Anthropic released Claude Opus 4.7 this week, but a far more capable model called Mythos went to a handful of institutions only. Plus: why "Bob" is every construction CEO's biggest AI problem, the ten-year data moat nobody is talking about, and a look inside our upcoming procurement report.

THE EXECUTIVE BRIEFING
THIS WEEK’S KEY TAKEAWAYS

Full episode write-up at the bottom↓

Shadow AI is already inside your business. Your staff are uploading company data to unsanctioned tools. Your teams are building peer-to-peer agents with systematic errors nobody audits. This is a governance crisis, not a productivity story. Most executives are still treating it like one.

Shock horror!! Data-readiness is the real moat. Three AI founders across procurement, scheduling, and drawings said the same thing this week. The value doesn't sit in the AI. It sits in the organised, connected, maintained data the AI needs to work.

The AI world split into two tiers this week. Anthropic's most capable model, Mythos, is not available to the public. Contractors with top-tier AI access will quote against you on the same jobs with a structurally lower cost of intelligence.

7 THINGS WORTH YOUR ATTENTION
ON THE RADAR THIS WEEK

POWERED BY:

The man who built Facebook's computer vision team, the system that classifies every photo and every second of every video across the platform, closed a ten-million-dollar seed round this week for a new company. Its focus: understanding construction drawings with AI.

I asked him how the difficulty compared to the work he did at Facebook.

His answer: Reading construction drawings is harder.

When one of the top AI researchers on the planet says your industry is harder than Facebook-grade AI, the question stops being is AI going to transform construction? It becomes who's going to benefit when it does, and who isn't?

That's what this week's briefing was about. Three pillars. One uncomfortable observation. One divergence moment that arrived live, in the wild.

The US construction industry spends around $600 billion a year on materials. Between $60 and $120 billion of that is lost to inefficiency every year.

Average overpayment sits at 5%. The same company, buying from the same vendor, can pay 25-30% more in one region than another. Not because of market differences. Because nobody has the data to compare.

For businesses operating on 2-3% margins, that's the difference between profit and loss on a job.

The full Bricks & Bytes procurement report drops on Friday, 24th April. In-depth interviews with technology founders. A peer survey of construction leaders across the US and UK. A structural look at a function most firms treat as a back-office afterthought.

A senior leader at a large US general contractor described his single biggest fear around AI. It wasn't the technology. It was Bob.

Picture someone in your pre-construction team. Smart, curious, quick on the uptake. Bob's been building agents on Claude Code or Cowork. One speeds up part of his job. He shares it with the rest of the team. Now the whole pre-con department uses Bob's agent.

The problem: Bob got it most of the way there. The agent is convincingly right; close enough that nobody notices the places where it's subtly wrong. An entire department is producing work based on an output with a systematic error, because it came from Bob, and Bob is one of us.

Bob is not a technologist.

Now layer on what Maryrose Lyons sees across the UK and Ireland training practice. She calls it shadow AI. The CEO says: "we don't really use AI in our business." Then she asks the staff. All of them are using it. Paid tools, free tools, personal accounts. Client data, pricing data, project data; all flowing out through apps IT never approved.

Her phrase: "all of your company data is leaving via your staff's phones."

Put the two together:

Most construction executives are still treating this like a productivity issue. It's a governance issue. Governance isn't something you buy. It's something you run.

Right now, across most of this industry, nobody's running it.

A founder who held a senior position in the Singapore government before starting a robotics company gave the most striking answer I heard all week.

Asked what he'd tell a construction leader about the next ten years:

Not the best tools. Not the best vendors. The ones with the data ready.

The same pattern shows up across three completely different domains this week:

Three domains. Three founders. Same answer. The value doesn't sit in the AI. It sits in the organised, structured, maintained, connected data the AI needs to be useful.

Claude Opus 4.7 was released this week. Solid, incremental.

But on the comparison chart, a model called Mythos sits in the same column. Significantly more capable than anything else on the market.

Mythos hasn't been released to the public. Anthropic judged it a security risk and did a private release to major institutions only. Some of the bugs it's uncovering are over 20 years old.

I welcome the caution. But pull back and look at what this actually is.

The AI world now runs on three layers:

The new currency is access. And this week, the leading AI company now has one tier of AI for itself and its largest clients, and another tier for everyone else.

Here's what keeps me up about it.

If Mythos today can refactor 20-year-old vulnerabilities inside the world's most complex institutions, what does the next version of that capability do inside a construction business that has access to it? Not in ten years. In three.

It prices work more accurately than any estimator. Schedules with precision your PMs can't match. Identifies risk, optimises procurement, runs the commercial side of a project in a way that drops overhead to a fraction of what you carry today.

And the contractors with access will be quoting against you on the same jobs, with a cost of intelligence that's structurally lower than yours. You'll lose bids. You won't understand why. By the time you do, you won't be able to catch up, because the data asset required to use that AI is the one you didn't build over the last three years.

This isn't about productivity. It's about survival.

The AI you're buying isn't the advantage. The business you build around it is.