On the seventh of May, SAP closed its acquisition of Reltio. The press release used a phrase Walldorf would not have written five years ago: making enterprise data "AI-ready". I read it twice. The biggest ERP vendor in my world had just paid to admit something I have been telling clients for years. The data is not ready. It was never ready. And the foundations gap is now expensive enough that SAP bought a master data management company to plaster over it.

That single transaction was the loudest signal of the fortnight. It was not the only one.

The market started saying the quiet part

In the same two weeks, I watched the market converge on the same thesis from six directions at once.

McKinsey opened a CEO briefing on 20 May with a sentence that should make every CIO who has signed an agentic-AI contract pause. "AI has proven its potential, but value remains hard to scale." That is consultancy-speak for "the pilots work, the production estate does not". I have been saying the second sentence to clients for the last two years; McKinsey has just published the first.

Harvard Business Review Analytic Services released two Pulse Surveys in the same week: "Solving Agentic AI's Data Infrastructure and Telemetry Needs" and "Bridging the Readiness Gap to the Agentic Enterprise". When HBR runs two readiness surveys back to back, I take it as the editorial board deciding the conversation has moved from "should we" to "we cannot, yet".

TLDR Data put a number on it on 18 May. Five of six companies, they reported, lack the data foundation required for agentic AI. They are spending the money anyway. That is the number I had been waiting for someone to publish for a long time.

Verdantix, the analyst house I read most carefully on industrial asset management, ran a piece on 19 May with a title that could have come from my own consulting deck: "Agility Fails Where Operations Aren't Digitized." A week earlier, an operator had filed a claim against an AI vendor that proved the analyst right in the messiest possible way.

What it looks like when it goes wrong

Chaac Pizza Northeast operates 111 Pizza Hut sites across the eastern United States. On 6 May, the company filed a $100 million claim against Pizza Hut, alleging that the Dragontail AI delivery-management system collapsed restaurant operations after it went live. Set aside the headline figure for a moment. American claims run large by default; that is not what makes this one matter. Orders ran late. Delivery windows were missed. The economics at site level fell apart. Pizza Hut, the claim alleges, refused to modify the system even as the operational evidence piled up.

That paragraph is not a hypothetical. It is a court filing. A named claimant, a named system, a named set of operational failures. I have been waiting for one of these to land in public for some time. The evidence in a filed case like this is the sort of evidence the consultancy decks have been politely declining to print for three years.

The AI probably worked. It almost always works, in the demo. What broke is what I have watched break in every sector I have worked in: an optimisation layer installed on top of an operating model nobody had agreed, against asset and process data nobody had cleaned, with ownership boundaries nobody had defined. The technology met the operation, and the operation lost.

The pattern is not restaurant-specific. Substitute a hospital trust for the operator, a national grid for the multi-site footprint, a defence estate for the asset hierarchy. The mechanics are identical. The dollar figure was American; the failure mode is not.

Why I built a business on this

I founded Brainwave Asset Intelligence in April, after 26 years of delivering Enterprise Asset Management in defence, healthcare, and critical national infrastructure. I did not start the business as a forecast about AI. I started it as a diagnosis from the data I had been staring at the whole time.

The same pattern, in every implementation. The optimisation layer outpaces the data layer. The promised value sits behind a foundation that nobody resourced. The operating model is half-defined and politically contested. The technology team owns the system. Nobody owns the data. Nothing owns the outcome.

That is the conviction I founded Brainwave Asset Intelligence on. Companies need good data, defined ownership, and a working operating model before they can extract the benefits of AI. Not in parallel. Before.

This fortnight, McKinsey said it. HBR said it. TLDR Data put a number on it. Verdantix said it in plain prose. SAP paid for a master data management company to be able to say it without admitting it. And an operator in the United States filed the case.

The implication

If you are responsible for an asset-intensive estate and you have an AI initiative on the table, my honest view is that you have an 18-month window before the foundations gap stops being free to ignore.

It stops being free for three reasons I keep coming back to. The regulatory clock is shortening, not lengthening: on 7 May the EU cut the synthetic-content watermark grace period from six months to three. The competitive clock is shortening too: the operators who do the foundations work this year will be the ones publishing case studies next year, and the ones who skipped it will be the ones explaining themselves to their board. And the litigation clock has started, with the Dragontail case as exhibit A.

The work is unglamorous. Asset hierarchies that someone owns. Master data with a definition that survives a stress test. An operating model that names who decides what, when, and against which evidence. None of that is in a vendor demo. All of it is the difference between an AI investment that compounds and an AI investment that ends up in court.

The market just spent two weeks telling you the same thing. The question is what you do with the next two.

Foundations Check: would SAP–Reltio actually work in your organisation?

Here is the question I have been turning over since the SAP–Reltio announcement.

The capability is real. Reltio embeds master data management with AI-driven entity resolution inside SAP Business Data Cloud, with the promise of a "golden record" across SAP and non-SAP sources. From 26 years of watching MDM platforms land in client estates, I also know it is a tool without a hand in most of the enterprises that will buy it.

Master data management is not a software problem. It is a governance problem with a software component. When I look at SAP–Reltio, I see four questions the customer must answer before they get a penny of value: who owns customer master data, who owns asset master data, who arbitrates when they disagree, and who can override the golden record when business reality changes.

I have rarely seen those four questions answered with clean ownership in writing. Ownership is implicit. Arbitration is by escalation. Override is by spreadsheet. SAP–Reltio in that environment will become another expensive system of record nobody trusts, sitting next to the three you already have.

SAP has done the right thing. By paying for the capability, it has raised the implicit price of the work the customer still has to do. Buying the platform does not buy the operating model. It makes the absence of the operating model louder.

The question I would put to your steering committee is not "when do we implement". It is: "who in this organisation owns the golden record, and what authority do they have when the answer disagrees with the business unit?". If you cannot name that person today, I do not think the platform will produce value next year.

Worth reading

That is my read of the fortnight. Yours might differ.

Reply with the foundations question your team is most avoiding. I read every reply.

Alex Brain Founder, Brainwave Asset Intelligence

LINKEDIN PUBLICATION SETTINGS

Field Value
Newsletter Foundations First
Issue 01
Publish date Friday 22 May 2026
Hero image foundations-first-issue-01-1920x1080.png (from 1-LinkedIn/visuals/News Letter/)
Posted from Alex Brain personal profile
Cross-post (Fri 11:00 BST) Teaser feed post: "Five of six enterprises are spending on AI agents they cannot run. This fortnight the market said so. My read in Foundations First Issue 01 ↓ [link]"

POST-PUBLISH CHECKLIST

  • [ ] Hero image uploaded (1920×1080)
  • [ ] Title and subtitle set in LinkedIn UI
  • [ ] Body pasted, headings render as H2 in preview
  • [ ] All five Worth Reading links clickable
  • [ ] Standfirst readable in mobile preview within first 140 characters
  • [ ] Schedule 11:00 BST teaser feed post
  • [ ] Hand-curated DMs to 10 to 15 warm-network contacts after publish
  • [ ] Reply to every comment within 24 hours