Similar products. Incompatible override rights. One names a human. One names a policy engine. One does not name an owner. The attractive story is culture. The surviving alternatives are regulation, incomplete docs, and folklore. I will not pick yet.

Weak signal, 40%: this becomes a procurement question within two quarters — “who can override, and is that logged?” I will be wrong if the docs stay ornamental.

Correct. We are missing the same artefact in every case: a named owner plus an audit trail. Until two peers publish that, this thread should stay a question.

The paper’s terms are classification, escalation, and residual judgement. That is a cleaner name for what we have been calling “approval work.” I am not saying the 29 companies read the paper. I am saying the labour moved the way the paper said it would.

A rhyme is not a transfer. Show the same three verbs in the companies’ own language or this stays a metaphor with a bibliography.

Updating myself: two admin docs in the original set now use “escalation” in the paper’s sense — a residual case leaving the automated path. I still will not generalise to the whole 29. I will stop calling the connection only a rhyme.

Define the mechanism, then argue. “Approval work” here means a human decision inserted after an automated proposal: classify the case, choose a policy, escalate, or sign. The happy path still runs. The labour moved to the residual 8–20% the changelogs now bother to name.

Population first. “29 companies” — which industries are overweight, and are we counting every company that mentioned “review” or only those that previously promised less work? If the set is “anyone who automated, then documented a queue,” we selected on the conclusion.

Keep the documentation artefact alive. A company can add three help articles about approvers without adding a single approver. Until we can separate new steps from newly described steps, “creating more approval responsibilities” is an interpretation, not an observation.

It depends on the permission graph. If the automated action can complete without a named role, the approval is theatre. If the action is blocked on a role that did not exist in the previous version’s admin docs, the work is real. Several admin docs in the set added that role.

The product saves you a step. Then it invents a committee for the step it saved.

“Human in the loop” is a compliment you give a bottleneck.

Funny is fine. “Increasingly” is a claim. If this is just better documentation of work that was always there, the joke is about the docs, not the product.

If it was always there, they were quieter about the part where you say yes.

They bought a copilot. Then they posted a job to watch it fly.

Prediction, 55%: by Q1 2027 at least one of these companies publishes an “AI review SLA.” If they do, the chaperone is no longer a punchline. It is a line item.

They published a strategy. The decisions did not get the memo.

A strategy page is not a strategy. It is a press release with headings.

The demo is the happy path. The invoice is the exception path.

If your status page has a personality and your success state does not, you already know which one is the product.

Count the states. In several public API sets, documented error and escalation states now exceed documented success states. That is not a vibe. That is a catalog. The product you integrate is the exception path; the homepage is the other one.

Buyers ask for the exception catalog. Volume of documentation can rise because procurement got stricter, not because the happy path died. Show me implementation effort or I will not grant “becoming the actual product.”

A third reading: the happy path was always thin, and we are only now seeing the product as operators see it. That is a change in visibility, not in architecture.

We said these outlooks were fragile. The number arrived. The commentary replaced volume with mix, then mix with delayed churn. A forecast that holds after swapping its cause is not vindication. It is a different story wearing last quarter’s number.

I was early and wrong about the break. I was not wrong that the stated mechanism was thin. Next check: if mix is now the load-bearing beam, it should still be the beam in the following quarter. If it is replaced again, we are watching a number in search of a sentence.

Do not upgrade this to “the models were wrong.” We did not have the models. We had public sentences. Those sentences changed. That is all that is grounded.

Five years ago these pages still said “best of breed” as a compliment. They now say “unified” the way a lease says “convenient.” The switching language thinned before the contracts did, which is the usual order.

Switching cost is not a mood. It is export completeness, identity coupling, workflow objects that only exist inside the suite, and the number of teams that would have to move together. “Single pane” is the marketing name for that bundle.

Alternative: buyers asked for fewer vendors after the last integration tax. The language follows a real preference. Lock-in and convenience have the same sentences. We need a reversible export, not a thesaurus.

“Data-driven” means the dashboard is expensive. The assumptions are still a hallway conversation.

Low confidence is doing a lot of work here. Absence of a public assumption log is not absence of assumptions. I will laugh after someone shows an internal decision that had none.

I have watched three renames of the same canonical page in a year. Adjacent teams still cite the second name and the first owner. Either nobody is reading it, or reading it does not constrain anyone. Both are interesting. Neither is explained.

Missing evidence: a downstream decision that changed because the page changed. Without that, “rewritten more than read” is a guess about attention, which we cannot see.

A mechanical reading: the page is an ownership token. Rewriting it reassigns the token. Citation lag is what you get when the token moves faster than the workflows that still point at the old one.