The best procurement teams aren’t making the same mistake twice

I received less of a lukewarm response to writing a market update after my last one… so my inner banker continues to have ideas.

A couple of months ago, Satya Nadella posted something that's been rattling around my head. His argument: the companies that win the AI era won't be the ones picking the best model: they'll be the ones that build proprietary "learning loops", where human expertise and AI compound together by tapping into institutional intelligence. They will depend on private evals. Internal reinforcement learning. A knowledge base that gets smarter with every use. He calls them "hill climbing machines."

It's a compelling vision. It also happens to be written by the man who sells Copilot to every enterprise on earth.

I have a different point of view.

I have the privilege of sitting across the table from CPOs, CFOs, and CIOs at companies like Elastic, T. Rowe Price, and Just Eat, and I keep seeing the same thing: enterprises seriously evaluating building AI technology in-house, and concluding it’s not the smart choice.

We’ve been here before

A few months ago I wrote about procurement's quiet ascent and how, with headcount cuts exhausted, the function everyone ignored became the key lever for resilience and growth. The companies that moved fast are in a materially stronger position today. Since then, agentic AI has begun delivering on promises, meaning procurement teams can finally make "do more with less" mean something other than another round of layoffs or cost-cuts. The operational grind that used to swallow the team — triaging intake, chasing suppliers, tracking renewals, and surfacing risk — can now be handed to agents that run it continuously.

However, inside procurement's history is a cautionary tale that's directly relevant to what's happening in AI right now.

Twenty years ago, enterprises looked at procurement technology and concluded it was:

1. simple enough to build in-house

2. too bespoke to be outsourced

They built tailored sourcing systems, custom approval workflows, and proprietary supplier databases.

And what they built became some of the most expensive, brittle, and user-hostile systems in the enterprise stack: routed around daily with spreadsheets and emails, while supplier data sat in silos nobody could query.

Part of what we do at Omnea every day is archaeology. Extracting data from those custom-built systems. Rebuilding data that's become fragmented because businesses have been compensating for systems with off-platform workarounds and human endeavour.

I’ve been impressed to see that history is largely not repeating itself: the enterprise market is remembering these mistakes. Menlo Ventures' latest data shows 76% of enterprise AI use cases are now purchased rather than built.

Do you want to be an AI company?

My undergraduate professors will be quietly pleased that I still remember Michael Porter, and his famous adage: the essence of strategy is choosing what not to do.

Building a proprietary AI learning loop assumes your organisation can run private ML infrastructure, maintain rigorous evals, and sustain a reinforcement learning environment in addition to its actual business… and do it well enough to outperform vendors whose entire existence / raison d’être depends on getting this right.

How many of your internally built systems are genuinely best-in-class today? Not functional. Best-in-class. For most organisations, the honest answer is close to zero.

Here's the real risk. A badly built procurement system is painful but inert: people route around it with spreadsheets. A badly built AI learning loop is not inert. Trained on bad signal, encoding flawed judgment, with no rigorous evals catching model drift, it can create an organisation that is confidently wrong at scale. The cleanup makes the procurement modernisation wave look straightforward.

Which is why the companies best placed to build in-house don't. Anthropic, Synthesia, OpenAI, and Cohere build frontier models for a living and still choose to buy the systems that sit outside their core rather than build them. If they won't go it alone, why would you?

So what should you actually do?

When Neo Finance, Ocado, and Perk invested in transforming their procurement functions, they weren't outsourcing strategic thinking. They were buying into a compounding system: run by specialists, trained on signal from thousands of enterprise customers, improving continuously. The vendor carries the infrastructure risk. The customer gets the benefit.

The same logic applies to AI. Satya is right that the learning loop will define the next era. The question is who should build it, and for most enterprises, the honest answer is: someone whose entire business depends on getting it right.

If you're wrestling with this in your own organisation, I'd love to hear where you're landing. And if the procurement archaeology we do at Omnea is relevant to where you are right now, my inbox is open.

My inner banker just wants to make sure we don't spend the next decade cleaning up the same mess twice!