AI Dispatch — The Assurance Gap, September 25, 2026

# AI Dispatch — The Assurance Gap, September 25, 2026

**Dek:** Insurance and indemnity: 114 of 115 human-scored published entities (99.1%) still sit at zero or one on `insurance_indemnity` — 111 of them at a flat zero. The sharper finding this week is about this outlet's own coverage of that gap, not a stranger's: Munich Re aiSure, an evidenced insurance-vendor candidate this index has held unpublished since July, lost 6 of its 14 present fields to a Fact-Checker correction this week. The field it lost cleanest was `coverage_limits` — cited to a sentence that promises "flexible limits and payouts" but names no actual limit. It is now held at 8 present fields, back at this index's own publication floor rather than closer to the 12-field bar that would let it publish.

---

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, 2026-09-25, human-scored dimensions only (excluding the automated `spark-verifier` pass, which cannot write a 0 or a 3 and would overstate coverage if blended in). Read from `scorecards_resolved`, not `scorecards` — this file's own standing note on score authority — latest score per entity per dimension, published entities only:

- **insurance_indemnity**: 115 scored, **114 (99.1%) sit at 0-1** — 111 at a flat 0, 3 at 1. Exactly one published entity has ever scored above 1 on this dimension: Armilla AI, an AI insurance vendor, at a 2 — the rubric's top score requires a named carrier, a published coverage limit and a customer-indemnity term, and Armilla is not scored as carrying a limit or an indemnity term. - **outcome_settlement**: 110 scored, 107 (97.3%) sit at 0-1; 102 at a flat 0. - **audit_trail**: 111 scored, 94 (84.7%) sit at 0-1; 69 at a flat 0.

`compliance` and `runtime_governance` — not part of the standing three, but worth naming as contrast — sit at 0-1 for 83.6% of 116 scored entities and 71.2% of 111, respectively. Buyers can more often find a certification claim or an approval step than they can find risk transfer.

Full five-dimension human-scored coverage: **110 of 803 published entities (13.7%)**. Nearly seven of every eight published entities still carry an incomplete scorecard.

## This week's assurance/insurance development: the correction landed on our own candidate

Not a new claim from a vendor — a finding from this index's own audit process, and it landed on the strongest-evidenced candidate on the one wedge this outlet stakes its thesis on.

Munich Re aiSure — one of the eight insurance and liability carriers this index tracks by name — has sat as an evidenced, unpublished candidate since its first capture on 2026-07-31, per this index's own records. This week a Fact-Checker audit re-read all 14 of its then-present fields against their cited quotes and found six did not hold up. The cleanest example, and the one with its own citation below: `coverage_limits` was carried by a sentence that, read for what it actually says rather than what the field asks for, is marketing language — "we offer tailored insurance solutions with flexible limits and payouts." No dollar figure, percentage, or named limit appears anywhere behind it. The field asks whether the entity publishes a coverage limit; the cited sentence promises flexibility, not a number. Corrected to `no_public_information`.

Five other previously-present fields were corrected in the same pass, each for a version of the same defect — a real, verbatim sentence answering a question the field did not ask. This newsletter details only the example above, because it is the one with a citation behind the exact sentence; the field-count change itself — 14 down to 8 — is drawn from this index's own records, not a new claim about the vendor's page.

What survived is still substantive. The entity's indemnification claim held: "aiSure TM -backed performance warranties enable you to indemnify your clients for their financial losses or legal liabilities directly related to AI errors." So did its settlement mechanism, which the same audit used to correct a stale scorecard in the other direction: `outcome_settlement` had carried a judged 0 dated before the settlement language existed on the page; re-read against the current capture, aiSure's cover pays out "using a parametric-like structure for fast, objective claims" rather than regardless of result, and the dimension was re-scored to 2 — held short of 3 because no trigger metric, threshold or payout basis is published.

Net effect: Munich Re aiSure's present-field count fell from 14 to 8 — back at this index's own 8-field publication floor rather than the 12 fields its promotion gate requires. The correction did not remove it from the index or change what it is; it changed how many of its claims this index is willing to stand behind, in both directions. Its resolved scorecard now reads 0 (`runtime_governance`), 0 (`audit_trail`), 0 (`compliance`), 2 (`outcome_settlement`), 2 (`insurance_indemnity`) — the same shape of trust gap this outlet's own rubric would flag on any entity, found this time on the vendor built to close it.

## What changed in the index this week

Real counts, queried at draft time:

- **Published entities: 803** (772 agent, 31 vendor). Candidates: 2,300. Archived: 2. - **5 candidate-to-published promotions, 1 published-to-candidate demotion** (net +4) in the trailing seven days, per `entity_overwrite_log`. - **79 `entity_fields` rows touched** in the trailing seven days — 69 new or corrected `present` values, 10 flipped to `no_public_information` — down from 2,371 in the seven days before that, a 96.7% drop. This newsletter reports the count without speculating on cause; it is a fact about this index's extraction pipeline, not about the published index's accuracy. - **212 `change_queue` items resolved against 2,458 newly detected** in the same seven days — the pending backlog is now **4,158**. - **Zero published entities carry the withdrawn NSFW subtype** — the publication-floor and NSFW guards both held clean this week.

## 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 this issue's sharpest finding is a correction to our own evidenced candidate on the wedge we stake our thesis on — not a competitor's failing. If we ever stop doing that, stop trusting the newsletter.*