, and lands on procurement more directly than anything else in either report. Nearly a third of respondents — 32% — report that their organisation has decided against buying at least one software product or feature because it could be built in-house using agentic coding tools. It runs from 41% in technology and 39% among healthcare payers and providers down to 17% in the public and social sector. That is a change in what the software category even is, arriving without an announcement, and it lands squarely on procurement.

The two honest cases

There is a real disagreement inside all of this, so both sides deserve their strongest form rather than a straw version.

The case for building the foundation first

Master data, supplier records, category structures and the connections between systems are what every downstream capability reads from. Get them wrong and every model, dashboard and agent inherits the error — at speed, at scale, and now with an explanation attached, which makes it harder to spot. In EBG’s Nordic surveys, 74% still report struggling with poor data quality and not one organisation called itself digitally leading. Deloitte’s proof-of-concept trap describes what happens when that work is deferred: the integration bill arrives later, and larger.

The case for piloting forward

Foundations are never finished, and a programme that waits for “ready” waits while the business stops believing in it. Working use cases reveal which parts of the foundation actually matter — a two-year data programme cannot know that in advance and will over-invest in the parts nobody ends up using. And the evidence that the binding constraint is workflow and role design rather than data quality cuts this way: an organisation that perfects its foundation while leaving 84% of its jobs unredesigned has bought itself a faster version of what it already had.

That is not something an article should settle. EBG’s reading is that it turns on scale, on complexity, on regulatory exposure, and on how much credibility the function has left to spend — but that is a reading, not a finding. The people who actually know are the ones who took one route and paid for the answer.

Which is why this is on the programme

None of the above resolves into an answer an article can hand over. It is the kind of question that gets tested when people running comparable problems compare what actually happened inside their own organisations — which is what the afternoon of 14 October is for.

It is not the only table on this question. The morning roundtables are hosted by organisations that see the same pattern from a different angle — across many procurement functions at once, rather than from inside one. Four of them sit directly on what this piece has been about: