AI Dispatch — The Assurance Gap, September 4, 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.

## The assurance gap, in real numbers

Queried directly against the live database, 2026-09-04, ~16:30 UTC. Of 780 published entities, 75 (9.6%) carry a complete five-dimension trust-gap score — unchanged from yesterday.

Among entities scored on each dimension:

- **insurance_indemnity**: 80 scored, and **all 80 (100%) sit at 0-1** — 73 at a flat 0, 7 at 1. No entity in the index has ever scored above 1 on this dimension. - **audit_trail**: 446 scored, 412 (92.4%) sit at 0-1; 47 at a flat 0. - **outcome_settlement**: 201 scored, 165 (82.1%) sit at 0-1; 70 at a flat 0.

`compliance` and `runtime_governance` — not part of the standing three, but worth naming as contrast — are less skewed: compliance sits at 0-1 for only 39.9% of its 429 scored entities, runtime_governance for 53.9% of 490. Buyers can find a certification claim or an approval step. They cannot find risk transfer.

## One sentence, two field keys, and a scorecard that shouldn't be a 2

That's the standing finding. What's new this week is where the evidence behind these scores comes from, and it does not hold up evenly.

A field is "on a reused quote" when the same verbatim sentence is filed under more than one `field_key` for the same entity — a page that says one thing about audit trails getting copied into `audit_trail`, `compliance_certs`, `regulatory_alignment` and more, each counted as if it were separate evidence. Queried directly at draft time: **4,129 of 12,645 present fields on published entities (32.7%) sit on a reused quote.** Cut to just the 24 assurance fields, the rate is worse — **1,109 of 2,945 (37.7%)** — and **674 of 780 published entities (86.4%)** carry at least one.

We can show exactly what that does to a published score, not just a field count. **Indemn** (published as `insurance-ai`) states on its homepage: "Compliance isn't an afterthought. Our architecture is designed to meet NAIC standards by default — with audit trails, escalation paths, and deterministic guardrails built into every Associate." That single sentence is filed verbatim under both `audit_trail` and `audit_trail_immutable`. Its `audit_trail` scorecard reads "Scored from 2 evidenced field(s): audit_trail, audit_trail_immutable" and comes out to **2 of 3** — one sentence, counted twice, buying a point it would not earn on its own merits: the sentence names no retention period, no tamper-evidence mechanism, and no customer access path, which is what a 2 on this rubric is supposed to require.

**Hapax** does the same thing in the other direction. "Enterprise-grade audit trails, role-based access, and compliance built in from day one" is filed under `audit_trail`, `compliance_certs`, `regulatory_alignment`, `data_handling`, `deployment_options` and `permission_scopes` — six fields, one sentence. Its `compliance` score reads "Scored from 2 evidenced field(s): compliance_certs, regulatory_alignment" for a 2 of 3, built from what is, on the page, one clause.

Neither entity did anything wrong; both pages say what they say. This is a defect in how we turned a sentence into six rows, not in the vendors' disclosures — and it means some of the trust-gap scores in this index right now are measuring the same evidence more than once, in the vendor's favor. We are publishing the finding rather than quietly re-scoring, because a newsletter that reports its own defect after the fact is more useful than one that reports a clean number it hasn't earned yet.

## The insurance count, checked again

`insurance_available` — can a customer deploying this agent actually get liability cover for what it does — has 4 present rows on published entities: **Coalition** ("Cyber Insurance ... Now Available: Active Cyber Insurance for Enterprises"), **RentAHuman** ("requester reputation and incident insurance funded from the fee"), and **ElevenAgents** / **ElevenLabs Agents** ("the first agent platform eligible for AI insurance through AIUC, once certified" — conditional eligibility, not cover in force). The last two are the same page (`elevenlabs.io/agents`), the same quote, the same evidence record, filed as two separate entities — a duplicate this newsletter first flagged on August 28 and it is still unresolved. Net of it, that's **3 distinct pages**, not 4.

Seven named insurance carriers we track — Armilla AI, Munich Re aiSure, Lloyd's of London, Chaucer Group, Embroker, Mosaic Insurance, Relm Insurance — remain candidates, not published, blocked by the same publication-floor tension flagged in prior issues: a general-purpose ≥8-field bar applied to a wedge category whose whole finding is that these pages barely mention AI by name. That has not moved.

## What changed in the index this week

Real counts, queried at draft time:

- **Published: 780** (2,283 candidates, 2 archived). - **16 entities promoted, 0 demoted** in the trailing seven days, per `entity_overwrite_log`. - **110 new `entity_fields` rows** in the trailing seven days — down sharply from the ~400/week pace of recent issues. This week's own operating log recorded the crawler stopping outright for over nine hours and no working deploy path from a merged fix to the machine that runs it; that is the most likely context, not a confirmed cause. - **329 `change_queue` items resolved**, against **1,500 newly detected**, in the same seven days — the backlog is growing faster than it's being worked.

## This week's assurance/insurance development

The standing claim holds on its latest sweep, re-verified today: of 3,354 combined listings across theaiagentindex.com (370) and aiagentsdirectory.com (2,984), **zero** name an insurance or liability carrier. Neither competitor has an insurance, assurance, or audit category. The insurance and liability layer — a named carrier standing behind an agent's mistakes — remains, on both public catalogues, entirely uncovered. It is the one part of the wedge that stays uncontested every time we check it.

## Sponsorship

No listing is sold 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.

## Methodology

Full rubric and sourcing standard: /methodology. Every count above was queried directly against the live database at draft time rather than carried forward from a prior issue.

---

*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.*