# Claim ledger — mid-2026 field report

Every material claim in the published report. Status reflects the REMEDIATED text (July 31, 2026). Actions taken against the July 30 draft are recorded. Types: F=verified fact · V=vendor-reported · T=third-party reported · E=estimate · C=Cadenflow calculation · O=case observation · I=interpretation/thesis · P=forecast.

| ID | Section | Claim (final wording basis) | Type | Source / support | Action vs. draft | Status |
|---|---|---|---|---|---|---|
| 01 | Findings/S2 | Published prices for AI-handled support work span ~$0.05 (token-metered conversation, Hugo) to $2.99 (verified resolution with human backup, Crescendo) — across DIFFERENT billing units | F+T | pricing-snapshot.csv, per-row sources | Reframed from "same unit of work, 40x spread" → units named, multiple removed | GREEN |
| 02 | S2 | Price rises with accountability wrapped around the model, not inference consumption | I | Labeled Cadenflow interpretation over snapshot | Relabeled as interpretation | GREEN |
| 03 | S2 | Inference cost per text resolution ≈ half a cent to a few cents (scenarios shown) | C | calculation-notes.md §1, model list prices Jul 31 2026 | Replaced "roughly a cent" with shown scenarios | GREEN |
| 04 | S2 | Human-handled ticket ≈ $5–12 (third-party estimates); offshore VA $900–1,300/mo all-in | T/E | DigitalGenius via corpus; OnlineJobs.ph/Hurupay guides | Labeled third-party estimates; per-ticket derivation removed | GREEN |
| 05 | S2 | HubSpot Breeze $0.50/resolved conversation, effective Apr 14 2026, cut from $1.00; "resolved" = no human handoff within 72h | T | SaaStr Apr 9 2026; Resolve247 | Kept; definition added | GREEN |
| 06 | S2 | Fin $0.99/resolution official; Gorgias $0.90–1.00 + per-ticket fee; Zendesk ~$1.50–2.00 third-party estimates, no official list price | F/T/E | Gleap; Chatarmin; Richpanel | Kept; Zendesk explicitly third-party | GREEN |
| 07 | S2 | Zendesk resolution definition includes 72h customer silence; Gorgias merchant reported $14K overage on $13.5K/yr contract; "success tax" | T/O | Pricing Conundrum analysis; merchant report via corpus | Kept, labeled reported cases | GREEN |
| 08 | S2 | Parloa (Forbes BrandVoice Jan 6 2026) and Siena published anti-outcome-pricing manifestos; Sierra/Yuma/Zendesk sell outcome models; Decagon: majority of customers choose per-conversation, resolution definitions have "gray areas" | F/V | Forbes/Parloa; siena.cx blog; decagon.ai blog | Kept; Decagon quote is vendor's own | GREEN |
| 09 | S2 | Redo sells AI support at $0.85/resolution inside a free returns platform; PolyAI opened its platform to self-serve May 2026 (free first two months) | F | redo.com pricing; PolyAI PR May 18 2026 | Kept | GREEN |
| 10 | S1 | Sierra ARR: $100M (Nov 2025), $150M (Feb 2026) — company-reported; $950M round at post-money above $15B (reported $15.8B), May 4 2026 | V/T | TechCrunch May 4 2026; Sierra posts | "Company-reported" label added; valuation wording fixed | GREEN |
| 11 | S1 | Agentforce $1.2B ARR, +205% Y/Y (Q1 FY27, company-reported ARR term confirmed) | V | Salesforce PR May 27 2026 | Term verified; labeled company-reported | GREEN |
| 12 | S1 | Zendesk ~$200M AI ARR 2025, targeting up to $500M 2026 — company statements to press, private co, unaudited | T | The AI Economy/Letter Two Dec 2025 | Label added | GREEN |
| 13 | S1 | Fin crossed $100M ARR; Intercom renamed company to Fin May 14 2026; Salesforce definitive agreement ~$3.6B Jun 15 2026; signed not closed; close expected by early 2027 pending clearances | F/V | Intercom blog; Salesforce PR | Kept | GREEN |
| 14 | S1 | Zendesk acquired six AI companies since 2024; Forethought announced Mar 2026, since closed; NICE–Cognigy $955M (~25x revenue, reported); Qualified closed Apr 1 2026 | F/T | TechCrunch; Zendesk newsroom; press | Kept; multiples labeled reported | GREEN |
| 15 | S1 | Shopify June 2026 edition: free AI sales associate in Shopify Inbox; Klaviyo Customer Agent available across ~196,000-customer base (Composer in beta) | F | Shopify App Store; Klaviyo PR Jun 30 2026 | Kept | GREEN |
| 16 | S1 | OpenAI launched Presence Jul 22 2026; design partners BBVA, SoftBank, IAG; deployed by OpenAI FDEs; "boots-on-the-ground prices" per The Register | F/T | openai.com; The Register Jul 22 2026 | Kept; Register quote attributed | GREEN |
| 17 | S1 | In Cadenflow's frozen tracker: top-3 2026 rounds = ~87% of tracked 2026 equity dollars; all three late-stage at $3B–$15.8B; largest first-institutional 2026 round in set = $5M | C | funding-tracker.csv; calculation-notes §3 | REBUILT: replaces 32.5%→3.0% seed-share and "$3.76B/46 deals" (unreproducible, removed); "race is over" removed as fact | GREEN |
| 18 | S1 | Entering as a new horizontal platform now means facing funded leaders, free bundles, and labs moving up-stack — Cadenflow's strategic read | I | Labeled interpretation | Relabeled | GREEN |
| 19 | S1 | Sierra ~79x forward (estimate basis shown); Decagon ~130x company-reported / ~375x on Forbes estimate; Wonderful ~110x; Freshworks ~3.7x P/S (Cadenflow calc Jul 31); Influx <1x on Dealroom EV estimate | C/E | calculation-notes §5 | Zendesk "trades 6.6x" REMOVED (private since Nov 22 2022 — verified); bases labeled | GREEN |
| 20 | S3 | Vendor-marketed resolution claims commonly 67–85%; must be read against metric definitions (containment/deflection/resolution not comparable) | V/I | Vendor materials; taxonomy table | "30-point gap" chart REMOVED; replaced with metric taxonomy | GREEN |
| 21 | S3 | Gartner: only 14% of customer service issues fully resolved in self-service (survey of 5,728 customers, Dec 2023) — a self-service baseline, NOT an AI-agent rate | F | Gartner PR Aug 19 2024 | Relabeled correctly with field date | GREEN |
| 22 | S3 | Zendesk "41.2% median deflection / 58.7% top quartile" | — | NOT FOUND in any primary source | REMOVED everywhere | REMOVED |
| 23 | S3 | Practitioner benchmark "70–85% mature deployments" | — | Unsourced | REMOVED | REMOVED |
| 24 | S3 | Intent-level rates 78%/69%/19% | — | Stat-aggregator only | REMOVED; replaced with directional statement (structured intents automate better than complaint-type) labeled as practitioner observation | REMOVED |
| 25 | S3 | Hugo markets "up to 60%"; its named case studies land at 40–60% | V | hugo.ai; Crisp case pages via corpus | Kept (vendor's own materials) | GREEN |
| 26 | S3 | Fin: Salesforce PR cites avg 76% of volume resolved end-to-end (vendor-reported); competitor teardowns estimate 45–53% in production (competitor-authored, incentives disclosed) | V/T | Salesforce PR; teardown via corpus | "Independent" label removed; incentives disclosed | GREEN |
| 27 | S3 | Intercom survey (2,400+ support professionals, Jan 2026): 87% plan AI investment 2026; ~10% describe deployment as mature. Hiver (700+ US support leaders, Mar 2026): 14% say AI significantly improved resolution times; ~50% no significant cost/ticket reduction | T | Intercom report; Hiver report | Kept with n/dates | GREEN |
| 28 | S3 | 11x.ai: logos of non-customers (ZoomInfo legal threat), full-year ARR on 3-month break clauses, ex-employees describe 70–80% churn (TechCrunch investigation Mar 24 2025) | T | TechCrunch | Kept | GREEN |
| 29 | S3 | Decagon company-reported $30–35M annualized vs Forbes independent estimate ~$12M (2025); Decagon and Fin both claim to win bake-offs against each other | T | Forbes Dec 2025/Feb 2026; Sacra; Upstarts | Kept; "nobody's numbers can be trusted" softened to "treat unaudited private metrics as claims until triangulated" | GREEN |
| 30 | S3 | MIT NANDA: 300+ initiatives analyzed, 52 org interviews, 153 surveyed leaders; ~95% of integrated pilots showed zero measurable P&L return in window; ~40% report deployment of generic LLM tools vs ~5% of task-specific tools reaching production; external partnerships reached deployment ~67% vs ~33% internal (interview sample); authors state limitations | T | NANDA report (mirror PDF), verified Jul 31 | CORRECTED per Blocker D (was: 52 interviews only; 40/5 as build-vs-buy) | GREEN |
| 31 | S3 | Speed-to-lead: 2007 Oldroyd study measured contact/qualification odds by phone; Optifai 2026 commercial benchmark (939 B2B cos, observational) shows 2.6x on close rate — different metric, not a formal replication; 23% (HBR 2011 audit) vs 63.5% (RevenueHero 2024) never-contacted, different samples/definitions noted | T | Oldroyd PDF; Optifai; HBR/RevenueHero via corpus | Scoping added; "best replication" removed | GREEN |
| 32 | S3 | Market-size forecasts for the category disagree 2–3x with undisclosed methodology; investors price against the $300–400B/yr contact-center labor pool (budget-frame interpretation, not substitutable TAM) | T/I | Corpus; labeled interpretation | Reframed | GREEN |
| 33 | S4 | Klarna: Feb 2024 — 2.3M conversations month one, "work of 700 agents," ~$40M projected profit improvement (company); May 2025 — CEO quality reversal quotes, human rehiring begins; Q3 2025 earnings call — agent "can now do the work of more than 853 full-time agents"; total headcount 7,400→~3,000 over several years, mainly attrition, not all AI-attributed, not support-only | F/V | Klarna PR; Forbes May 2025; CX Dive Nov 20 2025; Fortune Oct 2025 | CORRECTED: 800+→853 sourced; 7,000→7,400; attrition caveat added | GREEN |
| 34 | S4 | CBA reversed 45 call-center cuts within ~3 weeks with apology (Aug 2025); Ford rehired ~350 veteran engineers (engineering, not CX); IBM automated ~94% of routine HR requests and plans to triple entry-level hiring (HR, not CX); Salesforce support 9,000→~5,000 with redeployments | T | Invezz roundup Jul 1 2026; Fortune Sep 2025 | Cross-domain labels added (engineering/HR ≠ CX evidence); "case file, not dataset" framing | GREEN |
| 35 | S4 | Gartner predictions: >40% of agentic AI projects canceled by end-2027 (Jun 2025); 80% of common issues autonomously resolved by 2029 (Mar 2025) — both PREDICTIONS | F | Gartner PRs | Labeled forecasts | GREEN |
| 36 | S4 | Gartner survey (5,728 customers, Dec 2023): 64% would prefer companies didn't use AI in customer service; 53% would consider switching over AI use. Forrester 2026 prediction: one-third of brands will erode customer trust through self-service AI | F | Gartner PR Jul 9 2024; Forrester Oct 28 2025 | REPLACES unverifiable longitudinal tracker (83→85/69→73/41→68 REMOVED) | GREEN |
| 37 | S4 | Reviewed cases and major vendor designs increasingly converge on a hybrid operating model; human support repositioning as premium tier (Klarna CEO quote) | O/I | Case file; labeled synthesis | "Everyone converges" softened | GREEN |
| 38 | S5 | Frontier models score near ceiling on τ²-bench (Sierra's own vendor-authored benchmark) — benchmark performance ≠ production reliability | T | τ²-bench results via corpus | Vendor-authored label + limitation added | GREEN |
| 39 | S5 | Recurring failure factors across reviewed cases: KB rot, escalation design, scope selection, absent QA — pattern in case file, not proven causality | O/I | Case documentation (Air Canada, Cursor, CBA, drive-thru, corpus) | "Never the model" absolutized claim removed; "25% of help articles" number removed (unsourced); "single biggest rage trigger" removed | GREEN |
| 40 | S5 | Moffatt v. Air Canada: BC Civil Resolution Tribunal, Feb 14 2024, negligent misrepresentation, CA$812.02 total awarded, "separate legal entity… remarkable submission" quote; small-claims tribunal decision, not binding precedent; lesson = reliance liability | F | 2024 BCCRT 149 | CORRECTED from "your bot's words legally bind you" | GREEN |
| 41 | S5 | Cursor support bot invented a nonexistent policy (Apr 2025), users canceled; Chevrolet dealer bot $1 Tahoe (Dec 2023); DPD swearing bot | T | The Register; press coverage | "$150K promo-stacking loss" REMOVED (unsourced); "case law of the absurd" phrasing removed | GREEN |
| 42 | S5 | Gorgias, Fin, Yuma (and Triple Whale outside support) shipped the same approval UX pattern: review queue, draft-in-composer, per-intent autonomy dial, feedback affordance — from product docs/materials | T/O | Gorgias docs; Intercom Copilot docs; Yuma materials; TW Moby 2 PR | "Now the expected shape" labeled as read of shipped features | GREEN |
| 43 | S5 | Service-layer evidence: OpenAI Presence ships with FDEs; Sierra implementation can equal/exceed year-1 license (reported); Crescendo >$100M ARR selling resolutions with humans, bought PartnerHero; Wonderful 28 offices for ~80 customers + CEO quote; FDE among fastest-growing AI job titles (reported) | T/V | Presence PR/Register; OpenNash/Sacra; Crescendo PR; Wonderful press | Kept with labels | GREEN |
| 44 | S5 | Yuma: LinkedIn-visible headcount snapshot (Jul 2026, ~30 profiles): roughly two-thirds GTM/delivery roles, ~6–7 visible engineers — indicative snapshot, not audited headcount | C/O | Corpus GTM teardown (profile count method stated) | Method + limitation added | GREEN |
| 45 | S5 | In the 14-company teardown set, every vendor pairs the product with mandatory onboarding/implementation/account management at its core price point, per their own public materials (definition + list in method) | O | 14 teardowns; pricing-snapshot | "Plug-and-play" defined; scope bounded to the set | GREEN |
| 46 | S5 | ICONIQ 2026 snapshot (~300 execs): AI companies project ~52% gross margins for 2026; actuals 45% (2025), 41% (2024); classic SaaS benchmark 80–90% (industry benchmark, separate attribution) | T | ICONIQ report; SaaStr summary | Kept; causal "margin went into service layer" → labeled consistent-with interpretation | GREEN |
| 47 | S6 | EU AI Act Art. 50(1): providers of AI systems interacting with natural persons must ensure disclosure unless obvious to a reasonably well-informed person; applies from Aug 2 2026; EC final guidelines adopted Jul 20 2026 | F | AI Act text; EC guidelines page | Scoped correctly | GREEN |
| 48 | S6 | Digital Omnibus (Reg (EU) 2026/1744, OJ Jul 24 2026, in force Jul 27): postponed Annex III high-risk to Dec 2 2027 / Annex I to Aug 2 2028; did NOT postpone Art. 50; pre-Aug-2 generative systems have until Dec 2 2026 for machine-readable marking — four separate things, stated separately | F | EUR-LEX; Lewis Silkin; Jones Walker | Separated per Blocker E | GREEN |
| 49 | S6 | Art. 99(4)(g): transparency violations up to €15M or 3% of worldwide turnover, whichever is HIGHER; SMEs: whichever is lower (Art. 99(6)); actual enforcement depends on violation and authority | F | artificialintelligenceact.eu/article/99 | "Up to" + SME rule + enforcement caveat added | GREEN |
| 50 | S6 | Art. 50(4) human-review exemption concerns AI-generated PUBLIC-INTEREST TEXT with editorial responsibility — NOT a disclosure exemption for support replies; human review remains a governance/evidence practice, without that legal effect | F | AI Act Art. 50(4) | CORRECTED per Blocker E (compliance-instrument claim removed) | GREEN |
| 51 | S6 | California SB 243 and NY GBS §1702 are COMPANION-chatbot laws that expressly exclude customer-service bots; B.O.T. Act requires intent to mislead in commercial/electoral contexts — adjacent context, not general CS-bot law | F | LegiScan SB 243 text; NY GBS §1700/1702; BPC §17941 | CORRECTED per Blocker E | GREEN |
| 52 | S6 | "EU-hosted" as procurement trend: sovereign-AI is a growing European procurement category (e.g., SAP EU AI Cloud) — trend observation; vendor-audit claim removed | T/I | Press/corpus | "Almost no major US vendor" REMOVED | GREEN |
| 53 | S7 | Voice: overtook text as Sierra's primary channel Oct 2025 (company-reported); infra ~$0.05–0.10/min and human calls $5–16 (third-party estimates); SMB voice = Cadenflow thesis | V/T/I | Sierra/Forbes; corpus | Labels added | GREEN |
| 54 | S7 | Machine customers: Gartner warns of inbound machine customers; Zendesk exposes MCP as a channel; Shopify ships merchant-verified knowledge for agents; MCP 10,000+ public servers under Linux Foundation (Dec 2025) — early signals, not "already standard" | F/T/P | Gartner; vendor announcements | "Already standard" softened to signals | GREEN |
| 55 | S7 | Verification-as-product (Zendesk independent evaluation model billing) = observed pattern; "can't be bundled away" = forecast/thesis | T/P | Zendesk materials | Labeled | GREEN |
| 56 | S7 | Inference deflation ~10x/yr for constant capability over the measured window (a16z/Epoch series) — historical series, not indefinite projection | T | a16z LLMflation; Epoch AI | Projection language bounded | GREEN |
| 57 | Meta | Byline source counts generated from source-register.md; "40+ primary" REMOVED; method section answers period/selection/limitations | C | source-register.md | Per Blocker F | GREEN |
| 58 | Meta | Author positioning: builder, not platform vendor; commercial interest disclosed (implementation studio); two e-commerce pilot deployments, two months, no client data used | F | context.md; disclosure block | Per §9 opening rules | GREEN |