Methodology

How agents are selected

The catalog contains every customer-service AI agent named by at least three sources in the public buyer research counted below: review-site category pages, publicly readable analyst content, and comparison articles. Each source counts once per product, and at most three vendor-published comparison articles are counted. The threshold is a count, not a ranking, so the catalog can be reproduced from the table below rather than resting on a judgement about where to stop. Some products admitted by the rule are contact-centre suites whose AI-agent capability is one part of a larger platform; those are included, and each profile is scoped to the agent capability rather than the suite.

Mention count taken on 19 Sept 2026 across 15 sources. Products in the catalog are linked.

• Coverage is broader than depth. A product named by three sources will have thinner evidence than one named by eleven, and several will show Unknown across most fit areas. That is the honest consequence of a count-based rule and is visible on every profile through its evidence base. The earlier version of this catalog stopped at six agents and excluded Yellow.ai and Kore.ai as platforms rather than agents; the threshold rule supersedes that judgement and both are now included.
• The count is sensitive to three sources whose publishers sell adjacent services (Superkind, SuperDupr and Drag). Removing all three would drop Decagon and Ada out of the top eight and raise NiCE Cognigy and Gorgias into it. They are counted here, and the sensitivity is disclosed rather than resolved.
• G2's own category page returned an error to automated retrieval, so G2's editorial article reproducing its Summer 2026 Grid list stands in for it. Gartner Peer Insights and TrustRadius were read through a public text-extraction proxy after returning errors directly.
• Capterra's category page was captured for its first page of 25 products out of 12 pages; TrustRadius for its first 25 of 127. Both counts are therefore partial and favour products ranked highly on those sites.
• Forrester's own Wave report is paywalled. The publicly readable trade summary of it is counted in its place, and Forrester's vendor-free blog posts were excluded.
• Vendor-published articles were capped at three, chosen as the three framing their list as AI agents for customer service and naming the most distinct products. Counting the other vendor-published articles found (HubSpot, Manus, Fin.ai, Retell AI) would not change which products reach the top of the table.
• Product naming was normalised before counting: Fin, Intercom Fin and Fin by Intercom count as one product; Zendesk AI, Zendesk AI agents and Zendesk for Customer Service count as one; Cognigy and NiCE Cognigy count as one and are kept separate from NiCE CXone; Tidio and Lyro count as one.
• Two vendors in the catalog changed hands or names during the count window. Intercom renamed its corporate entity to Fin in 2026 and Salesforce signed an agreement to acquire it on 15 June 2026, which was not closed as of the count date. Zendesk completed its acquisition of Forethought in March 2026, so Forethought's 5 mentions arguably belong to Zendesk; they are counted separately here because the sources name them separately.

How sources are classified

Every claim is classified by its source type and freshness, and no claim is presented as stronger than its best source. Conclusions read first; one click reveals the type, the link, and the date checked.

Vendor-reported

Stated by the vendor in its own documentation, website, or case studies. Not verified by us.

Independent report

Reported by a party with no commercial tie to the vendor: a customer, an analyst, or a journalist.

Editorial assessment

Our reading of the public evidence, with the basis stated. An opinion, and labeled as one.

Unknown

We could not find a public source that settles this. Shown as unknown, never as a zero or a score.

In a comparison, a cell where the two agents’ best sources do not support a like-for-like statement is marked unknown or not comparable. Missing data never looks like a zero.

How products are positioned

The product positioning map groups documented offerings by their intended work and surrounding tools. Every placement is an editorial assessment with a cited source and a check date. It does not measure tested performance, autonomy, deployment effort, cost, or revenue.

Product focus, from left to right

Support operations

The reviewed offering centers on resolving customer requests and running support workflows.

Service + sales

The reviewed offering explicitly covers support and selling, such as shopping guidance or upselling.

Sales-led

The reviewed offering centers on acquiring customers or completing purchases.

Product breadth, from bottom to top

Focused agent

A targeted agent product that connects to other tools to complete its role.

Connected suite

An agent with connected tools for building, testing, monitoring, and managing its work.

Broad platform

An agent offering embedded in a broader native help desk or CRM workspace for human teams.

We assess the offering covered by the review, including its native deployment when stated, rather than every product sold by its parent company. Service + sales requires explicit documentation of both. Broader scope is not inherently better; a focused product may be the right fit. Categories are not numerical scores, and products within a cell are alphabetical. Missing evidence produces an unknown placement, never an invented point.

The map covers our current catalog, not the whole market. Empty categories do not imply that no such products exist. Factual evidence that changes a placement can be submitted through the correction route.

How fit verdicts are derived

Each entry on a profile is labeled by what the cited sources establish, using one rule applied the same way to every agent, across the use cases billing, orders, returns, product questions, account issues, complaints, human handoff, and across business sizes (small businesses, mid-market, enterprise). It is not a judgement of product quality, and it is never a score.

Vendor-documented

The vendor documents this workflow or audience. This does not establish success in your setup.

External report

An external source discusses this use case. Its scope and attribution matter; it is not our test result.

Caveat documented

A cited source describes a limitation or failure in a specific context. Read the conditions alongside any capability evidence.

Not established

The reviewed sources do not establish this workflow or audience. This does not mean the product cannot support it.

The verdict is computed from the cited evidence, so it is repeatable and contestable through the correction route rather than an unexplained opinion.

Shortlists by business size and industry, and why they are grouped rather than ranked

Buyers arrive asking which AI customer service agent is best, usually for a business their size or an industry they work in. The honest answer is that nobody here has tested them, so this portal will not name a winner. What it publishes instead is a shortlist per segment: the whole catalog, grouped by what the cited sources establish about that segment, with the ordering rule printed beside the result.

1. Agents are grouped by what the cited sources establish for this segment, not scored against each other.
2. An outside source reporting the product in this segment comes first, then the vendor's own documentation, then a documented limitation, then silence.
3. Within a group the order is alphabetical, because the evidence does not support ranking one against another.
4. No agent is dropped: the ones with nothing published are named too.

Every term in that rule counts sources. None of them is a judgement about a product, and no record stores a position: the order is recomputed from the claims on every build, which is why a shortlist changes when the evidence changes and never because somebody moved a row.

Groups rather than places, and the reason is a mistake this page used to describe differently. An earlier version ordered every agent in a segment by who had published about it, on the argument that a documented limitation should never count against a product. Applied to small businesses, that put an agent first on the strength of a Trustpilot reviewer saying they keep its AI switched off because the per-conversation price would run to a four-figure bill. The finding is real and it is still published — as a caveat, in the group for documented limitations, where a reader looking for evidence of fit will read it correctly.

So a caveat sits below vendor documentation, because it is not evidence of fit, and above the agents nobody has written about, because somebody at least looked. Within a group the order is alphabetical: the evidence supports saying which group an agent belongs in, and does not support ranking twenty-two products against one another. A group describes how much has been established, never how well a product performed.

Why a review-site rating does not order a shortlist

The obvious way to separate two agents in the same group is the star rating their listing carries. We cite those ratings where a record has one, with the count and the date, and we do not order anything by them. A marketplace rating scores the product you install, and the products being compared are not the same shape: one listing is a chat widget with an AI agent bundled into it, another is a full help desk that happens to include one. Neither number measures the AI agent, which is the thing this portal is about.

Volume and venue move them too. The same vendor can hold 4.8 on a marketplace its customers are prompted to review and 3.7 on a complaints site, from populations with different reasons to write, which is why a record showing both states both. Ordering on either would publish a ranking whose real subject is where a vendor's customers were asked, not how well its agent handles a refund.

How an industry gets onto a profile

An industry entry uses the same four states and the same rubric as a workflow, and it answers a narrower question: where do the cited sources place this product, not how well did it work there. A vendor page listing an industry it markets to is Vendor-documented. A named deployment, customer or reviewer in that industry, reported by somebody outside the vendor, is an External report. A documented limitation specific to the industry is a Caveat.

Industry entries do not carry their own sources. They cite a claim already published elsewhere in the same record, by its claim id, so a source is stored once and the profile, the shortlist, and the machine-readable record cannot end up quoting different versions of it. An industry with no entry is Not established, which is a gap in the public record and not a finding against the product. Most cells are empty, and a well-established product that nobody has written a sector story about will look no different from one that has never been deployed.

A limitation we know about

The states above describe what the reviewed sources establish, not how well a product performs. That distinction removes a problem an earlier version of this page had to warn about at length: when the site graded products good or poor, an agent whose public record happened to be angry customer reviews scored worse than one whose record happened to be trade press, even if the two were equally capable.

The underlying asymmetry has not gone away, it has only stopped being dressed as a verdict. Some products are covered mainly by their own documentation, some by analysts describing what a product is designed to do, some by end users writing after something went wrong. Those sources answer different questions. So a row reading Vendor-documented where another reads External report tells you what kind of source exists, and nothing about which product is better.

Read the distribution the same way. Of the ten workflows we track for each product, most cells across the catalog are Not established. That is the state of the public record rather than a gap in our reading, and a product named by three sources will always look emptier than one named by eleven. Every profile and every comparison names the kinds of source behind its states so you can see when two products are not documented to the same depth.

A benchmark is what would fix it. Running the same scenarios against every product is the only way to get evidence that is symmetric by construction, and it is the next thing this portal intends to build. Until then, read Caveat documented as "a cited source described a limitation in some context", not as "this product is worse than one without a caveat" — the product without a caveat may simply be the one nobody has written about.

How to request a correction

Use the correction link on any page. It opens a pre-addressed email tocorrections@ai-agent.review. We reply within 14 days.

• A correction concerns a factual claim and comes with a link to a public source that shows it.
• Requests about wording, tone, or positioning are declined. Which claims to publish is our decision.
• Every accepted change appears with its source in the change history at the bottom of the affected page.

Who owns and funds this

Independently owned and self-funded. No vendor relationships, sponsorships, or paid placements. No vendor can pay for a profile, a placement, or a badge, and no vendor is consulted on what is published.

No personal data is collected from visitors beyond what they choose to put in a correction email. There are no accounts and no tracking of individuals.