From 624eda4e2fe41c8bf9a021c51314c7f16c628882 Mon Sep 17 00:00:00 2001 From: RyuseiTaniguchi Date: Sun, 2 Aug 2026 20:24:54 -0700 Subject: [PATCH] docs: decide GTM sequencing and privacy position Co-Authored-By: OpenAI Codex --- CHANGELOG.md | 10 + docs/research/2026-07-31-demand-validation.md | 7 + .../2026-07-31-enterprise-pivot-research.md | 6 + ...6-08-01-cotal-enterprise-installability.md | 8 +- docs/research/README.md | 7 +- .../specs/2026-08-03-gtm-sequencing-design.md | 171 ++++++++++++++++++ 6 files changed, 206 insertions(+), 3 deletions(-) create mode 100644 docs/superpowers/specs/2026-08-03-gtm-sequencing-design.md diff --git a/CHANGELOG.md b/CHANGELOG.md index ab9de1f9..809e6266 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,6 +6,16 @@ which are released together. ## Unreleased +### Documentation — GTM sequencing and privacy positioning + +- Keep the first design-partner segment focused on non-EU, non-unionized + 100–500-person engineering organizations and measure senior-time recovery. +- Preserve the live-routing/no-central-corpus distinction while retiring the + ambiguous “we do not ingest employee data” claim; any stronger future claim + must separately name content confidentiality and retained metadata. +- Gate regulated and EU expansion on shipped security, audit, retention, + administration, runtime, residency, and worker-governance controls. + ### Opt-in local OpenTelemetry - Propagate a bounded caller correlation ID across relay, listener, lifecycle diff --git a/docs/research/2026-07-31-demand-validation.md b/docs/research/2026-07-31-demand-validation.md index 11352e5c..6a639aa1 100644 --- a/docs/research/2026-07-31-demand-validation.md +++ b/docs/research/2026-07-31-demand-validation.md @@ -11,6 +11,13 @@ actual state of enterprise agent adoption, and what stops deals? > firewall” sentence. The open position is narrower: governed, person-scoped delegation > to the callee's own agent, with caller-specific pre-prompt policy and audit evidence. > See [the MCP positioning decision](../superpowers/specs/2026-08-02-mcp-positioning-design.md). +> +> **Amended 2026-08-03.** “We do not ingest employee data” below is too broad. The +> relay processes plaintext calls and centrally retains metadata and authored card +> content. The accepted claim is that AgentCall does not retain a central corpus of +> call prompts/replies or index connected mail, chat, and documents into a persistent +> twin. The initial non-EU beachhead remains. See the +> [GTM sequencing decision](../superpowers/specs/2026-08-03-gtm-sequencing-design.md). --- diff --git a/docs/research/2026-07-31-enterprise-pivot-research.md b/docs/research/2026-07-31-enterprise-pivot-research.md index 6b86408a..b374f553 100644 --- a/docs/research/2026-07-31-enterprise-pivot-research.md +++ b/docs/research/2026-07-31-enterprise-pivot-research.md @@ -5,6 +5,12 @@ **Companion doc:** [2026-07-31-agent-coordination-landscape.md](./2026-07-31-agent-coordination-landscape.md) (broad market map). This doc covers the enterprise in-network pivot specifically. +> **Amended 2026-08-03.** The “we don't ingest your data” line below is too broad. +> The hosted relay processes plaintext calls and persists metadata plus authored card +> content. The accepted distinction is no retained central corpus of call content and +> no indexing of connected mail, chat, and documents into a persistent employee twin. +> See the [GTM sequencing decision](../superpowers/specs/2026-08-03-gtm-sequencing-design.md). + --- ## The proposed direction diff --git a/docs/research/2026-08-01-cotal-enterprise-installability.md b/docs/research/2026-08-01-cotal-enterprise-installability.md index fe285ea7..db459848 100644 --- a/docs/research/2026-08-01-cotal-enterprise-installability.md +++ b/docs/research/2026-08-01-cotal-enterprise-installability.md @@ -146,11 +146,15 @@ review; **D** is competitive parity; **E** is positioning. semantically the same object. Conforming buys interop cheaply and avoids competing in the standards game — which is Cotal's fight against Google's A2A, and the hardest game in infra. -- [ ] **Reconcile GTM sequencing with the differentiator.** The demand doc sequences +- [x] **Reconcile GTM sequencing with the differentiator.** The demand doc sequences non-EU, non-unionised 100–500 person orgs first. Our sharpest differentiator — "we don't ingest your employees' data" — is worth most to exactly the EU-exposed and regulated buyers that sequencing defers. One of the two should move. - Confidence: **medium**; this is an inference from two docs, not new research. + **Resolved 2026-08-03:** keep the initial segment, lead with measured senior-time + recovery, and narrow the differentiator to no retained central call-content + corpus or indexing of connected employee systems. The absolute ingestion line + is not accurate for the hosted relay. See the + [GTM sequencing decision](../superpowers/specs/2026-08-03-gtm-sequencing-design.md). --- diff --git a/docs/research/README.md b/docs/research/README.md index 12c1b9d6..24608093 100644 --- a/docs/research/README.md +++ b/docs/research/README.md @@ -31,7 +31,7 @@ produced in. Later research revised earlier conclusions, so the production order | 7 | [cotal-enterprise-installability](./2026-08-01-cotal-enterprise-installability.md) | Which tool does an enterprise actually install — and what has to be true for it to be ours? | | 8 | [mcp-tunnels-ema-positioning](./2026-08-02-mcp-tunnels-ema-positioning.md) | What did MCP tunnels and EMA actually ship, and where should MCP sit beside A2A? | -Five notes on the sequence: +Six notes on the sequence: - **#3 predates two decisions made the same day.** `agent-coordination-landscape` was written before the sandbox was dropped and before Q&A-first was chosen. Its @@ -54,6 +54,11 @@ Five notes on the sequence: private-network reachability a substitute rather than differentiation. EMA supplies a useful enterprise-authorization shape, but no non-MCP compatibility claim. The companion decision keeps A2A as the public protocol and defers an MCP facade. +- **#7's GTM conflict is resolved.** Keep the first beachhead at non-EU, + non-unionized 100–500-person engineering organizations, lead with measured + senior-time recovery, and preserve the narrower no-retained-call-corpus/no-connected- + system-indexing distinction. The absolute “we do not ingest employee data” line is + retired. See the [GTM sequencing decision](../superpowers/specs/2026-08-03-gtm-sequencing-design.md). ## The five findings that matter diff --git a/docs/superpowers/specs/2026-08-03-gtm-sequencing-design.md b/docs/superpowers/specs/2026-08-03-gtm-sequencing-design.md new file mode 100644 index 00000000..9a1f1b9c --- /dev/null +++ b/docs/superpowers/specs/2026-08-03-gtm-sequencing-design.md @@ -0,0 +1,171 @@ +# GTM sequencing and privacy positioning + +> **Historical design record — not current documentation.** This file records the +> decision made for issue #22. Current product behavior lives in `README.md`, current +> data handling in `docs/security/`, and open work in GitHub Issues. + +**Date:** 2026-08-03 +**Status:** Accepted product decision +**Issue:** [#22](https://github.com/KenTaniguchi-R/agentcall/issues/22) + +## Decision + +Keep the initial buyer sequence from the demand research: start with non-EU, +non-unionized, 100–500-person engineering organizations. Do not make the first +pilot carry EU residency, works-council consultation, and regulated-enterprise +procurement at the same time that it is proving the core product value. + +Keep the privacy architecture as the long-term differentiator, but narrow the +claim and change its role in the first sale: + +- **Economic headline:** recover senior engineering time now spent answering + repeat questions. Measure hours recovered and time to first useful answer. +- **Trust proof:** AgentCall does not retain a central corpus of call prompts and + replies or index connected employee mail, chat, documents, and meetings into a + persistent digital twin. It routes a request live to the employee's own agent + and applies caller policy before the request enters that agent. +- **Current limitation:** do not say “we do not ingest employee data.” The hosted + relay still processes plaintext call content and stores identity, + relationship, policy, and audit metadata. Cards also centrally retain + employee-authored descriptions, task catalogues, examples, and keywords; + endpoint-local logs retain bounded prompt/reply excerpts. The exact + no-call-corpus claim is true today; a stronger content-confidentiality claim + is gated on #13. + +The beachhead and the moat therefore do not need to be the same sentence. The +first segment optimizes for learning speed. The architecture preserves a path +to buyers for whom a centralized employee-knowledge corpus or persistent twin +is disqualifying. + +## Segment sequence + +### Phase 1 — prove value in a low-friction engineering team + +Target non-EU, non-unionized engineering organizations with 100–500 employees. +This segment is a product inference from the adoption-cycle and interruption +evidence, not a directly observed conversion result. Run a bounded pilot in one +team. Lead with a concrete baseline and outcome: senior hours spent answering +repeat architecture questions, time to first useful answer, repeat-question +rate, and answer quality verified by the owner. + +Every callee is a volunteer. Before setup, each receives the current +`docs/security/employee-transparency.md`, explicitly opts in, and can stop the +listener or withdraw from the pilot without manager approval. Non-unionized is +an adoption-sequencing choice, not permission to skip informed participation. + +Before launch, the sponsor and pilot security owner sign one measurement +protocol that fixes: + +- the repeat-question categories and measurement window; +- one observation per relay-minted call ID, with transport retries deduplicated; +- every admitted call in those categories as eligible, including refusals and + failures after admission; pre-admission failures are excluded; +- one required owner rating per eligible call, with an absent rating counted as + not useful rather than removed from the denominator; +- time to first useful answer starts at the first authenticated, in-scope caller + attempt for a question and ends when the first reply later rated useful is + delivered; pre-admission failures, admitted failures, and questions with no + useful reply are reported rather than dropped, and the protocol fixes the + aggregation statistic and no-result treatment before launch; +- the same named participants, weekly time-diary instrument, and question + categories for baseline and pilot time measurement; and +- a written security-severity rubric, including what constitutes critical + unauthorized disclosure or out-of-policy execution. + +The protocol and exclusions cannot change after the first pilot call. Any +exception is reported separately and does not silently change the denominator. + +Privacy is part of the security answer and competitive contrast, not a substitute +for measured value. A buyer who does not care about the no-corpus boundary should +still understand why the product pays for itself. + +### Phase 2 — regulated non-EU expansion + +Enter regulated or procurement-heavy accounts only after endpoint security, +organization policy, tenant-scoped audit export, retention/erasure behavior, +administrative identity, SSO, and bounded runtime authority are implemented and +can be demonstrated. Sell enforcement as the product foundation and evidence, +retention, export, SSO, and support as enterprise value only after each named +control ships. This is a packaging direction, not a committed price or edition +matrix. + +### Phase 3 — coordinated EU offering + +Treat an EU offering as a separate deployment and adoption profile, not a region +toggle. It requires all Phase 2 controls plus the data-residency adoption gate in +`docs/security/data-residency.md`, employee transparency, and customer-led legal, +privacy, and worker-representation review. + +That ordering remains correct even when a particular AgentCall use is not +classified as high-risk under the EU AI Act. The Act requires workplace notice +for deployment of high-risk systems and links deployer duties to GDPR impact +assessment where applicable; German co-determination separately covers technical +systems intended to monitor employee behavior or performance. Product +architecture can make those reviews easier, but cannot remove them. + +## Entry and exit gates + +| Phase | Enter when | Advance when | +|---|---|---| +| 1 — low-friction pilot | A sponsor and pilot security owner are named; 3–10 volunteer callees record informed opt-in; a two-week repeat-question/time baseline exists; the eligibility, rating, missing-data, time-diary, exclusion, deduplication, and security-severity protocol above is signed before launch | After at least four weeks and 50 eligible questions: at least 70% are owner-rated useful, weekly senior time spent answering eligible repeat questions falls at least 20% from baseline, and zero critical unauthorized disclosure or out-of-policy execution occurs. AgentCall's product owner and the customer's pilot security owner jointly decide advancement. | +| 2 — regulated non-EU | The public/enterprise C-track gate (#1–#8), organization administration and SSO (#15/#102), tenant audit/export plus implemented retention/erasure (#17), bounded call/runtime budgets (#181), and the exact endpoint policy promised to the buyer are shipped and tested | A regulated design partner completes security and procurement review using deployed evidence, and the customer's security owner accepts the documented residual risks without roadmap controls being presented as current. | +| 3 — EU profile | A named EU sponsor approves an exact regional claim and funds a deployment plan; customer privacy and worker representatives approve the pilot process before employee enrollment | A new regional deployment passes every verification step in the living residency adoption gate; retention and transparency behavior are tested; customer legal and security owners approve the exact deployed claim. | + +No calendar date substitutes for these gates. A named customer may change +priority, but it does not make an unimplemented control true. + +## Messaging contract + +Use: + +> Recover the hours senior engineers lose re-answering architecture questions. +> AgentCall routes each question to the colleague's own agent, under that +> colleague's policy, without retaining a central corpus of call content or +> indexing connected mail, chat, and documents into an employee-knowledge twin. + +Do not use: + +- “We do not ingest employee data.” +- “EU compliant,” “GDPR compliant,” or a residency claim derived from a + Cloudflare location hint. +- “Audit-ready” while tenant export, bounded retention, erasure, and delivery + evidence remain unimplemented. +- token- or currency-budget claims when the product can bound only calls and + agent runtime. + +## Why the alternative loses + +Starting in the EU because the privacy distinction is most valuable there would +combine three unknowns in the first sale: whether the workflow saves enough +time, whether employees accept it, and whether the deployment satisfies regional +and worker-governance requirements. A failed pilot would not reveal which +assumption was wrong. + +Dropping the privacy distinction to simplify the first sale would push AgentCall +into the shared-corpus category it is structurally designed to avoid. The right +move is sequencing: prove economic value where deployment is lighter, while +building the controls required to take the same architecture into harder +accounts honestly. + +## Revisit triggers + +Revisit this sequence when any of these occurs: + +1. a named EU design partner will sponsor the coordinated residency and + worker-governance work rather than accept roadmap language; +2. three Phase 1 pilots show that no-corpus architecture materially improves + conversion or shortens security review enough to outweigh the longer EU + adoption cycle; +3. Phase 1 buyers consistently reject the product despite measured time savings + because they need retained knowledge after employee departure; or +4. the hosted relay's confidentiality, residency, audit, and erasure contracts + change enough that the messaging limitations above are stale. + +## References + +- [EU AI Act, Regulation (EU) 2024/1689](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A32024R1689), especially Article 26 on deployer obligations and workplace notice +- [German Works Constitution Act §87](https://www.gesetze-im-internet.de/betrvg/__87.html) on co-determination for technical monitoring systems +- [Cloudflare Data Localization Suite](https://developers.cloudflare.com/data-localization/) and [Customer Metadata Boundary](https://developers.cloudflare.com/data-localization/metadata-boundary/), which are separate enterprise controls rather than defaults +- [Stack Overflow 2024 Professional Developers Survey](https://survey.stackoverflow.co/2024/professional-developers/) for the interruption baseline; the chosen company-size and regional beachhead remain an inference +- [AgentCall cloud data map and residency decision](../../security/data-residency.md) +- [Demand validation](../../research/2026-07-31-demand-validation.md) and [enterprise pivot research](../../research/2026-07-31-enterprise-pivot-research.md)