Eltropy

Applying the removal test leaves a working product: a unified omnichannel conversations platform carrying text, voice, video, chat and co-browsing into more than 50 core and ancillary systems, which is what 750 institutions originally bought and what the company sold for years before the agentic line existed.

The artificial intelligence layer is real, shipped and growing quickly, spanning chat and voice agents, an employee assistant and an intelligence layer, and the 2026 agentic platform and marketplace suggest the centre of gravity is moving. It has not moved yet.

This is the same structural position as other channel platforms in this index and the useful distinction for the category is that the assistant native vendors, where the model is the entire product, sit a full grade above the channel platforms that added models to a communications spine. Revisit if the agents become the reason institutions buy.

The most complete oversight description located in the conversational banking agents lane. Four mechanisms are named rather than asserted. Institutions get visibility into what an agent did, why it did it, what data it used and how it reached its decision, which is provenance and reasoning at the level of the individual action rather than a general explainability claim.

Role based control determines which employees can access and direct which agents, so agent authority is scoped to a person rather than granted to the institution as a whole. Agents are created and deployed inside stated operational and compliance guardrails under a named framework. And third party agents entering the marketplace must pass certification before they can be distributed, which places a gate between an outside developer and a member conversation.

The chief executive frames the design as innovative but controlled, powerful but predictable, open but always safe, and unusually for that kind of sentence the product description supports each half of it.

Governance is treated as product rather than policy, which is what earns this grade. The audit position is specific: an institution can see what an agent did, why, on what data and by what path to its decision, and that is the raw material a model risk function needs to review an automated interaction after the fact. Third party agents are certified before distribution, so a review step exists between an external model and a member.

Held off the top grade because none of it is evidence about model behaviour. No accuracy figure, containment rate, error rate, escalation rate or validation summary is published, no artificial intelligence management system certification is held, and the framework itself is described in announcements rather than in a document an institution could put in a vendor file.

The installed base is the evidence and it is substantial and checkable: more than 750 community financial institutions, more than 100 added during 2025, and a stated 43 percent of new customers buying multi channel plus artificial intelligence platform deals rather than single point products.

Named customers exist with a named executive quoted, the president and chief executive of InTouch Credit Union speaking specifically about the governance framework, and OnPath Federal Credit Union named on a consolidation contract. Four appearances on a national fast growth company list add an external data point. What is missing is outcome.

No resolution rate, containment rate, handle time reduction, cost saving or member satisfaction figure was located for any customer or in aggregate, which for a conversations platform is the obvious measurement and its absence is conspicuous.

A named safety framework, a certification gate on third party agents, and dedicated laboratory environments configured to mirror live institution setups so partners develop and test against realistic configurations rather than in production. That is more structure than most of this category offers. Two questions hold it off the top grade and the second is created by the marketplace itself.

Eltropy sits across member conversations at more than 750 institutions, which is an aggregate view of community banking member intent that no single participant could assemble, and nothing states whether conversation data from one institution informs agents serving another. The marketplace then introduces third party developers into that position, and nothing states what member conversation context a certified partner agent can see or retain.

No data processing agreement, subprocessor list, retention schedule or lawful basis statement was located publicly. The material at stake is unusually rich: member conversations across text, voice, video and chat at more than 750 institutions, which includes voice recordings and the substance of discussions about loans, hardship, disputes and account problems. Federal financial privacy rules apply directly to that content.

The marketplace sharpens the question rather than answering it, since certified third party developers gain a position adjacent to those conversations and nothing states what they may see, store or use.

No certification, attestation, trust centre or penetration testing statement was located in this pass. The early access programme states that partner fintechs receive compliance and security documentation, which indicates such documentation exists but describes a partner facing artifact rather than a published one.

Credit union vendor due diligence expectations make it very likely that a service organisation control report is held, so the finding here is about disclosure rather than about controls, which is precisely what this axis measures. Worth confirming on a later pass.

Eltropy is a technology supplier to credit unions and community banks and holds no financial licence, which is the correct posture for its shape and carries no penalty. Its positioning leans heavily on regulatory sensitivity, describing credit unions as operating in one of the most regulated and trust sensitive environments in financial services, and its buyers are subject to supervisory expectations on third party risk that a vendor in this position is routinely assessed against. None of that is formal standing. No supervised regulator test, sandbox participation or programme admission was located, and no statement of regulatory position is published.

No fairness testing, differential outcome monitoring or impact assessment was located. The exposure is real and specific because of where the agents operate. Eltropy names collections and loss mitigation as priority workflows, and an agent conducting a payment plan conversation with a member in financial difficulty is operating inside consumer protection territory covering unfair practices and debt collection conduct, with the member on the other side of it.

Lending and member financial wellness are named alongside it. The credit union carries the regulatory duty rather than the platform, which is why this is not graded lower, but nothing public describes how agent behaviour in those conversations is tested for consistency across members or monitored for outcomes that fall unevenly.

No error rate, remediation commitment, liability position or correction path is published. The marketplace creates the sharper version of the question, and it is the second time in this session that an agent marketplace has raised it without answering it: when a certified third party agent gives a member wrong information about a loan, a payment plan or a dispute, responsibility sits somewhere between the fintech that built the agent, Eltropy which certified and distributed it, and the credit union whose name the member sees. Certification implies Eltropy vouches for something, and nothing states what it covers or what happens when a certified agent is wrong.

No model provider, family, version or country of processing is disclosed for the chat agents, voice agents, assistant or intelligence layer. The marketplace makes the supply chain structurally more complex rather than clearer, because certified third party agents will run their own models inside the platform, and nothing states whether a partner must disclose what it uses, whether the institution is told, or whether certification examines the question at all. An institution deploying a marketplace agent would inherit a model supply chain it cannot see.

More than 50 native integrations span core banking, lending, collections, account opening, customer relationship management and contact centre systems, which for a community financial institution is the whole battle: a conversation platform that cannot see the core is a separate inbox rather than a service channel. The company describes its position as sitting where member conversations happen and connecting the intent expressed there directly to execution in the systems of record.

The marketplace architecture extends this by making the platform a meeting point for three parties, the institution defining workflows, the fintech building agents and the core provider holding the data, which is an integration strategy rather than a feature list.

Delivered as a hosted platform and no region list, residency commitment, single tenant option or self hosted path was located. Exposure is narrower than for a global vendor because the footprint is North America only, so domestic processing can reasonably be inferred, but inference is not disclosure.

The laboratory environments offered to marketplace partners, described as configured to reflect live institution setups, raise a related question nothing addresses: what data those environments contain and whether it derives from real member interactions.

No pricing, tier structure, billing basis or indicative range is published for the platform, and the route to a number is a demo request. This is the index norm and is measured against Sumsub, which publishes per verification rates on a public page.

One partial exception is worth recording because it is unusual: the marketplace programme discloses the shape of its commercial model, stating that early partner pricing and revenue sharing terms are available to fintechs joining the early access programme. The shape is disclosed and the numbers are not.

More than 750 credit unions and community banks across North America, with more than 100 added in 2025, which is the largest installed base in this category by a wide margin. Functional coverage runs across the whole institution rather than one department, spanning marketing, lending, servicing, collections, fraud and dispute handling, business banking and branch operations.

Held at the middle grade because the breadth is functional rather than institutional: the platform serves one tier of one market. There are no large banks, no international footprint outside North America, and no insurance, wealth or capital markets exposure. A large count of similar institutions is depth, not reach.