Release Communication, Strategy & Coverage Self-Audit

Summary

The leverage sits on top of the basics

Everyone who ever shipped frequently failed the same way: the update was distributed but never consumed. So flip the first metric, stop counting updates sent, start counting % adopted. Then the differentiation, which builds on the spine, not instead of it:

The leverage · build 2nd (weeks 30–90)

The spine · build 1st (weeks 0–30)

1 · AI-release comms. A model / retrieval / ACL change shifts answers with no feature to announce, including the occasional regression.

On-intent surfacing of "you can now do X" (needs intent-matching + a confidence signal)

2 · Receiving-side fit. Clients ingest change through gates (change boards, freezes, security/legal). Fit their intake or it never lands.

One in-product nudge + one segmented digest, single lead at the top

3 · Counterfactual ROI. Usage totals die at procurement's "would that have happened anyway?" A comparison cohort survives.

Split every release: Action-Required (tracked) vs Now-Available

4 · Agent-owner notifications (Glean-specific). Tell the person who built an agent that their agent now does Y. Under-exploited by peers, not unique.

Quarterly ROI moment for the buyer, never the biweekly beat · plus the adopted-vs-shipped metric

Don't skip the spine to chase the leverage. Build the basics first (weeks 0–30), then the differentiators (weeks 30–90). They only work on top of the basics.

Part 1 · Start here

Coverage Self-Audit

Click 29 rows (≈20 min, no prep, hand it to one CSM if you like). The bar fills as you mark what you already do. Reds and blanks are your roadmap. Bold groups are the differentiators most teams skip; each differentiator row names the capability it needs. Nothing is saved or sent.

Live coverage

Mark each: We do · Partial · We don't

0/29 0%

Mark the rows above, this line will read your result back to you.

How to read it

Strong in A–D, blank in E–G → solid standard program; the differentiation is the AI-release lane + receiving-side gates + agent-owner notifications (H = the operational backbone). Blank in A–D → start with the MVP spine below.

The reframe

Three audiences + an overlay + a gating layer

Product ships every 2 weeks; admin attention is monthly; renewal is quarterly-or-yearly. By renewal ~6 releases per quarter (~26/year) have piled up, absorbed by no one. The audience isn't one group, three audiences, an asset-ownership overlay, and a gating layer that can block the whole release.

Audiences · must adopt / renew

Adminenable + goEnd-userchange behaviorBuyersee ROI

comms must REACH these

Overlay

Agent owner / buildera power-user hat

a targeting dimension on end-users

Gating layer · can block

Security Legal / Privacy Procurement IT change board / freeze

comms must PASS these

The gating layer is what standard release-comms thinking overlooks: you can do everything right for the three audiences and still fail because a security review or change-freeze blocked the release. The agent-owner overlay is Glean-specific, its knowledge graph + agent library let a release map to owned assets, but the owner is usually a power-user end-user, not a new headcount.

Part 2 · The differentiators

What "ship a better changelog" can't fix

Eleven laws sit under this playbook. Laws 1–8 are mined from 23+ historical parallels (legal publishing → aviation); 9–11 are proposed for the AI / receiving-side context. Every differentiator below cites one.

#	Law	What it forbids
1	Distribution ≠ consumption: measure adopted, not announced	"We sent it" counting as "they got it"
2	Curation is mandatory, the raw feed always under-serves	Shipping the full changelog as the primary channel
3	Encode importance in the artifact	A flat, equal-weight list
4	Bifurcate mandatory vs optional; track the mandatory lane	One stream where act-now sits beside nice-to-have
5	Invisible value is a renewal liability: narrate with a comparison group	ROI claims from telemetry totals with no counterfactual
6	The relay dies the moment it reads as a sale	Letting the end-user/admin relay become an upsell channel
7	Habit forms on a bounded, fixed-format artifact	An unbounded, append-only feed
8	Proactive nudges obey the Clippy rule (relevant + dismissible + controlled)	Unsolicited interruption at a focus moment
9	Communicate the invisible change: an AI release can alter behavior with no visible feature	Letting a silent model/retrieval/ACL shift reach clients unannounced
10	Comms must pass the client's gates, not just reach the audience	Designing only for the sender; ignoring the client's intake machinery
11	Target the owned asset, not just the segment	Generic "what's new" when you could say "the thing you built changed"

AI-release comms (rows 14–16, Law 9), a major gap

Glean is a RAG-grounded AI product, so many releases are invisible behavior changes. The silent-change vectors, in rough order for Glean:

Retrieval / index (the dominant one): re-crawls, connector updates, embedding refresh.
Permission / ACL recomputation.
Ranking, then model, then prompt, then tool-schema.

Two paths, kept separate:

A quality regression is product comms: "answers may differ; here's the eval bound." For genuinely unexplained drift, the honest line is "we detected a shift and are monitoring," not "here's the guardrail."
A permission/ACL change is a security incident. It routes through legal and security, not the comms team, because surfacing a doc to the wrong person is among the most renewal-threatening releases Glean can ship.

Build: an "AI-change" release class with a trigger (eval-delta threshold OR touches retrieval/permissioning), its own template, and an incident/rollback lane. Closest living analogs to copy: Notion's model-drop notes, Dropbox Dash release notes, and Microsoft Copilot's admin Message Center, not Bloomberg. Scoped task: owner = PMM; by day 30, pull the Copilot/Notion/Dash model-update notes and extract their AI-change disclosure template.

Receiving-side gates (rows 17–19, Law 10)

Enterprises ingest change through machinery you don't control:

IT change boards, change calendars, freeze windows, security/legal/procurement review.
Format the must-act items to a change-record schema (change type · risk · rollback · window · action-required · security-review-needed) that drops into their process.
Pre-clear a standing data sheet + model card per release class (increasingly expected as a baseline procurement artifact; verify EU AI Act obligations for your markets).
Add a freeze-declaration intake, and adopt the off-the-shelf pattern: Copilot's admin Message Center + usage-based targeting.
A deprecation or removal needs ≥60–90 day notice.

Counter to pre-empt: some enterprises won't expose their change machinery to a vendor. Fallback: emit the standardized change-record format and let them map it; add the low-cooperation client to the maturity tiers (Part 4).

The counterfactual ROI ledger (row 13, Law 5)

"Saved ~Y hours" dies at procurement's one question. Only a comparison group that didn't get the thing answers it (the pharma difference-in-differences design, Larkin et al., JAMA 2017).

Tier	Method	Defensible?
B · workhorse	Within-subject before/after (needs telemetry from before adoption)	Most defensible thing you can produce unaided
A · bonus	Staged-rollout hold-out, cohort A vs not-yet-enabled B	Only if rollout order isn't confounded by champion self-selection
C	Pre-stated Success-Plan goal	Only if the baseline was locked before the period
D · color only	Telemetry totals ("saved Y hrs")	Never the attribution line

Caveats: you measure usage, not outcome, hours-saved needs a client-validated factor or it's an estimate. Tier-B needs history, so it can't be built until the 30-day instrumentation has aged. If telemetry is too poor, fall back to Tier-C qualitative, never ship an assumed number as Tier-A. The expansion pitch lives in the buyer/CSM channel only, never the end-user relay (Law 6).

Procurement asks	Surviving tier
"Would this have happened anyway?"	A or B (comparison group)
"Is the time-saved number audited?"	Only if client-validated; else label directional
"Can finance defend it?"	B/C tied to a pre-stated goal

Agent-owner notifications (rows 20–24, Law 11), Glean-specific

Glean's knowledge graph + agent library let you tell an agent owner the agent they built is now better, "release R makes your agent X do Y better." Asset-level, not segment-level. An agent owner is a power-user hat a subset of end-users wear (Glean roles = viewer/editor/owner), not a new audience. Build it product-generated, not champion-relayed:

Why build it (not delegate)	The honest caveat
Champions can't compute the release→agent map; only the platform can.	It escapes editorial cost but needs an engineering build, cheap to run ≠ cheap to build.
Cheap for static-config deps (tools, data sources).	Eval-compute-real for model/ranking changes (per-agent eval diffing).
Lower privacy risk, the owner's own asset.	But shared/admin-owned agents → route by sharing graph, digest to moderators, don't leak agent existence.
Eval-backed only, no "better" without a number.	Glean has capability evals but likely not customer-agent evals.
Dual-sided, gains AND "re-check Z."	Fatigue-gate the "re-check" side or owners mute it. Not unique, Copilot Studio has the targeting spine.

Part 3 · The playbook

MVP spine, start here

Cheap, high-transfer. The differentiators sit on top of this. Each piece names the capability it needs. Cost key: L = low · M = medium · H = high effort, build = one-time, run = ongoing.

Engaged · M build / L run

On-intent surfacing

"You can now do X" inside a query. Needs intent-matching + confidence, or it's an infra build and risks Clippy.

Disengaged · L

Nudge + digest

One in-product nudge + one segmented digest, single lead at the top.

Admin · M

Required vs Available

Split every release; track ack on the mandatory lane + a "was it configured?" signal.

Buyer · M

Quarterly ROI moment

Themed, one lead number, never the biweekly beat.

Routing test for any release: push only if breaking/security OR (material AND in-context AND dismissible). Passive (no user action) → auto-deliver silent, narrate later. Behavior-change → surface at intent. AI-change quality shift → its own template + eval note. Permission/ACL → security-incident path (legal + security). Incident/rollback → push immediately. The buyer never receives the biweekly beat.

Routing table, release shape → channel → cadence

Release shape	Channel	Cadence
Admin · breaking/security · needs action	Console (ACK-tracked) + one email, ≥30-day notice	Continuous
Admin · material · optional	Console "Now Available" lane	Biweekly digest
Any · roadmap/unshipped	Preview/Labs only, no clock	Opt-in drip
End-user · needs new behavior	In-context, point-of-use (not email) + GIF + one-click try	Continuous, on-intent
End-user · minor	Ambient / pull (splash, role feed)	Biweekly, below fold
End-user · passive (no action)	Auto-deliver silent, narrate later	Monthly / quarterly
AI-change · quality shift	AI-change template + eval note	At ship + monthly
AI-change · permission/ACL	Security-incident path → legal + security gate	Immediate, gated
Incident / rollback	Incident template, push immediately, gating layer pre-notified	Immediate
Release touches an agent's deps	Per-agent owner notification (sharing-graph routed, eval-backed, dual-sided)	At ship, batched per owner
Buyer · batched	CSM/EBR relay + quarterly report	Quarterly only
Any · minor/fix	Comprehensive changelog (pull)	Always-on reference

Two hard rules: one release → a separate artifact per audience (never one note for three). The buyer never receives the biweekly beat (they may pull the changelog). Notice standards: urgent security fix = ship + notify; breaking change = ≥30-day; deprecation/removal = ≥60–90 day.

Cadence stack, ship cadence ≠ comms cadence

Layer	Beat	Audience	Contains	Owner
L0 in-context drip	real-time / on-intent	end-user	"you can now do X"; passive value silent	Product (free)
L1 bounded artifact	biweekly, numbered	admin + power-user	this cycle's deltas: Highlights → TL;DR → changelog	Release Editor
L2 synthesis	monthly, per-segment	admin + user	cumulative "what's newly possible for your role"	PM-as-editor
L3 buyer momentum	quarterly, themed	buyer	~6 releases → one ROI narrative + ledger	CSM / account
L4 recompilation	quarterly	all	fold deltas into a clean baseline (kills backlog)	Product + editor

Degraded mode (decide funding FIRST): L0–L4 each name a funded human. If no Release Editor is funded, the stack silently collapses, the loose-leaf "labor evaporated and the publisher didn't know" trap. Degrade to: L0 (automated) + an auto-generated bounded changelog + the buyer quarterly.

Part 4 · Execution

What makes it a plan, not an essay

Window	Move	Branch if data is poor
0–30	Run the self-audit; instrument adopted-vs-shipped on top 5 features; ship benefit-framed copy + one lead	–
30–60	Required-vs-Available + tracked ack; pilot on-intent surfacing; draft the AI-change template	If intent-matching unavailable, demote on-intent from MVP
60–90	Quarterly ROI moment with a Tier-B ledger for 1–2 lighthouse accounts; pre-clear the security/legal data-sheet + model card	If telemetry can't support Tier-B, fall back to Tier-C, don't ship an assumed number

Kill-criteria (starter thresholds)	Change or kill the channel when…
Digest unsubscribe rate	> 2% / quarter
Nudge dismiss rate	> 40% (or 2× the 30-day baseline)
Mandatory-lane ack	< 80% acked + configured within 30 days
Adoption decline	> 10% relative drop (or > 5pp) on a broad feature, release-over-release
Agent-owner opt-out	> 15% of notified owners mute
On-intent self-announce	precision < 95% (false self-promotion inside answers)

The one call only the manager can make: who owns the curated artifact each cycle (a "Release Editor")? Unowned, it silently collapses, the way the loose-leaf legal-binder industry lost its distribution (updates shipped, nobody filed them). Default: vendor curates one artifact centrally; champions only relay. Resourcing: full stack ≈ ~0.5 FTE Release Editor + ~1–2 engineers for a quarter (then low maintenance) for the dependency-map/eval pipeline + ~0.25 FTE data/analytics + CSM time. MVP ≈ 1 owner, no new eng. Choose the tier you can staff (bands are rough placeholders).

Owners, who does what

Layer	Does it	Accountable	Consulted
In-product on-intent (L0)	Product/Eng	PM	–
Biweekly artifact (L1) + AI-change template	Release Editor	PMM/PM	Eng (eval notes), Legal (gating)
Monthly synthesis (L2) + recompiled baseline (L4)	Release Editor	PMM	CS
Quarterly buyer ROI (L3) + ledger	CSM / Account	CS lead	Data/analytics, Finance
Agent-owner notification pipeline	Eng (the dependency map)	PM	Data, Legal
Adopted-vs-shipped metric	Data/analytics	PM	–

Customer-maturity tiers → row subsets

Tier	Posture	Rows in scope
Enterprise, high-maturity	Full stack + design-partner early access + gating fit	all
Mid-market	MVP spine + champion kit	1–13, 20–24, 25–26, 29
Regulated (finance/health/EU)	AI-change + ACL + model-card + gating mandatory; localization & EU-AI-Act = launch-gates	1–19, 25–29 + EU gate
Low-maturity / SMB / low-cooperation	Safe-default ring + managed-on-behalf + passive receipt; standardized change-record	1–8, 11, 12, 25, 29

Ground-truth intake checklist, do this before trusting any recommendation

Current release-comms channels inventory (what you send, where).
Adopted-vs-shipped on the top features (the real baseline).
Top-3 renewal objections CS actually hears.
Which historical releases needed behavior change vs were passive.
Current freeze/change-board exposure across the top accounts.

Part 5 · Preconditions

Capabilities each piece needs

Built without ground truth, so several mechanisms assume a platform capability. Verify each exists before promising it; some may be roadmap items, not latent features.

Mechanism	Required capability	If absent
On-intent surfacing	Query-intent classification + a confidence signal	An infra build; demote from MVP
Agent-owner notifications	A release→capability→agent dependency map; per-agent evals for model changes	A roadmap item, not a comms feature
Tier-B ROI ledger	Per-cohort usage telemetry retained from before adoption	Falls to Tier-C/D
AI-change class	A release-tagging / eval-regression detector	No trigger for the lane
Timing dial + freeze	Per-tenant rollout rings + a client-freeze intake	A roadmap ask
Tracked-ack downstream	Config-state telemetry joined to the ack	Ack proves a click only

Part 6 · What this is

What this is, and isn't

Built without ground truth. Three things this can't give you: which of these you already do (the self-audit fixes that), your real adoption baseline, and whether your org can staff the Release Editor and the eng build. Get those three and this becomes a real plan.

Part 7 · Worked example

One release, end-to-end

Three releases ship in cycle 2026.14, trace each through class → route → metric → agent owners. The machine has to compose.

	(a) Retrieval re-index (silent)	(b) Quality regression on long queries	(c) New "summarize thread" skill
Class	AI-change · retrieval · passive	AI-change · quality-regression	feature · needs new behavior
Route	Auto-deliver; narrate in monthly + buyer receipt	Product comms: "answers may differ; eval bound + rollback"	In-context on-intent + GIF + one-click try
If it touched ACL	→ security-incident path (legal + security gate)	n/a	n/a
Metric	answer-acceptance before/after (Tier-B)	regression-eval pass rate; complaint rate	% users who applied it ≥3× (not views)
Agent owners	notify owners of agents that retrieve from the re-indexed source	notify owners whose agents regressed, with the eval delta	–

Passive value gets narrated; a regression gets the honest product note; an ACL change escalates to legal; behavior-change surfaces at intent.