What moved since July
Eighteen new entries, a correction to my own chart, the figures I keep coming back to, and the calendar the pilot briefing reads. Nothing here is for sale.
This is the first issue, so it covers the whole summer. Twenty-five entries in the catalogue carry July dates; eighteen have been added since, thirteen in August and five so far in September. Those additions are fourteen verified briefings and four practice notes. One chart I had already published needed a correction.
Correction, September 10 (clarified September 13): this issue originally described seventeen new verified briefings. The counts above now show eighteen additions: fourteen verified briefings and four practice notes.
From here on, an issue reports only what moved since the last one, in four parts: something I saw firsthand, what is new, the numbers I keep using, and the dates the evidence carries. It lands when there is enough movement to be worth your time, not on a calendar.
1. A scheduled check corrected my own chart.
In August I ran the source check on the AI Adoption, Drawn to Scale charts early, expecting a newer Deloitte edition. There was none, and the pilot-to-production split behind the enterprise chart was unchanged. What had changed was the source: Deloitte had folded its quarterly generative AI series into its annual State of AI in the Enterprise report, so the refresh cadence stated on my own page was wrong. I corrected the page and moved the next check to the annual edition, expected around January 2027.
Nine days earlier, a figure that circulates everywhere, that 45 percent of the Fortune 500 run AI agents in production, failed provenance on the way into the same chart and was cut before publication. Both events sit in the site's proof record, which is the point of keeping one: a refresh trigger that only ever confirms the page is not a check.
2. Eighteen new entries, and the six I would start with.
All eighteen are in the catalogue; the fourteen verified briefings carry claim ledgers. These six settle the decisions I hear about most.
The Tool Curve Is Not the Value Curve. Across seven technology eras, tool adoption keeps accelerating while the organizational redesign that produces value has never been shown to speed up. Read it before treating an adoption number as a value number.
Delegating work to AI, in three parts. Bound the work before you expand the autonomy; the scarce resource is the right to act; no single vendor owns the enterprise action layer. Together they are the sign-off path for agents that act inside your systems. Also: the right to act · the action layer.
Be Callable Before You Become Autonomous. Make one transaction flow callable by software agents before granting any commercial authority, and scale only when the economics and the assurance both pass.
The Data Is Reachable. The Answer Still Is Not. Access is the visible complaint in regulated enterprises; shared definitions, lineage, and valid joins often bind the defensible answer. Two companion briefings cover the entitlement mechanisms and the constraint analysis. Also: entitlements · constraint analysis.
AI Activity Is Not a Performance Score. Prompt counts, output counts, time in tool, and model-scored quality are traces of tool use, not measures of contribution. Read it before anyone puts telemetry into a review.
Approval Is Not Verification. A first-person note from my own working system: where I had been approving what I could not evaluate, and what I changed so that an approval means something.
3. The numbers I keep coming back to, with their grade.
Each figure links to the ledger row that carries its source and its limits. The grade is the evidence grade, not my confidence, and the pilot-to-production briefing now carries all of its figures in one table.
Read down the rows and the rule writes itself. Expect the upside where the work is codifiable and the people are new to it. Budget for a slowdown and a heavier review load where the work is complex and the people are your best. And treat any adoption percentage as a usage number until someone shows you the redesigned work behind it, which is why the last row was cut.
4. The calendar, as the pilot briefing reads it.
No countdown here, only the dates the evidence carries. The pilot-to-production briefing reads the EU AI Act as a staged calendar: transparency obligations from August 2026, most high-risk obligations moved to December 2027, embedded products in August 2028, with high-risk fines capped at 15 million euros or 3 percent of global turnover. If you run a regulated function, a moved deadline is runway, and the first deliverable is the same either way: an inventory of the AI systems you already run.
Two refreshes are scheduled. The pilot briefing's next source check is due around October 2026, and the adoption charts wait for Deloitte's next annual edition, expected around January 2027. When either lands, it will be in the next issue, along with whatever else moved.
One thing before you go. Insights grew out of the questions my enterprise work kept raising. If the decision you are facing is not on the shelf, tell me on LinkedIn what it is and it goes on the list. And if someone you work with is deciding one of these right now, send them the briefing rather than this issue; the briefing carries the ledger.
Travis
What changed · Issue 1 · September 2026