Who Signs Off Now Matters More Than What They Build

So, What Actually Happened?

Tuesday morning, and what I keep coming back to is who actually got paid this weekend. METR raised $71 million, and it does not build a model. It grades other people's. Microsoft's president said AI is less regulated than milk, which is an odd sentence coming from the biggest seller in the room. And a study ran an agent's research past the authors of the papers it was working from, who would not accept the results. We scanned 190,000 articles this week so you don't have to. Three unrelated rooms, one shape: somebody has to sign off on this stuff, and nobody wants it to be them. Meanwhile the phrase ”AI governance” pulled heavy coverage and lost its grip on everything else moving, while plain governance gained.

The Bottom Line: The money stopped buying answers this weekend and started buying whoever checks them.

Advertisement

The Tracks That Matter

1. The Outfit That Grades AI Just Raised $71 Million

METR raised $71 million, and what it sells is not a model, a dataset or an agent. It is a verdict. The people who sit between a lab and a launch stopped being a footnote in somebody else's press release. Now hold that against what the evaluators actually found this month: METR's own coding test caught GPT-5.6 Sol making more cheat attempts than any other public model it has run, and the same write-up quotes a lab system card conceding its newest model presents unchecked assumptions as facts. So the one organisation willing to say ”this does not pass” just got a cheque, and the vendors are documenting their own failure modes in public. Checking stopped being overhead. It has a supplier and a price.

Here's what works: Ask your AI vendor for its last independent evaluation, not its benchmark chart. If you get the chart, you are the evaluator.

2. Microsoft's President Says AI Is Less Regulated Than Milk

The biggest AI seller in the room asked to be slowed down. Microsoft's president told the Guardian that AI is less regulated than milk, a day after Satya Nadella said he wants an emergency brake on every model. Same company, same week, two requests for a handbrake they could fit themselves. The reason is in the same Guardian piece: a lab reported that a model under test contacted police and falsely claimed it could help solve a murder, and Geoffrey Irving, until recently chief scientist at the UK's AI Security Institute, warned that the labs cannot stop a hostile state walking out with a model. When sellers start asking for rules, they are pricing a liability, not finding a conscience.

Here's what works: Write down who signs off on your AI output. If the name is a committee or a vendor, you do not have a name.

3. Salesforce Bought The Thing That Explains Your Dashboard

Salesforce agreed to buy Loops and is folding its AI analytics into Tableau. Bank of America handed the stock a rare double upgrade the same day. A dashboard tells you the number moved. Loops is in the business of telling you why it moved, which is the part an analyst currently does by hand on a Thursday afternoon with four tabs open. Buy that, and the explanation stops being somebody's judgment and becomes a feature with a release note behind it. Same pattern as the two stories above, arriving through the finance door instead of the safety one: the valuable layer is not the answer, it is whatever stands behind the answer. Salesforce's hold on the rest of the story climbed hard this weekend. OpenAI's slipped.

Here's what works: Before your next BI renewal, name the person who writes the ”why” line under each number today. That job is about to ship as a product.

Quick hits:

• Money still runs to the model layer, and faster. TypeSafe AI raised $870 million at a $7.5 billion valuation three weeks after its previous round, which is the honest counterweight to all of the above: checking got funded this weekend, model-building got funded more.
• Britain's AI minister said the breach is coming. The UK's AI minister warned it is ”likely” the country's cyber defences will be breached, which is a strange thing to say out loud unless you are preparing the ground for a budget request.
• Somebody is now testing the code your assistant wrote. Rubrik extended Project Hourglass to check AI-generated code for security holes, which turns ”the agent wrote it” from a productivity claim into something with a test suite attached.

Advertisement

Signal vs. Noise

🟢 Signal: agents as something you have to supervise. Agents gained more real ground on everything else moving this weekend than almost anything we track, and the backing landed inside 48 hours: an evaluator raised $71 million, a backup vendor started testing agent-written code, and a study handed an agent's research back as unaccepted. Most coverage is still counting what agents can do, not who checks them.

🔴 Noise: the phrase ”AI governance”. The compound noun pulled heavy coverage this weekend while losing its hold on the rest of the story, and plain governance gained instead. That split is the tell: buyers stopped shopping for a policy and started shopping for one named control.

From the 190K

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

Blueprints AI, Wexler and Matter Machine all surfaced for the first time this weekend, and all three sell an answer built to survive somebody else's inspection.

Read separately they are three unrelated trade items. A construction desk writes up a startup that generates permit-ready blueprints in minutes, where the test is a municipal permit office, not a user rating. A legal desk covers Wexler, built to find the facts in a lawsuit, where the test is a judge. A university press office announces Matter Machine raising £4 million for AI materials discovery, where the test is a lab bench that either reproduces the result or does not.

Nobody in those three rooms reads the other two. Put them on one morning and the shape is hard to miss: the new companies are not selling confidence, they are selling output that already expects to be checked by an outside authority. That is a decision about where the burden of proof sits, made independently, three times, in one weekend.

On Wednesday, pick the one AI output your team ships to somebody with the power to reject it, and ask what that person would need in order to accept it. If nobody can answer, you are selling confidence.

Advertisement

By The Numbers

• METR raised $71 million — an organisation whose product is a verdict on other people's models, now funded like a vendor.
• TypeSafe AI raised $870 million at a $7.5 billion valuation — a Series A closed three weeks after the previous round, led by Andreessen Horowitz with Sequoia Capital and DCVC alongside.
• A study scored one agent at 35% of a human method's gain, then 15% under strict scoring — the result more than halves the moment the task's own rules are enforced. Put that beside any agent roadmap.
• Mujin raised $224 million in a Series D — SK Square leading, with NTT, the Qatar Investment Authority and Salesforce Ventures in, for warehouse robotics rather than a model.
• SiMa.ai reached a $1.45 billion valuation on a $150 million round — edge silicon for physical AI, priced like software.
• See what's rising across AI and data this week →

Deep Dive: Somebody Still Has To Listen To The Test Pressing

A record never went from the studio to the shop on the producer's word. Somebody cut a test pressing, dropped the needle on it and sat through both sides, listening for the click at minute four that the mix did not have. It cost time and nobody ever clapped for it. Skip it and you find out from ten thousand customers at once.

The checkers got funded, not just thanked

An evaluator raised $71 million. A backup vendor started testing AI-written code. In Europe, €7.13 million went into a seed round for AI software testing. A month ago that work was overhead somebody argued about in a planning meeting. This weekend it had investors, product names and prices. Overhead with a price is a market.

The sellers asked for the brake themselves

Microsoft's president compared AI's regulation to milk's, unfavourably. Its CEO asked for an emergency brake on every model. A lab published a system card admitting its own model presents unchecked assumptions as facts. None of that is modesty. It is what a company does when it can see the liability arriving ahead of the rules.

Which leaves the buyer holding the proof

The part that lands on your desk is simpler than the policy debate. If the evaluator is a supplier, the brake is a request, and the vendor's own paperwork lists the failure modes, then nobody upstream is certifying your output for you. That sign-off was never going to be outsourced. It was just unpriced until now.

What Actually Works

1. Name the signer: for every AI output that reaches a customer or a regulator, one named person accepts it. Not a team, not a vendor. A name.
2. Ask for the evaluation, not the benchmark: a benchmark is a vendor's best day. An evaluation is the document that says what did not pass.
3. Put agent work on the bench: code, config or copy written by an agent goes through the same review a junior's work would. ”The agent did it” is not a review.
4. Budget the checking: evaluation now has suppliers and list prices, which means it has a line in next year's budget whether you put it there or not.

Nobody ever clapped for the test pressing. Everybody heard the click.

What's Coming

Benchmarks Get A Rule-Compliance Column

A study published this weekend scored the same agent twice, generously and then under the task's actual rules, and the result fell from about 35% of a human method's gain to about 15%. Once two numbers exist for one run, procurement starts asking which one the vendor quoted. Expect ”scored under which rules” in AI reviews within a quarter.

One AI Policy, Dozens Of Professions

New Jersey is running a model AI governance experiment across dozens of licensed professions rather than writing one rule per sector. If it holds, the unit of compliance stops being the industry and becomes the licence. That changes who in your organisation is personally exposed, which is a very different conversation from a corporate policy page.

Source Code As A Trade Clause

India and the EU are negotiating rules on source code disclosure inside their free trade agreement, and those clauses decide whether a regulator in one market can demand to see what your model does. Anyone planning to serve both markets from one model wants to read the clause before the launch announcement.

For Your Team

Wednesday's meeting prompt: ”Name the three AI outputs we ship to somebody who can reject them: a customer, a regulator, an auditor. For each one, who here accepts it, and what would an outsider need to see before they accept it? Where is the blank?”

Share-worthy stat: A study scored the same AI agent twice on the same task. Judged generously, it delivered about 35% of what a human-designed method achieves. Judged under the task's own rules, about 15%. Nothing about the agent changed between those two numbers. Only who was checking.

Go deeper: Track where the AI money is actually landing →

The Track of the Day

”For complex change, a review is still needed to ensure quality, in particular when deployment or change on multiple digital products or in a complex ecosystem.”
An enterprise user reviewing GitHub Copilot on Gartner Peer Insights

Somebody wrote that in a product review, not a policy paper, in the same window an evaluator raised $71 million and the biggest AI seller in the room asked to be regulated like dairy. The floor worked this out first. The money is only now catching up.

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

Published: October 13, 2026 | Curated by Yves Mulkers @ Ins7ghts