One mechanism. Several symptoms.
The cost of capturing a judgment just fell to zero. That single change turns organisational memory from a contribution tax nobody pays into a byproduct of work people already do. Everything else on this page follows from it.
The market just priced the bottleneck, and its answer is bounded by hiring.
In July 2026 a consortium including Anthropic, Blackstone, Goldman Sachs and General Atlantic committed 1.5 billion dollars to the thesis that enterprise AI fails at deployment rather than at the model. We agree with the diagnosis. The remedy they funded is a hundred engineers, and they state themselves that this talent stays oversubscribed for eighteen to twenty-four months. A market whose best funded answer scales by recruitment is a market waiting for one that compounds.
One change, and the consequences fall out of it.
Every knowledge system before this one asked a person to stop working and write things down, for a benefit that went to the organisation and never came back to them. Rational people declined, so the systems emptied, and everyone blamed culture. The tax is what failed, not the idea.
The same job, delivered by two different kinds of thing.
It arrives, learns the operation, ships, and leaves
Excellent work, genuinely. And when the engagement ends, the craft and the understanding of that particular house leave with the team. The firm rebuilds it at the next client, which is why it must bill by the day. Capacity grows only by recruiting, in a market that says the talent is oversubscribed.
It arrives, learns the operation, ships, and stays
The same four moves, with the fourth one changed. What it learned about that house remains, and each decision makes the next arrival cheaper. Capacity grows by deployment rather than by hiring, and the accumulated understanding is the thing a competitor cannot backfill.
Three layers, and only one of them is defensible.
Services firms hold all three, in people, and lose the third at the end of every engagement. That is not a management flaw, it is the nature of their container. It is also the reason their revenue cannot compound.
The parts that are still being built, said before you ask.
We are early, with paying design partners. The compounding effect is the whole thesis and we are measuring it rather than asserting it: the first engagements establish the baseline, not the curve. Automatic reopening and writeback are being wired and are labelled as such everywhere they appear. And on the storage mechanism for conflicting beliefs, parts of the funded memory tier are at parity with us today. What is ours is custody at the moment of decision, and the buyer who has to answer for it.
Every wave of automation moves the value to whoever holds the new system of record.
Agents are producing decisions faster than any services army can account for. We are building the layer that will have to answer for them.