What we got wrong, and what changed.

1 October 2026, methodology and standing pages

The methodology, the tier specification and the standing pages were rebuilt on one design. Every statement on them was read again against the methodology itself and against our list of verified third party facts. These are the material changes.

The AIUC-1 crosswalk carried codes that are not AIUC-1's. Methodology v2.0 mapped each dimension to control codes such as SR-1, DIQ-1, ASC-1, OR-1, GA-1, SIA-1 and AOC-1, and described AIUC-1 as an industry reference written by a consortium of underwriters. Neither matches what the Artificial Intelligence Underwriting Company (AIUC) publishes. The column now names the nearest of AIUC-1's six pillars, marked as our own reading, and the note describes AIUC-1 as AIUC describes it: 51 requirements and 130 controls across six pillars.

An illustration broke the methodology's own floor rule. Example A in Section 3 of v2.0 was labelled Advanced at an overall score of 63, with Context Integrity at 5. The Advanced floor is 6, so the floor rule places it at Certified. The example now says so and shows why.

Claims about a body of assessments and about insurer behaviour were removed. No organisation has been assessed under this methodology as of the last confirmed date, 17 August 2026. The v2.0 tier descriptions nevertheless said that Advanced was "the most common tier for operators seeking their first certification", that most In Progress operators reach Certified within three to six months, and that v3.0 would draw on incident data from certified deployments. They also said that some carriers apply more favourable terms to higher tiers, that most AI liability applications require at least Certified, and that underwriters treat Elite agents as a separate risk category. The tier specification said Certified was "the minimum bar most European insurers will begin asking for by the second half of 2026". None of this could be supported, and all of it is gone.

Legal references were corrected.

• Article 27 of the EU AI Act was described as a duty to register high risk systems in the EU database. It concerns the fundamental rights impact assessment that certain deployers must carry out.
• The proposed AI Liability Directive was cited as an emerging instrument. The European Commission withdrew the proposal in 2025. The reference now points to the disclosure of evidence provisions of the revised Product Liability Directive, Directive (EU) 2024/2853.
• Two references to an EIOPA supervisory statement of 2024, with section numbers, now cite the Opinion on AI governance and risk management published on 6 August 2025.
• Titles given to ISO/IEC 42001 Annex A.6 and A.8 did not match each other or the standard; the clause numbers are kept and the titles removed. Two section references to NIST SP 800-207 and IEEE 7000-2021 were narrowed to what could be matched.
• A statement that regulators are extending NIS2 accountability to AI by supervisory guidance was removed.

A review date that has passed. Methodology v2.0 gave 1 October 2026 as the target publication date for v3.0, after consultation in August and September. v3.0 has not been published. The date was removed; v2.0 remains in force until v3.0 is published, and the review dates will be announced on the methodology page.

The methodology page had carried version 1.0. It showed the 720 denominator and the one to ten scale that v2.0 replaced in April. It now presents v2.0, with the formula on a 0 to 10 scale over a maximum of 1,000 weighted points.

Commercial statements were withdrawn. The partners page listed indicative prices for partner tiers, said that carriers, investors and procurement teams use the methodology for due diligence and underwriting, and mentioned licensing of a certification dataset. The partners and editorial standards pages also referred to assessment services at published rates. No partner programme has launched, no rates are published and no dataset exists. All of it was removed.

The request page offered a Q3 2026 assessment window that has closed. It now describes the four steps and the four to six week cycle without a window.

Contact addresses. Mail addresses on the agentcertified.eu domain were listed on several pages, and they do not receive mail. They were removed, and every enquiry now goes through the request page.

17 August 2026, currency and evidence review

A second review, this time of whether what the site says is still true rather than whether it was ever true, found one class of error running through most of the site and a second, smaller class of invented evidence. Both are recorded here in full.

The regulatory calendar was out of date everywhere. The EU AI Omnibus entered into force on 27 July 2026 and moved the high risk obligations under Regulation (EU) 2024/1689 to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems embedded in physical products. Until that day this site said otherwise in several places.

• The site wide banner, present on every page, read that the Omnibus was at trilogue on 28 April 2026 and that the deadline "may shift". It was labelled Live. It was rewritten to state the position as it stood, with the dates.
• Nineteen editor's update and regulatory note panels described the Omnibus as a proposal that had not been adopted, several of them stating that the 2 August 2026 date "remains legally binding". Every one was rewritten and dated.
• The article published on 3 August 2026 under the headline "What certification evidence deployers need now that enforcement is live" was built on the premise that the Omnibus deferral had failed to arrive in time and that the original 2 August 2026 date had governed. That premise was wrong on the day it was published, six days after the Omnibus entered into force. The article's substance, the five elements of the Article 26 operator file and how they relate to certification, was correct and was kept. Its framing, headline, standfirst, key takeaways, FAQ answers and references were rewritten. It is now titled "What evidence an EU AI Act deployer needs to hold in 2026" and states the current dates. The URL did not change, so anyone who bookmarked or cited the old page lands on the corrected one.
• Four further articles stated that high risk or conformity assessment obligations applied from 2 August 2026 without qualification. All four now carry the current dates. Statements that Article 50 transparency obligations apply from 2 August 2026 were checked and kept, because the Omnibus did not move them.
• A deadline was published two months later than it falls. The Article 50(2) transition, which lets systems already on the market before 2 August 2026 delay only the machine readable marking of AI generated content, was given across this network as ending 2 February 2027. That was the European Commission's proposed six months reported as though it were law. The transition ends 2 December 2026. Anyone who planned marking work against February had two months less than they were told. This site described the transition without a date; it now carries the correct one on both pages that discuss it.

Claims about a body of assessments that does not exist were removed. No organisation had been assessed or certified under the Agent Certified methodology. Four passages implied otherwise and were corrected.

• "In assessments conducted against the Agent Certified framework, prompt injection vulnerabilities are the most frequently identified high-severity finding" was removed. There are no such assessments. The passage now attributes the ranking to the OWASP Top 10 for Large Language Model Applications, where prompt injection is LLM01, which is the actual basis for the dimension's weighting.
• "Most AI agent deployments this framework has assessed did not enter the process with governance documentation" was removed and rewritten as an assumption the framework makes, no longer presented as an observation it has recorded.
• "Reviewers routinely cite Elite agents as reference profiles when producing sector guidance" and "Used as a reference profile for insurer underwriting models" were removed. No agent has been assessed at the Elite tier, and no insurer uses this framework as an underwriting input. The comparison table entry "Priceable risk for insurers" was removed for the same reason.
• Two statements that an assessment takes a given number of weeks "for most operators" were rewritten as the scoped duration of the process, because there is no population of operators to average.

A third party was described more strongly than its own record supports. The cost article cited "the EEOC's 365,000 US dollar 2023 settlement over a discriminatory AI hiring tool". The settlement is real: EEOC v iTutorGroup, 365,000 US dollars, announced 11 September 2023, more than 200 affected applicants. The EEOC's own announcement describes tutor application software programmed to reject female applicants aged 55 or older and male applicants aged 60 or older, and does not call the system AI. Every reference on this site now says what the EEOC says.

Two dates could not be confirmed at source and are marked as such. The adoption of the Council of Europe Framework Convention on Artificial Intelligence in May 2024 and its opening for signature in Vilnius in September 2024 are stated on the Council of Europe's own published record, but coe.int, rm.coe.int and eur-lex.europa.eu all refused automated access during the review, so neither date was read against a retrieved document. Rather than remove the dates or quietly leave them, the article carries a dated sourcing note saying exactly that, and warning that a search engine confirming the dates may be drawing on that page.

The carrier alignment language removed on 15 August was checked again across every page, every JSON-LD block, the sitemap and llms.txt. It was gone and had not returned.

15 August 2026, site wide factual review

A review of every published page removed statements about named insurers and reinsurers that could not be verified at the administering company's own domain.

• Statements that the Agent Certified assessment report is formatted to align with the supplemental AI questionnaires of named carriers were removed. The report follows the Agent Certified methodology and no insurer's questionnaire. Agent Certified has no arrangement with any insurer or reinsurer.
• Descriptions of specific policy terms, schedules, notification periods and coverage triggers attributed to named carriers were removed, because they could not be verified at source.
• Coverage limits and funding round details attributed to named carriers were removed for the same reason. The qualitative statements they sat inside were kept.
• Statements that named carriers give preferential pricing to operators assessed under this framework were removed. No carrier has adopted the Agent Certified methodology as an underwriting or pricing input.
• The full name of the Artificial Intelligence Underwriting Company was normalised across the site, and unverified dates attached to the first AIUC-1 backed policy were removed.
• References to Agent Certified as a registry were replaced with accurate wording. Agent Certified publishes a methodology and issues assessments; it is not a registry, an issuing authority or a governing body.
• Brussels was removed as a location in page footers. Future Proof Intelligence has no Brussels presence.
• Two citations to European Insurance and Occupational Pensions Authority material were corrected to the Opinion on artificial intelligence governance and risk management published on 6 August 2025, replacing a placeholder document number.

15 August 2026, second pass on the AIUC name

A later review the same day found five further places where the acronym AIUC was expanded incorrectly, all in reference lists and footnotes, which is why the first pass missed them.

• AI Underwriting Consortium, in the certification assessment guide, the recertification article and the methodology reference table, was corrected to Artificial Intelligence Underwriting Company.
• AI Underwriting Code, in the dynamic behaviour audit article, was corrected to the same. The citation in that entry pointed to aiuc.org, which is not the company's domain; it now points to aiuc.com.

A network wide correction issued the same day removed the description of Munich Re aiSure as a parametric product, because Munich Re does not use that word for it. No page on this site made that claim, so nothing here changed. Where this site uses parametric it refers to the insurance model in general, which is ordinary industry vocabulary and remains accurate.