Five Supermarkets Stopped Paying Twice for the Same Answer

So, What Actually Happened?

Tuesday, and the number I keep coming back to is not a valuation. Gartner says 60% of organisations will fail to govern AI by 2027, and the reason it gives is not budget. It is that people will not let go of their data. We scanned 190,000 articles this week so you don't have to, and two more things landed in the same window making the same point from opposite ends. Accenture took majority control of Dragos in a $4.1 billion deal, and Dragos immediately closed two acquisitions of its own. Then five supermarket groups who fight each other for the same shopper funded one shared food-risk index instead of five private ones. Both are ownership decisions wearing technology clothes.

The Bottom Line: What is breaking AI programmes right now is not the model and not the money, it is that nobody will say out loud who owns the data underneath.

The Tracks That Matter

1. Gartner Says Culture Kills AI Governance, Not Budget

Gartner's forecast is that 60% of organisations will fail to govern AI by 2027, and the qualifier is the whole story: the 60% are specifically the ones that never dealt with the cultural side of data ownership. Underneath the forecast sits a March 2026 survey of 223 data and analytics leaders, where cultural resistance beat funding constraints 60 to 40 as the reason governance programmes fail. Put that next to TechTarget's reporting that most companies now run more than 100 SaaS products, each quietly collecting a slice nobody else sees, and the word ”culture” turns concrete. Governance means a named person hands over a system they currently control. No procurement cycle has ever fixed that.

Here's what works: Before buying a governance platform, list the five systems your AI needs and the person who would have to give up control of each. That list is the project.

2. Accenture Took Dragos, And Dragos Took Two More

Accenture acquired a majority stake in Dragos in a $4.1 billion transaction, and Dragos closed its NetRise and runZero acquisitions straight after, so three deals in industrial cyber security resolved into one structure inside a single news cycle. The part worth sitting with: the firm that inspects your plant floor and the firm that sells you the remediation programme are now the same P&L. This lands in the same week that CISA guidance on AI in operational technology started circulating among grid operators, which means the rulebook and the referee arrived together. Independence in OT assessment just became something you have to write into a contract rather than something you get by default.

Here's what works: Check whether your OT assessor and your remediation supplier now share a parent. If they do, buy the assessment from somebody who cannot bid on the fix.

3. Five Rival Supermarkets Bought One Risk Model

Helios AI raised $3 million through W23 Global, and the backer list is the news: the venture arms of Tesco, Ahold Delhaize, Woolworths, Shoprite and Empire, five groups that compete on the same shelf. What they bought is a Hunger Exposure Index scoring countries nought to 100 on climate risk, fertiliser and energy exposure, foreign-exchange constraints and cereal-import dependence, with anything over 70 flagged critical. The interesting bit is not the model, it is the arithmetic the buyers did. Sourcing risk is not a competitive advantage any of them can hold alone, so they stopped paying five times for the same answer. Watch for that shape wherever the input is shared and the margin is not.

Here's what works: Name one risk your competitors face identically. If you are funding a private view of it, you are paying a premium for information nobody differentiates on.

Quick hits:

• A data-centre builder is going public against a contract book. Nscale is seeking an IPO to meet $103 billion of data centre contracts, which makes the listing a financing event for obligations already signed rather than a victory lap.
• A care provider drew the line at the data, not the tool. Anglicare Victoria blocked its most sensitive records from any open AI platform and ran a three-month staged pilot before scaling, which is the governance posture Gartner says most organisations skip.
• The bottleneck moved to the cable. Marvell will demonstrate 2nm optical interconnect for AI data centres at ECOC 2026, a reminder that the constraint is now moving between racks rather than inside the chip.

Signal vs. Noise

🟢 Signal: who owns the data. Data governance and risk management took the firmest hold on this week's real events of anything we track, and Gartner's March survey of 223 data and analytics leaders says why: cultural resistance, not funding, is the top reason governance programmes fail. Most coverage still files governance as a software category, which is the budget half of a problem that was never about budget.

🔴 Noise: ”AI agents” as a label. The phrase pulled one of its loudest weeks and lost more of its grip on what actually happened than anything else we watch. Meanwhile the real deployments this week described themselves by workflow: fund administration, clinical notes, plant-floor inspection. The label is going slack exactly as the practice tightens.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

A private-equity firm with $110 billion under management is running 1,600 live AI projects across its portfolio, a trust-administration platform called Quantios shipped an agent and then got sold, and the data-engineering press quietly started renaming ETL to ”extract, context, link”.

Three different desks, three filings, no overlap. The private-capital wires wrote up Hg extending its three-year Anthropic collaboration across its portfolio, where those 1,600 projects carry $260 million of budgeted EBITDA impact, five times the count of early 2024. Quantios launched its first agent in March, ahead of Hg's exit to Vista Equity Partners. And the plumbing press is rebuilding the pipeline around context and lineage rather than movement. Read them together and the same fact appears three times: the people specifying how AI gets wired into a business are increasingly the owners, not the builders. AI stopped being an experiment on the CTO's slide and became a line in somebody's value-creation plan.

What changes today: find out whether your AI roadmap appears in your owner's or board's value plan. If it does not, somebody is about to write it for you.

By The Numbers

• 60% of organisations will fail to govern AI by 2027 — Gartner's forecast, aimed precisely at the companies that treated data culture as somebody else's problem, and backed by a March 2026 survey of 223 data and analytics leaders.
• Accenture took majority control of Dragos in a $4.1 billion deal — with NetRise and runZero closing straight after, so industrial cyber security consolidated in one motion rather than three.
• Hg is running 1,600 live AI projects worth $260 million of budgeted EBITDA — across roughly 60 businesses, five times the project count of early 2024, which is what AI looks like once it reaches a value-creation plan.
• Nscale is seeking an IPO against $103 billion of data centre contracts — a contract book that makes the listing a funding requirement, not a milestone.
• Most organisations now run more than 100 SaaS products — each collecting data that mostly never leaves it, which is the honest shape of the ”data foundation” everyone says they already have.
• See what's rising across AI and data this week →

Deep Dive: Nobody Wants To Give Up Their Listing

I am a child of the Commodore 64, and I taught myself to write a listing to keep track of my vinyl. The collection outgrew one listing, so it became three, and from that moment I could not sort everything in one go. That is where most companies are sitting right now, except the listings number a hundred and each one has an owner.

The silo is a person, not a system

Every one of those hundred SaaS products was bought by somebody who needed it, defended it in a budget round, and is now measured on it. When governance asks for the data, it is not asking a database. It is asking a manager to make their own work legible to people who might disagree with it.

That is why the tooling keeps failing

A catalogue can map a hundred sources. It cannot make a regional director volunteer the one field that makes their numbers look worse. Gartner measured this and found culture outweighing money 60 to 40, which is not a software finding at all.

The owners have started answering instead

When a private-equity firm puts 1,600 AI projects on a value-creation plan with EBITDA attached, ownership of the data stops being negotiable inside the business. That is a blunter fix than any governance programme, and it is arriving whether the middle layer likes it or not.

What Actually Works

1. Name the person, not the system: for each source your AI touches, write down who loses control. That name is your actual dependency.
2. Trade something for the data: give the owner a report, a metric, or a decision right in exchange. Governance without reciprocity is expropriation.
3. Start with the source nobody defends: the boring, unowned system gives you a working pattern before you spend political capital.
4. Put it in the value plan: if AI is not attached to a number somebody is measured on, it stays optional, and optional loses every budget round.

Three listings you can actually sort beat one listing nobody will hand over.

What's Coming

Governance Starts Getting Sold As Change Management

The argument that actionable data and governance are what unlock agentic AI is turning up in operations coverage, not compliance coverage. Once the named failure mode is culture, the fix gets priced as advisory work rather than licences. Expect governance software vendors to start buying advisory shops within two quarters, because that is the only way their category sells against a Gartner finding like this one.

The Grid Gets Its AI Rulebook First

CISA guidance on AI in operational technology is landing with power operators before anyone else, which is the usual order: the sector where failure is visible goes first. Water, rail and hospital estates inherit the same language within a year, usually copied rather than rewritten. If you run anything with a physical consequence, read the grid version now and save yourself the panic later.

Production Readiness Replaces The Demo

Vendors are publishing AI production-readiness checklists aimed at infrastructure teams rather than innovation teams, and vendor content follows buyer questions with a lag. The question in the room has moved from what the model can do to what breaks at 3am. Procurement documents will catch up next, and the demo-led sale gets noticeably harder.

For Your Team

Wednesday's meeting prompt: ”Pick the AI project we are most committed to. Now name every system it needs data from, and the person who currently controls each one. Have we actually asked them, or have we been assuming access we do not have?”

Share-worthy stat: Gartner expects 60% of organisations to fail at governing AI by 2027, and the cause it names is culture, not funding. Its March survey of 223 data and analytics leaders put cultural resistance ahead of budget 60 to 40. Everybody is buying tools for the smaller half of the problem.

Go deeper: Track where data governance and AI risk are actually moving →

The Track of the Day

”The objective is to direct human attention toward exceptions, relationships and decisions requiring expertise.”
Intralinks, on automating processes without automating judgment

Automate the typing. Leave the arguing about who owns what to the humans, because that is the part the machine cannot settle and the part that decides whether any of this ships.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: September 22, 2026 | Curated by Yves Mulkers @ Ins7ghts