The source-integrity rule
Every claim on this site traces to a named source and a date — a vendor pricing page, a study, a named builder's published numbers. If a claim can't be traced, it doesn't ship. Where sources disagree or look stale, the page flags it instead of picking the convenient version.
How figures are labeled
- Vendor-claimed — numbers from a vendor's own marketing or case studies. Promising, not proof, and labeled as such wherever they appear.
- (est.) — estimates — prices triangulated from secondary sources because the vendor publishes no list price. Never quoted as firm.
- (secondary) — reported by secondary press or trackers rather than a primary vendor page; flagged for re-verification before anyone relies on them.
- Independent evidence — peer-reviewed studies and large-sample research (MIT/InsideSales, JMIR, Invoca, CallRail, Yelp/YouGov, AnswerConnect, US Census) are the anchors we lean on hardest.
- Self-reported — builders' own published figures in the newsletter research are flagged as self-reported; the flags carry over from the watchlist.
Things we never claim
- ❌ "62% of small business calls go unanswered" — it's a 2016 study of 85 businesses, misdated everywhere.
- ❌ Vendor case-study numbers as facts ("850% ROI," "90% of calls booked") — label them vendor-claimed.
- ❌ Enterprise agent prices as firm quotes — Agentforce, ServiceNow, Ada, Sierra, Harvey publish no list prices.
- ❌ "Intercom was acquired by Salesforce" — unverified, do not repeat.
- ❌ Settled TCPA one-to-one consent advice — sources conflict; refer to counsel.
Compliance content
Compliance flags on platform and use-case pages come from Part 5 of the Landscape Report:
- TCPA / FCC (US federal) — the big one. Feb 2024 FCC ruling: AI-generated voices = "artificial" voice → outbound AI-voice calls/texts need robocall-level consent. Marketing texts need prior express written consent (47 CFR § 64.1200). Penalties: $500–$1,500 per violation. ⚠️ Source conflict: the FCC's 2023 "one-to-one consent" rule was vacated by the Eleventh Circuit on Jan 24, 2025 (Insurance Marketing Coalition v. FCC) — per aloware (Sep 2026) and compliancehub (Jun 2026); FCC deleted the vacated language. One Sep 2026 vendor guide still describes one-to-one as in force. Treat as "verify with counsel" — do not advise on it as settled.
- Call recording consent — 12+ all-party/two-party states (CA, CT, DE, FL, IL, MD, MA, MT, NV, NH, PA, WA). Transcripts count as interception in several states. Default: disclose AI + recording at call open.
- AI-identity disclosure — EU AI Act Article 50 (effective Aug 2026): AI audio output must disclose its AI nature audibly at the start. Increasingly best practice in the US too.
- HIPAA/BAA — any PHI handling: signed BAA before data flows, TLS 1.2+/AES-256, RBAC, audit logs, minimum-necessary, retention/purge policies.
- Vertical-specific: attorney advertising/ethics (ABA Model Rules 1.1/1.4/1.6/5.3/7.1/7.2/8.4 + state bars); fair housing (real estate); PCI DSS for card-on-phone/chat (hosted payment links); FDA/FTC for med-spa treatment claims.
- Security posture — 51% of service leaders say security concerns delayed/limited AI initiatives (Salesforce State of Service, cited Sep 2026). Verify vendor data handling before any customer data flows.
This is research, not legal advice. Where the research itself found conflicting sources — the TCPA one-to-one consent rule is the live example — the site says "verify with counsel" rather than presenting it as settled.
Corrections
Prices and platform facts are re-checked as the research updates; the daily Atlas routine scans for launches, price changes, and deprecations (two are already on the calendar: Google Gems retire Nov 17, 2026, and Zendesk legacy bots end Dec 10, 2026). Found an error? It gets fixed publicly, with the same number shown publicly and privately — that rule is borrowed from our directory score guardrails.