Agents, platforms, applications or not yet: which road to AI?

Some procurement teams are building their own AI agents. Others have tested AI inside their applications and put the effort into the shared data platform instead. These are not opposites: an agent is only as good as the data it can reach, and a platform only earns its cost when something uses it. The more useful question is who is allowed to build, and how each company reasoned its way to where it is.

A CPO Outlook 2026 discussion piece · EBG | Network

Could someone in your procurement team build an AI agent next week? In most organisations the honest answer says less about the technology than about permission: who is allowed to build, who has to be asked, and who carries the result afterwards.

That is the ground the Day 2 panel at CPO Outlook 2026 stands on. Its three panellists have reasoned their way to different starting points: from inside procurement, from between the business and IT, and from within IT. No study names a winning road for AI in procurement. What the evidence does show is how fast the ground has moved, and how unevenly.

Where the Nordics actually stand

EBG’s Xperience 2026 survey asked 48 senior procurement professionals, one per company, how far AI has come in their function. More than half, 55%, are exploring. 38% have basic adoption in specific areas, such as contract analysis or spend data. Only 2% describe AI as embedded in how they work and decide.

The same respondents were asked where meaningful AI adoption in a large organisation has to begin. No single answer dominated.

The free-text answers say the same thing from different angles. One respondent added “individual enablement and citizen development in parallel to longer-term projects” to this question. Another, asked what they most wanted to discuss, wrote: “Don’t believe that there is one perfect way of rolling AI out — pros/cons to all approaches.”

A year is a long time in this field

Many of the figures still quoted about AI in business come from 2025. Set next to what has been published in 2026, the change in a single year is a finding in its own right.

Large organisations have moved quickly: McKinsey’s 2026 survey found 40% of respondents from companies above $1 billion in revenue scaling AI agents, up from 27% a year earlier, while smaller organisations stayed flat at 22%. Gartner’s 2026 Hype Cycle for Agentic AI places agentic AI at the Peak of Inflated Expectations: only 17% of organisations have deployed AI agents, yet more than 60% expect to within two years. Nordic procurement’s own read of its maturity has barely moved.

EBG’s reading is that the figures will keep moving, quarter by quarter. The questions behind them are what stays.

Build or buy has moved too

In 2025, MIT NANDA found that internal AI builds succeeded about one-third as often as tools bought from specialised vendors and partnerships. In McKinsey’s 2026 survey, 32% say their organisation decided against buying at least one software product or feature because it could build the functionality in-house with agentic coding tools. The two studies measure different things. In EBG’s reading, building in-house no longer looks like the outlier it did a year ago.

Four roads, and the decision behind each

The panel frames four roads: agents built in the function, AI through a shared data platform, AI inside the applications you already have, or deliberately not yet. Large organisations often run more than one at once. None of them is a technology choice alone. Each rests on a decision someone made, or did not make.

Read together, the research suggests the choice of road may matter less than whether the questions behind it are answered. In 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, and gave three reasons: escalating costs, unclear business value and inadequate risk controls. None of them is about which road was taken.

Who is allowed to build

The people doing the work, or IT? The programme text for the panel sets out both failure modes without a verdict. When the business moves faster than IT, compliance and security tend to be solved last, going backwards. When IT sets the pace, the business may build around it. The 2025 MIT NANDA research, as reported by Fortune, has a name for the second pattern: “shadow AI”, meaning unsanctioned tools used outside official channels.

The appetite to build is there. At Xperience in Stockholm, leaders were asked what they would do with eight hours won back next week. Five of the eleven who answered said they would put the time into building agents. Whether that can happen is mostly decided elsewhere: by who grants access, who funds it and who signs off. In the Xperience survey, 36% named unclear mandate as a reason action slows down, even when the opportunity is clear.

Who owns the data, and who keeps it correct

Data readiness is not one state across a function. At Xperience in Stockholm, tables placed six AI use cases on a grid of problem clarity against data foundation. Contract analysis, RFQ drafting and supplier discovery tended towards ready to deliver. Supplier risk monitoring and tail spend tended the other way: clear problems, weak data.

If readiness differs by use case, the road may differ by use case too. And someone has to own the data each road runs on, after the build is finished.

And who answers for it

The third question follows from the first two. When the person who built an agent moves on, who owns what it produces? Research on automated decision aids suggests this is not only a governance question. In Parasuraman and Manzey’s review, people told they would have to justify their performance checked the aid’s recommendations more often, though the authors call that evidence not fully conclusive.

Between two keynotes

The panel sits between two keynotes that come at AI from different directions. At 13:15, Hilde Solheim, Director at Deloitte and previously CPO at Eviny, asks what AI changes in procurement and what it does not. Her question includes what still does not stick unless the foundations are repaired. At 14:25, Mauricio Bozzo, Group CIO of Arrive, closes the summit with “Lean in”, on the company you get when you stop piloting.

The panel does not have to settle that. It puts three organisations on the stage that have reasoned their way to different roads, and asks the room where it stands.