Roster updates
What changed, and when
Most directories are written once and left to rot. This one is tended. Every addition, removal, and verification sweep is logged here with a date, so you can see for yourself that the list is maintained rather than abandoned.
414 vetted providers on the roster.Last updated July 2026RSS feed
- Verified
We re-checked every listing that came from one bad source
A firm was removed on 27 July for not implementing AI at all, and it had reached us through a DesignRush list that mixes agencies which market AI in with agencies that build it. Eight other listings came from that same page, so all eight were re-checked. The good news first: none of them is a repeat. Six clearly build AI systems for clients. Two are borderline, and both now say so in their own descriptions rather than in a note only we can see. One is the AI brand of a 1999 outsourcing group whose eight independent reviews all describe data-labelling work, not implementation. The other opened its AI practice in March 2026 and has published exactly one AI case study, from 2023, with no figures in it. We also learned why that source misleads: it lets providers tick their own AI expertise boxes, and the borderline firm ticked nine. The list is claimed, not vetted, which is the whole reason this roster exists. The records themselves were another matter. Seven of eight were materially wrong. One profile named American Express, Whole Foods and Starbucks as a firm's clients when the firm only says products it developed are used there, and names none of them on its own client page. Another listed five certifications that turned out to be the regulatory standards its clients operate under, read off homepage graphics. A third described an independent San Francisco firm founded in 2016; it is a Los Angeles software agency trading under a brand whose domain was registered in April 2025. Corrections ran the other way too, and further. One firm's own website footer links its Clutch, GoodFirms and G2 profiles, and our harvest followed none of them, so 54 reviews at five stars were recorded as zero. Another had 31 case studies one click behind a menu item while we published two homepage testimonials in their place. Across the roster, the share of case-study links that reach an actual case study rather than a homepage moved from 43 percent to 47.
- Removed
Two removals: one sells no AI, one publishes no English
Coefficient is an Airtable and no-code consultancy, and the word AI does not appear anywhere on its website, its Airtable partner directory profile, or its LinkedIn specialties. The floor asks for a firm that offers AI implementation, so it does not clear it. This one stings, because on evidence they were better than most of the roster: Airtable Gold Services Partner confirmed in Airtable's own directory, and 66 dated reviews there at five stars from named people at Warner Bros. Discovery, Yale Ventures and YPO. The fault is ours for listing them, not theirs. Pymes AI publishes entirely in Spanish, with no language switcher and no English route, so a reader arriving from this roster cannot read it. That is our language rule, and they would clear it the day a complete English mirror exists. The same audit found the ISO/IEC 42001 badge in their footer links to a page describing a commitment to the standard, with no certificate number, scope, audit date or certifying body anywhere on it, and that the productivity figures we had published were unattributed homepage counters the cited case study never states. Their one documented engagement, by contrast, is genuinely well evidenced.
- Verified
Twenty more listings fact-checked, and the same three defects keep turning up
Eight of the twenty were materially wrong, eleven had smaller errors. The pattern is now boringly consistent. First, staff lists that are actually client lists: on eight of the twenty, the people we named as the firm's team were executives at its customers, lifted from testimonial bylines. Second, evidence links that point at a homepage instead of at the evidence; on one firm every one of eight case studies had its own page and our record cited none of them. Third, numbers that drifted: a review count of 150 that is 39 at the source, 50 Google reviews that are 22, a Make.com Gold partnership the vendor grades Silver, an Odoo Gold badge the vendor now grades Ready, a CMMI Level 3 appraisal that returns nothing in CMMI's own published results. Corrections went both ways. Several firms were better than our copy said: one we described as disclosing no case studies has a client-named build corroborated independently by the vendor, and another turned out to be an OpenAI Advanced Partner and an Anthropic services partner while we scored it as if it had neither.
- Removed
Two removals, neither of them about quality
Aimers is off the roster because it does not do the thing the roster is for. It sells paid media, conversion optimisation and landing page design, and nothing on its site describes building AI systems for a client; the only AI mentioned is automated bidding inside an ad platform. It is a good agency with 44 client-named case studies and 101 reviews across four directories, and it should never have been listed. The fault was ours: it came from a third-party list that mixes agencies which market AI products in with agencies that build them, and the seven other listings from that same source are being re-checked. Fractional AI is off because it no longer exists as a firm you can hire under that name. It was acquired in May 2026 and relaunched in July as a different company, and its own site now sends visitors elsewhere. Its case-study evidence was among the strongest here, which is exactly why the listing had to go rather than be quietly left to rot.
- Added
58 firms added, and the bar for being listed was lowered on purpose
The directory used to require a named person on the site and at least one piece of published evidence. Both are gone as requirements. A firm with a plain site, no case studies and no named staff now gets listed and scores low, because that describes a lot of real practices in a category this young, and a score of zero reads as failure when the truth is usually 'has not published much'. Nothing on the roster scores zero any more. What tightened in exchange is honesty: publishing invented evidence is now an explicit reason to be excluded. Of 22 firms rejected under the old rules and re-checked under the new ones, 10 should have been listed all along. The 12 that stayed out all publish something fictional, and each rejection records the URL and what renders there. Those 10 are now listed, scoring between 6 and 36, alongside 48 from the main intake. Their case studies link to the work rather than the homepage 96% of the time, against 28% for the existing roster, because the standard was written before the records were, not applied afterwards.
- Verified
Scoring rebuilt around what a buyer can check
Vendor partner badges used to be worth a fifth of the score. Only 9 firms of 360 could produce one confirmable in the vendor's own directory, and checking twelve claimed partnerships found twelve defective ones, so the signal was ranking almost nobody and doing it unreliably. Named client work now leads at 35 points, and is scored by strength rather than presence: a case study that links to the work counts double one that only points at a homepage. Published technical output, meaning code, marketplace listings and templates a reader can open, is second at 20. Two smaller components exist so honest firms are not stranded at zero, one for anonymised work carrying a real metric, which is what an NDA looks like, and one for simply stating what you do and who you do it for. Testimonials that name a person but no company were also dropped from every profile: 10 quotes across 6 providers that nobody could have checked.
- Removed
Four listings removed, for four different reasons
Re-checking listings one firm at a time has started producing removals, and the reasons are worth separating because they are not the same kind of problem. Pixelette Technologies published itself as "Official Secretariat of the British Government's AI policy body" in a homepage badge and, more consequentially, inside the structured data that search engines and AI assistants read. Parliament's own Register of All-Party Parliamentary Groups, as at 18 May 2026, names a different organisation in that role and does not mention Pixelette at all, and the APPG link inside Pixelette's own markup redirects to that other organisation's website. A false institutional credential in machine-readable form is not a wording slip. Allegations about its reviews were investigated separately and form no part of the decision: they trace to one pseudonymous account, which is not evidence. The AI Automation Agency listing was a website template nobody finished editing, still thanking the theme vendor in its testimonial and presenting the theme's stock placeholder names as its co-founders. DataTech.ai had not updated its site since 2019 and the one link we published as its case study never mentioned the firm. Monino Solutions is the odd one out and we want to be clear about it: a capable, well-reviewed firm with partner status we could confirm, removed only because it does not do AI work, which is the whole point of this list. Every reason is recorded in full, including what we deliberately did not count.
- Verified
We were ranking firms on the wrong thing, so we changed it
Until today a vendor partner badge was worth a fifth of a firm's score, more than anything except client work and reviews. Checking the roster properly showed that was measuring the wrong thing. Only nine of our listings, and one of the first sixty-one new firms we assessed, hold a partner status that can actually be confirmed in the vendor's own directory. A signal two percent of a category can obtain does not rank that category. And the programmes themselves largely reward sales volume: one platform describes its own partner scheme as a closed pilot. AI implementation is a young trade with barely any real credentials yet, and pretending otherwise flattered the firms with the best vendor relationships rather than the best work. What separates firms here is work you can go and inspect. One firm on this roster publishes nine components on a vendor's marketplace, one of them downloaded 56,400 times, which means thousands of other companies run its code. Another has no case studies and no reviews at all, and a public code repository it has maintained for three years. A third has the worst website we assessed all week and eighteen published automation templates with a current release cadence. Meanwhile a firm with eleven named staff, thirty polished pages and four directory listings publishes nothing and labels its own case studies representative. So published technical work now carries as much weight as reviews, named client work leads, engagement count counts on its own, and partner status is a tiebreaker. Client work counts wherever it is published, a post naming a client included, because where evidence lives matters less than whether it is specific enough to check. One honest caveat: we are auditing every listing one firm at a time, and until that finishes the ranking partly reflects how much we have verified rather than how strong a firm is. The firms at the top today are the ones we have looked at hardest.
- Removed
Partner badges now count for nothing until we have seen them at the vendor
A vendor partnership is the easiest thing on a profile to check and, it turns out, the least reliable thing we published. Of the first dozen firms whose partner claims someone actually opened, twelve were wrong in some way: absent from the vendor's own directory, a cited listing that returns a 404, a tier the vendor does not award, a certification program that does not exist, and in two cases a partner status the firm itself never claims anywhere on its site. We had 333 such claims on file across 151 firms, and they were carrying up to a fifth of each score. So the default is inverted. A partner claim now earns nothing until someone has found it in the vendor's directory, rather than earning credit until it is disproved. Nine firms currently clear that bar. That number will climb as the audit works through the roster, and every one of them will have a link you can follow. The immediate effect is that 65 firms no longer sit above the evidence line, and the count of firms we describe as having independently corroborated evidence fell from 167 to 102. The roster got smaller in what it claims, not in what it holds.
- Verified
The evidence score is now calculated, not stored
Every listing carries an evidence score, and until today that number was written down by whichever pass last touched the record. Which meant it went stale. Yesterday we corrected 81 facts across twenty listings and every one of those scores instantly described data that no longer existed, in both directions: one firm sat near the top on points awarded for a partner status that turns out not to exist in the vendor's own directory, while another sat near the bottom despite carrying sixty-one third-party reviews nobody had recorded. Nothing is stored now. The score, and the tier that follows from it, are calculated from the evidence on file every time this site is built. Correct a fact and the ranking moves on its own, with no judgment call in between. Recalculating moved 43 listings: 37 up, 6 down. Depth counts properly too, so a firm with sixty reviews now outranks one with six instead of both clearing the same bar.
- Removed
DataTech.ai removed, and one listing was filed under the wrong company
We have started re-checking every listing against the firm itself, one at a time, and the first ten produced a removal and a correction worth explaining. DataTech.ai is off the roster: its site has not been updated since February 2019, it names no client, it mentions no language-model or agent work anywhere, and the single link we published as its case study turned out to be a marketing article on someone else's site that never mentions the firm at all. It failed the requirement to show activity in the past year, and the evidence we had cited for it was not evidence. Separately, the listing we had under "AI Automation Consultancy" was not that firm's name. The site belongs to Uppit AI, a one-person automation practice in Chicago, and our record described a fifty-person San Francisco company founded in 2015 with two staff who do not exist. The profile now carries the right name at the right address, and the old link redirects. Both corrections came from checking, which is the only reason we know.
- Verified
We audited 20 listings against the firms themselves, and corrected 81 things
The roster promises evidence you can check. So we started checking our own, twenty profiles at a time: each firm searched, its site read page by page, and every claim in our record held against what the firm actually publishes. The first twenty produced 81 corrections. Fifteen were based in a different country than we said. One profile named a firm that had been renamed a year earlier. Two carried a metric that does not appear anywhere on the source page, including one where two separate case studies had been merged into a single false claim. Four certifications did not exist in the vendor's own partner directory, so they came off. Six profiles named the wrong people, one of them naming a company's marketing lead as its founder and another giving a co-founder and CEO someone else's job title. One firm's entire evidence record turned out to be unedited template text from its own website, placeholder names and all, and has been emptied. Corrections ran in both directions: nine firms had real third-party ratings we had missed, one of them 61 reviews deep, and they are now on the record. Separately, a check across all 367 listings found that only 18 percent of our case-study links point at the case study itself rather than the firm's homepage. That is the next thing we fix, and until it is fixed a link on a profile may take you to a home page rather than the proof. The audit continues at twenty a day.
- Removed
People sections were showing clients, not colleagues
Every profile has a section for the people behind a firm. An audit found that 235 of the 553 entries in it, spread across 106 firms, were not that firm's people at all: the harvest had copied the authors of their client testimonials into it. On 69 firms every single name was a client. One profile listed a template placeholder as a company's chief executive. Nothing in that section is now shown unless it can be told apart from the firm's own testimonial authors, which removed 238 entries and the section entirely from 70 profiles. Showing nothing is the right answer here. Naming the wrong person as someone's colleague is worse than leaving a gap, and a gap is honest about what we know. Placeholder names are filtered from client quotes for the same reason.
- Verified
Location QA: every provider's stated base is now resolved, not reprinted
Fifty-seven profiles were showing their location exactly as it came out of the harvest, and some of it was unreadable: a literal "null", a "Not specified", a fragment left behind where a parenthesis had been split off. Worse, a handful were filed under the wrong country, because a record whose country field read "delivery team in India" was being counted as India when the firm is in London. Locations are now resolved from the record rather than printed from it: the city leads, secondary-office notes no longer override a firm's actual base, and the country appears in one canonical form. Ten records whose city and country contradict each other outright now show only the part we can stand behind, and are queued for re-verification against each firm's own site. The same resolved values feed the structured data search engines and assistants read, so nothing published claims a country we have not confirmed.
- Expanded
The AI-native generation: agent development companies get their own segment
A third intake round judged 76 firms from the 2024-2026 AI-native generation, the boutiques that build on LLM APIs and agent frameworks directly rather than no-code automation, and accepted 42 against the evidence rubric. The roster now stands at 414 vetted providers. Every provider carries a new approach tag (AI-native, hybrid, or automation agency), the home roster can be filtered by it, and there is a dedicated ranked page for AI agent development companies. Seven production-stack signals, LangChain, LangGraph, CrewAI, MCP, RAG, vector databases, and evals, are now first-class, browsable tools, with evals being the sharpest marker of a firm that engineers and tests its AI rather than reselling automation.
- Removed
Language QA sweep: English-serving floor enforced
We audited every language-suspect record (non-English domains, en. subdomains, /en/ paths, and non-Anglophone home markets) for whether their English surface is real. Six firms whose sites are only in German, French, or Swedish, with no usable English mirror, were removed for failing the English-serving floor. Sixty-four keepers were classified and tagged as English-primary or full English mirror, and where a firm's real content lived on a translated mirror we now link straight to it. The roster stands at 414 vetted providers.
- Verified
Case-study re-harvest and sampling disclosure
A re-audit of Automation House found four documented case studies on their site where we had recorded one, so the profile was corrected and re-tiered on the fuller evidence. To stop implying our profiles are complete, every profile's Verification panel now says evidence was sampled from the provider's public site on a given date, with a one-click way to report anything off.
- Expanded
Roster passes 331 providers, and location pages go live
A second intake round judged 242 candidate firms against the evidence rubric and accepted 107, taking the roster to 414 vetted providers. On top of it we launched country pages: browse the roster by where firms are based across the UK, United States, Australia, Ireland, Canada, the Netherlands, and the DACH region, each with its own read on the local market.
- Expanded
Geo discovery: UK, Ireland, Netherlands, ANZ, Canada, DACH, Nordics
A second discovery pass surfaced 242 new candidate firms across seven regional sweeps, every one extracted from a page actually fetched. Candidates enter the roster only after an independent judge pass against the evidence rubric, so the roster now stands at 414 vetted providers and is still growing.
- Verified
Full verification sweep
Every provider was re-checked against the five evidence signals and given a public last-verified date. Evidence older than 24 months now counts at half weight in the ranking, so stale listings drift down on their own.
- Removed
Deduplication and floor filtering
The initial harvest of 227 candidates was deduplicated by domain and name and filtered against the Listed floor, removing anonymous shells, product-only vendors, and label-slappers that could not clear a single evidence signal.
- Launched
The Expert Roster launched
An independent, evidence-vetted index of AI implementation consultants and automation agencies, segmented by industry, use case, and tool. Ranked by evidence, never paid.
How vetting works, and what earns each tier, is documented in the methodology. New providers arrive through an ongoing harvest and judge pass, so this log grows as the roster does.
Browse the roster