Methodology: How vettedaiagents.com Vets AI Agent Vendors

The trust register of an editorial directory depends on transparency. We publish the methodology for three reasons: so buyers can audit the process and verify that we applied the criteria consistently; so vendors know exactly what bar they need to clear; and so the site is auditable if a vendor disputes a profile.

Editorial vetting is not certification. We are not Schellman, BSI, or Bureau Veritas. We do not conduct on-site audits, do not run penetration tests, and do not review vendor documentation under NDA. We apply a published set of criteria using public sources. The boundary of what we claim to have done and what we have not done is named explicitly throughout the site.

The seven criteria are available at /vetting-criteria. The source index is at /sources. The watchlist is at /watchlist.

We identify candidate vendors from industry publications, third-party software marketplaces and buyer-guide aggregators, Schellman and BSI certified-vendor lists, GitHub awesome-AI-agents lists, the Agentic List conference directory, sister-site research on aiagentforcustomerservice.com and aiagentforsales.com, and submit-your-agent form submissions. At this stage we are asking a simple question: is this a real AI agent vendor with a real product in production? We are not applying the quality bar yet.

For each candidate we verify the seven criteria using public sources only. Certifications: vendor trust center URL cross-checked against Schellman, BSI, or Drata public registries where available. Named references: vendor customer page cross-checked against the customer's own blog post, press release, or case study. Pricing: the vendor pricing page if public, otherwise the vendor's publicly disclosed pricing model on its own site or docs; third-party aggregator listings are consulted only as context and are not quoted as a vendor's price. Founding date and team: LinkedIn, Crunchbase, and vendor about page. Every verified claim receives a source URL stamped in the dataset.

We apply the seven criteria from the vetting-criteria manifesto. Each criterion is scored pass, partial, or fail per vendor. A partial is used where a criterion is met in part: for example, a vendor with SOC 2 + ISO 27001 but not ISO 42001 passes the security criterion partially. Five or more passes lists the vendor. Four or fewer passes places the vendor on the watchlist with the specific gap named. Watchlist placement is not a rejection; it is a transparency mechanism.

The vendor profile is written or the matrix mini-card is prepared. Every profile carries a Vetting last reviewed [Month] [Year] stamp. At v1 the profile is reviewed by a single editor before publication. The profile is reviewed again when the rolling re-vet reaches it, or sooner if a material change pulls it forward. The dataset schema includes a vetting_last_reviewed field, and the stamp rendered on each profile is read from that field rather than written by hand.

What we cite for prices:a vendor's own pricing page or a public marketplace listing (for example AWS Marketplace or a public cloud marketplace, with date accessed) is the only surface we quote a price from. A disclosed pricing model on the vendor's own site, docs, or blog qualifies a vendor for the pricing criterion even without a list price. Third-party aggregator buyer guides and analyst commentary are used only as indicative context to sanity-check a claim; we do not present their figures as a vendor's price. We also cite vendor trust centers, vendor customer pages, press coverage in named publications, standards body references (iso.org, aiuc.com, schellman.com), and sister-portfolio sites where directly relevant.

What we do not cite:posts on X (formerly Twitter) unless from the vendor's official account in a factual announcement, Reddit comments, anonymous LinkedIn posts, AI-generated content (including from this site's own tools), and any source we cannot independently verify the author of.

Aggregator and paywalled sources:figures from third-party aggregators are not vendor-published prices, so we do not quote them as a vendor's price even where they are public. Where a vendor does not publish pricing, we say so and describe the pricing model, rather than substitute a third-party estimate.

Vendor-claimed numbers: deflection rates, accuracy benchmarks, customer-count ranges, and similar metrics that come from the vendor and are not independently audited are flagged vendor-claimed in amber on the profile. We report them because they are procurement-relevant information; we flag them because they have not been independently verified.

Building effective agents reference: the patterns and reliability framework at buildingeffectiveagents.com is used in the methodology context where the Anthropic agent patterns reference applies.

Vendors that pass four or fewer of the seven criteria are placed on the watchlist. The watchlist is a transparency mechanism, not a rejection. It exists so buyers can understand why a vendor they may have heard of is not yet on the main list, and what would qualify them.

Every watchlist entry names the specific criteria that are unmet and what evidence would move the vendor to the main list. The four watchlist vendors at v1 launch are 11x, Crescendo, Hebbia, and Adept / research-stage. See /watchlist.

A vendor moves off the watchlist by passing five or more criteria at its next re-vet. Major changes (acquisition, certification loss, pricing model shift) pull a re-evaluation forward out of turn.

Re-vetting is rolling, not a single scheduled sweep of the whole dataset. Each profile carries its own Vetting last reviewed [Month] [Year] stamp, and that stamp, not any site-wide date, is the one a procurement team should read. Of 20 entries in the current dataset the most recent review is October 2026; the oldest is April 2026, which still covers 7of them. Where a stamp is old, treat the profile as a starting point for your own diligence rather than a current attestation, and check the vendor’s trust center directly.

A material change pulls a profile forward out of turn: a vendor loses a certification, gets acquired, shifts pricing model significantly, or discloses a new certification that changes its criteria score. Worked examples of each on this site: NiCE closed its acquisition of Cognigy on 8 September 2025, Zendesk completed its acquisition of Forethought on 26 March 2026, and HubSpot moved its Breeze Customer Agent and Prospecting Agent to outcome-based pricing effective 14 April 2026.

Affiliate and referral relationships: where a listed vendor runs a public partner program (Intercom, Apollo, Clay, Anthropic, Aisera), the site uses neutral non-affiliate links in the vendor profile body. Affiliate links are confined to a clearly labelled Buyer Assistance sidebar block with visible disclosure. Editorial scoring is never affected by affiliate availability.

Sister-site relationships: we own aiagentforcustomerservice.com, aiagentforsales.com, and related portfolio sites. These are disclosed here. Customer-service vendor profiles link once to aiagentforcustomerservice.com for pricing depth. Sales vendor profiles link once to aiagentforsales.com for pricing depth. These links are one-way and contextually relevant, not reciprocal.

No display advertising at launch: no Carbon Ads, EthicalAds, or network display advertising at launch. The trade-publication register requires editorial independence from display advertising. Reconsidered only at twelve months if traffic is at scale and placement is genuinely contextual.

The boundary of editorial vetting: we are not a certified compliance auditor. We do not conduct on-site audits, penetration tests, or private security reviews. We apply published criteria using public sources. The disclaimer on every page of this site makes this explicit. Buyers must verify directly with vendors before purchasing.