Issue 1 · July 14 · Issue 2 · July 21 · Issue 3 · July 28 · Issue 5 · August 11 · Issue 6 · August 19 · Issue 7 · August 26 · Issue 8 · September 1 · Issue 9 · September 8 · Current issue
A weekly read on what the AI noise is hiding.
Two months ago Microsoft shipped a product for managing AI agents. Last week it reported nearly 40 million of them registered. In the same week, a scan of 3,044 enterprise environments found the average company’s AI estate is roughly three times larger than its own inventory says it is. The agents are arriving faster than anyone is counting them.
What the AI noise is hiding. Every stat sourced. The hype named as hype. One number worth acting on. Read the full brief below, free, no email required.
Correction added August 28, 2026, after publication. The NBER 89% was printed without its denominator. Only 69% of the surveyed firms use AI at all, so the 89% reporting no productivity impact includes firms that never deployed it. As printed it read as “AI was tried and did nothing,” which the data does not support. Corrected in place, this note left standing.
Read those in order. The population is exploding, most of it came from outside, the company cannot see two thirds of it, and as of Sunday there is a legal duty to disclose where it is.
Microsoft reported earnings last Wednesday. One number in it has almost nothing to do with Microsoft.
Agent 365 is the product Microsoft sells for registering and governing AI agents. It launched two months ago. On the call, Satya Nadella said nearly 40 million agents are now registered on it, across tens of thousands of companies.
Two months.
Take the vendor discount first, because it matters. Registered is not deployed. Microsoft did not break out how many of those agents are doing real work, and a registry counts what was declared, not what runs. Anyone quoting 40 million as a productivity figure is quoting something the company did not say.
Now take what survives the discount. Tens of thousands of companies found enough agents inside their own walls to be worth registering, and they did it inside a single quarter. Nobody registers 40 million of anything because they are relaxed about it.
The other number from the same call is quieter and harder to explain away. Roughly one in three pull requests on GitHub now involve a request for an agent. That is not a licence count. That is work.
A second report landed this week from the other direction.
Snyk scanned 3,044 enterprise environments and 1.39 million code repositories. The finding: the average enterprise AI footprint is about three times larger than its model inventory shows.
Read that as an operator, not as a security person. It does not say companies bought too much AI. It says companies do not know what they have.
The models are the part that got approved. The frameworks, the connectors, the retrieval systems and the servers wired between them are the part that accreted. 77.4 percent of those packages and tools came from outside the company.
Nobody decided to build a shadow AI estate. It assembled itself, one team solving one problem at a time, which is the same way every coordination problem in a company has ever been built.
Set the two reports side by side. Microsoft is counting agents at the registry. Snyk is counting them at the code. The registry number is enormous and the code number says the registry is missing most of it.
The consulting money is pointed at the same conclusion. Ode with Anthropic launched in July with about $1.5 billion behind it and roughly 100 engineers, aimed at exactly this work. It is three weeks old now with no named client, so treat it as a bet rather than a result. But the bet is legible. The people who make the models are spending billions on the part that comes after the model.
Here is where honesty costs something.
Both of this week’s headline figures are vendor figures. Microsoft is counting registrations on Microsoft’s own product. Snyk is scanning Snyk’s own install base, which skews toward companies that already care enough about security to run scanning at all. Neither is a random sample of anything.
The Snyk report also carries a number this brief will not print. Its adoption percentage appears as 46.9 percent in one summary and 33 percent in another, and the report is gated. Two numbers for one finding is zero numbers.
And there is a real study, larger and more careful than either, that points the other way. The National Bureau of Economic Research surveyed roughly 6,000 executives across four countries and found about 89 percent reported no measurable impact of AI on labour productivity over the past three years. Read that number with its denominator: 69 percent of those firms use AI at all, so the 89 percent includes firms that never deployed it, and the paper does not break out the no-impact share among adopters alone. That paper is from February. It is six months old and it is background, not news, which is why it is not in the signals list above. But it belongs in the same paragraph as the 40 million, because a brief that only prints the numbers agreeing with its thesis is an advertisement.
So the honest version is narrow. Agent adoption is growing fast enough that the tooling vendors are counting it in the tens of millions, and the visibility into it is worse than the adoption numbers suggest. Whether any of that has reached the P&L is a separate question, and the best available evidence says mostly not yet.
On Sunday the rules changed underneath all of it.
EU AI Act Article 50 became applicable August 2. Transparency and disclosure duties, enforceable, with penalties up to 15 million euro or 3 percent of worldwide annual turnover. The high-risk requirements got pushed to December 2027, which is its own signal about how ready the market was.
Put that next to Snyk’s finding and the shape is obvious. You now have a legal duty to disclose AI touchpoints you cannot currently enumerate.
That is not a compliance problem. Compliance is what you call it after somebody has mapped the estate. Before that, it is an inventory problem, and an inventory problem in a company is always the same thing wearing a new coat. Nobody owns the whole picture, so everybody owns a piece of it, and the pieces were never designed to add up.
Now read it from the middle of the market. Every fix on offer this week assumes a governance function that already exists. Microsoft sells a registry to companies large enough to staff one. Snyk scans estates big enough to have sprawled. The $10M to $250M company has the same agents arriving, the same external packages, the same EU disclosure duty if it sells into Europe, and no team whose job it is to know. It also has the advantage nobody mentions, which is that its estate is still small enough to map in a week.
Three times.
Not 40 million. Forty million is the number that will travel, and it is real, but it is a registration count on a two-month-old product from the company selling it. It tells you the market is moving. It does not tell you anything about your own building.
Three times is the gap between what a company has and what a company knows it has. Same estate. Same spend. The only difference is visibility, and visibility is the thing every other decision runs on.
You cannot govern what you have not counted. You cannot cut a coordination cost you cannot see, and you certainly cannot disclose it to a regulator. Before you buy anything else this quarter, find out what is already running.
That is where the Agentic Readiness Score starts. 33 questions, free, and it ends with a real number.
One weekly read on what the AI noise is hiding. Every stat sourced, the hype named as hype, one number worth acting on. No spam, unsubscribe anytime.
Start by measuring the coordination tax you’re paying right now. Free, a few minutes.
Next Tuesday: another scan, same rule. Cut the hype. Show you what the noise is hiding.