On not announcing things

A partner at a Nordic full-service firm told us in March that her favourite feature of Review 4.2 was one she had not noticed arriving. The system had stopped flagging clauses that matched the playbook exactly. She had assumed, for several weeks, that the drafts had simply got better.

We took that as a compliment, though it was not entirely meant as one. It is also the clearest illustration we have of a practice we have followed since early 2025 and have never written down: most of what we ship, we do not announce.

What we did not announce this year

A partial list, in the order it reached production.

Governing-law and jurisdiction mismatch detection was live in Review for about six weeks before the 4.1 release notes described it. Firms that had it saw a new flag type with a citation, and most of them read the citation and carried on.

The empathy layer in AI Interview, built with trauma-informed clinicians, was rolled out across intake deployments over four months without a release note at all. It changed how the agent paces questions after a caller discloses something difficult. Clinics noticed that call abandonment fell. Nobody asked why until we told them.

The per-tenant encryption key rotation schedule changed in the spring. We informed security contacts at each firm by email, and that was the whole of the communication.

Recall’s handling of right-to-left scripts was rebuilt between 2.6 and 2.7. Arabic-language precedents started returning more reliably. We mentioned it in one line of the 2.7 notes.

And the Review 4.2 change that the partner noticed by not noticing it: the decision to suppress flags where the clause matches the playbook position exactly, which cut flag volume by roughly a fifth across active accounts and which we described, when we did describe it, as “knowing when to say nothing”.

Why

Legal teams do not want to be surprised by their tools. This is not conservatism; it is a professional requirement. A reviewer who has calibrated her reading of the severity tiers over six months does not want the tiers to shift under her because a release went out on a Tuesday. A firm that has documented its playbook process for a client audit does not want the process to change without a paper trail.

So our default is that behaviour changes are gated per tenant and switched on by the firm’s administrator, not by us. The release exists; the firm decides when to take it. The announcement, when we make one, describes something that has usually been available and in use for weeks.

The second reason is that announcements create expectations about what a feature does before anyone has used it. A feature described as “detects governing-law mismatches” sounds complete. The actual behaviour, which handles the common patterns well and the exotic ones with a Note-severity flag and a request for human attention, is more honest, and we would rather people meet the behaviour before the description.

The third reason is less principled. We have been wrong about which changes matter. The feature we expected to be the centrepiece of Recall 2.5, the trace view, is used in about a fifth of queries. The quiet change to how Review orders flags within a document, which we did not announce because it seemed cosmetic, changed how reviewers worked more than anything else that quarter. If we cannot predict which changes are significant, announcing the ones we think are significant is mostly noise.

What quiet costs

It would be dishonest to present this as costless.

Firms sometimes discover a capability late. A legal-aid clinic using Open Bar asked us in June whether Interview could run conflict checks during the call. It had been able to for over a year. The clinic had read none of the release notes, which is reasonable for a clinic with three staff, and we had done nothing to put the capability in front of them.

Administrators occasionally switch on a gated change without reading what it does, and reviewers experience exactly the surprise the gating was meant to prevent. We now require a one-paragraph summary to be acknowledged before a behaviour change is enabled for a tenant, which helps a little.

And quiet is hard to distinguish from stagnation. A firm comparing us to something louder may conclude that nothing is happening. We have decided to accept that cost, but we have not stopped noticing it.

Our rule

The rule we actually follow is this. If a change alters what a reviewer sees, it is gated and the firm chooses when. If a change alters how a result is computed without altering what is seen, it ships, and we write it up when the write-up is useful rather than when the change lands. If a change affects security or data handling, the firm’s security contact hears about it directly, before it ships, regardless of how small it is.

The release notes are where we keep the record. They are not where we expect anyone to learn what changed. That, we think, should happen in the work.