AI Dispatch — Daily Dispatch, September 15, 2026

An evidence-bound index of AI agents and the AI assurance/supplier stack — insurance, evaluation, audit, observability, guardrails, escrow, legal. Every field carries a source URL, a retrieval timestamp, and a verbatim quote. No estimates, no growth-hack framing.

## Sponsor disclosure

No sponsored placement in this issue. `listings` holds zero rows of any tier or status, checked directly at draft time. There is nothing to disclose because nothing has been sold.

## What published in the last 24 hours

**The Trust Gap, Vol. 7** ([full report](https://aidispatch.news/articles/trust-gap-vol-7-2026-09-14)) is the story of the week: we built a promotion gate to screen candidates on the way in, then ran it backwards against the 792 entities already published. **382 (48.23%) fail it today.** The largest single blocker, covering 230 entities, is a citation that no longer verifies against our newest capture of the same page — the gate's freshness check only ever fires on the transition into `published`, so nothing has re-tested an entity once it clears that door. A rebuilt staleness detector found that of 7,291 re-fetched citation values across the published index, 1,563 (21.4%) no longer verify, up from 18.1% a week ago on a smaller sample. Two findings worth sitting with: the six rarest fields in the 63-field catalogue, on an identical 4,752-slot denominator, are the same six risk-transfer fields in an unbroken block — 787 of 792 published entities (99.4%) show no public evidence of insurance availability at all — and the automated scorer that fills gaps between fact-checker passes has now gone six volumes without ever writing a 3, and this week without writing a 0 either, on a scale meant to run 0–3.

**Ozor**, an AI video-generation platform, added programmatic access for other AI systems: "Ozor ships a keyed REST API for generating and exporting videos programmatically, an official n8n node for automation workflows, and an MCP (Model Context Protocol) server so AI agents like Claude and Cursor can create Ozor videos directly," the company says. Its public materials do not say what happens once that server is used by an agent instead of a person clicking through a browser: no stated runtime spend or scope limit on agent-driven sessions, no audit trail for generated content, no insurance or indemnification terms, no compliance certification. ([full item](https://aidispatch.news/articles/ozor-mcp-access-no-audit-trail-insurance-2026-09-15))

**FastPhoto**, a headshot generator, trains a personal image model on photos users submit of their own face and says "Your photos are encrypted with AES-256 and used solely to train your personal AI model. We never share your data or use it for public model training." That claim is the extent of its published data-protection posture. Its dedicated data processing agreement page — the document built to spell out exactly this kind of obligation — names no insurance coverage, liability cap, indemnification term, or compliance certification either. ([full item](https://aidispatch.news/articles/fastphoto-no-insurance-liability-terms-2026-09-15))

**Field-level correction, not a wire item but worth flagging here:** today's fact-check pass found 91 of its 112 downgrades were policy boilerplate — a nav link reading "Security and Compliance" or "Privacy Policy" filed as evidence that a compliance certification or safety policy existed. Four already-published pages (`avoma`, `adcreative-ai`, `agent-creator`, `kay-ai`) had been carrying an automated `compliance` score of 1 sourced from exactly that kind of heading; those scores are now corrected to a judged 0. The pages themselves are unchanged — what changed is that a section title is no longer read as a finding.

## What changed in the index this week

Queried directly against the live database at draft time:

- **Published entities: 794** (764 agent, 30 vendor). Candidates: 2,309 (2,278 agent, 31 vendor). Archived: 2. - **17 candidate-to-published promotions, 1 published-to-candidate demotion** in the trailing seven days (net +16). In the last 24 hours: 2 promotions (Ozor, FastPhoto), 0 demotions. - **4,861 `entity_fields` rows touched** in the trailing seven days — 998 new or corrected `present` values, 3,863 flipped to `no_public_information`, most of that the fact-check pass described above. - **2,030 `change_queue` items resolved against 1,406 newly detected** in the same seven days — pending backlog stands at 2,568, up slightly from 2,533 in the previous issue. Resolution still outpaced detection this week; the backlog figure moves on its own rolling seven-day window and is not directly comparable day to day. - Full five-dimension scorecard coverage, human-scored dimensions only: **98 of 794 published entities (12.3%)**.

## One assurance/insurance development

**Armilla AI now names a carrier.** A capture retrieved September 11 supports a new `insurance_carriers` field: "Armilla Insurance Services is a Coverholder at Lloyd's. Affirmative AI Liability Insurance is underwritten by certain underwriters at Lloyd's." It is the first documented link in the index between the AI assurance layer and a named insurance market, and closes part of a gap this outlet has flagged before: no prior published entity named coverage at Lloyd's, Chaucer, or Mosaic specifically.

Read it for what it is, not more. Armilla is an insurance business describing its own product to its own customers — the same shape as ElevenAgents' page, which says it is "eligible for AI insurance through AIUC, **once certified**" — conditional eligibility, not cover in force. Across the index's human-scored dimensions: **`insurance_indemnity`, 101 of 102 (99.0%) at 0 or 1**, 98 of those a flat zero; **`outcome_settlement`, 95 of 98 (96.9%)** at 0-1; **`audit_trail`, 86 of 99 (86.9%)** at 0-1. Separately, `coverage_limits` — six straight volumes of this outlet's own Trust Gap report at a flat zero — moved to 1 of 792 this week, entirely because Coalition, a cyber insurer, published its own "Up to $100K additional Funds Transfer Fraud coverage" sub-limit. No AI agent in the index discloses a coverage limit of its own. Both of this week's apparent insurance movements are the same artefact: we published an insurer, not evidence that an agent got insured.

The competitor wedge holds, reconfirmed at draft time: across theaiagentindex.com (378 listed, 0 featured) and aiagentsdirectory.com (3,041 listed, 13 featured) — 3,419 combined — zero name any of the insurance or liability carriers this outlet tracks (Armilla, Munich Re aiSure, Lloyd's, Chaucer, Embroker, Mosaic, Relm, Testudo, Coalition, Vouch). Neither site runs an insurance, assurance, or audit category.

## Methodology

Full rubric and sourcing standard: /methodology. Every figure above was queried directly against the live database at draft time.

---

*AI Dispatch is not a growth-hacked directory. Our reader is someone doing diligence on an AI agent or its supplier stack before money moves. We publish absence as a finding, we cite everything, and we publish our own errors. If we ever stop doing that, stop trusting the newsletter.*