EnFi

Built model first in 2024 rather than assembled around an existing platform, and the removal test leaves nothing. Agents read borrower leverage, collateral positions and credit histories, flag inconsistencies between documents, screen incoming deals and monitor portfolios after booking, each tuned to the individual institution's book. There is no underlying loan origination system, spreading tool or workflow engine that would survive the models being taken out.

The company describes the offering as agentic human infrastructure rather than software, which is the correct description of a product whose unit of value is analyst capacity rather than a feature set.

Two accounts of what the agents actually do are in circulation and they do not agree, which is itself the finding. Press reporting describes agents that analyse and make decisions on credit applications, while the lead investor frames the thesis as artificial intelligence not replacing human judgment but creating the capacity for humans to exercise that judgment at scale. Both cannot be a complete description of the same product.

The company itself publishes no approval gate, confidence threshold, escalation path, sampling regime or statement of which decisions an agent may take unaided, and agents are stated to complete end to end tasks. For a system operating inside regulated credit approval, the boundary between recommendation and decision is the single most important thing to publish and it is absent.

No accuracy figure, validation evidence, error rate, documentation package or backtesting capability was located. Transparency and governance are named as design emphases in announcement material without a mechanism attached, and the distinction the index draws matters here: describing a system as explainable is not the same as exposing which variables drove a particular assessment, which is what earned Unit21 its grade.

The correctness question is sharp because the agents flag documentation inconsistencies and assess collateral and leverage, so a missed inconsistency is a credit file approved on a defect and a false one is analyst time spent on nothing, and neither rate is published.

One institution is named and it went public voluntarily: Citadel Credit Union disclosed its deployment at the time of the Series A, with an executive quoted on responding to rising demand without increasing the risk profile by extending credit team capacity. Beyond that there is no volume, no application throughput, no cycle time and no outcome figure of any kind. Roughly 30 staff and 22.5 million dollars raised across a seed and a 15 million dollar Series A led by FINTOP.

The investor network is the tempting number and it should not be read as adoption: the syndicate is stated to connect to more than 150 financial institutions, with FINTOP alone reaching around 90 community and regional banks and Patriot Financial Partners having invested in 66 banks. That describes distribution potential, not deployed footprint. Same position as Flagright, which also carries a single named customer.

One scoping property is stated and it is a good one: agents are tuned to each institution's specific portfolio, which implies customer scoped configuration rather than a single model trained across the customer base. What is missing is the boundary itself.

Nothing states whether credit outcomes, default experience or documentation patterns observed at one bank inform models serving another, and that question carries unusual weight in this lane because the syndicate behind the company connects to more than 150 institutions that compete for the same commercial borrowers in overlapping regional markets. Compare Rulebase, which forecloses training on customer data in a single sentence.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The payload is commercial credit files, which is a shorter privacy chain than consumer lending, but it is not a clean one. Small business credit routinely involves personal guarantees, principal credit histories and owner financial statements, so individual consumer credit information enters commercial underwriting through the back door, and nothing published addresses how personal data inside a commercial file is handled differently or at all.

No attestation, certification, trust centre or dedicated security page was located, and no service organisation control report or international information security standard certificate is announced or offered on request. For a two year old company this is unsurprising, and it is also the first item a bank vendor risk programme will demand before agents touch credit files. The syndicate's reach into more than 150 institutions makes it commercially consequential rather than merely procedural, since each of those relationships begins with the same questionnaire.

No supervisor, statute, rule or guidance instrument is named in vendor material. Emphasis on reliability, transparency and governance as requirements for regulated institutions appears in announcement coverage, which describes the market rather than stating a position.

The gap is specific and closeable: a product placing model output inside commercial credit approval at supervised banks and credit unions sits directly against model risk management expectations, and its buyers are examined against them. Bretton AI shows what naming the instrument looks like in this index. EnFi holds no licence and needs none, which is the correct posture for a technology supplier.

Credit decisioning with no fair lending disclosure from the vendor, which is the sixth instance of this pattern after Alloy, Sardine, Taktile, Oscilar and UPTIQ. Commercial rather than consumer lending weakens the protected class analysis but does not remove it: business credit is inside the federal equal credit opportunity regime, small business lending carries its own data collection rule, and deal screening decides which applications advance before a human sees them, which is where systematic exclusion by industry, geography or borrower profile would occur invisibly.

No fair lending testing, adverse action reason code handling, disparate impact analysis or monitoring commitment appears in vendor material. Third party commentary asserts that an explainability first architecture addresses reason code, audit trail and disparate impact requirements, and that claim is not made by the company itself and should not be credited to it.

No guarantee, indemnity or falsifiable commitment on assessment quality was located. What sits above the floor is the stated design emphasis on transparency and governance, which implies an institution can reconstruct what an agent concluded and why, though no audit trail or reasoning artifact is described concretely enough to confirm it.

The party with no route is the commercial borrower, and the shape is particular to deal screening: an application filtered out before a human reviews it produces no decision to appeal, no adverse action to explain and no record the applicant ever sees. Rejection by triage is the quietest failure mode in credit and nothing addresses it.

Nothing is disclosed. No foundation model provider, hosting arrangement, subprocessor, credit bureau, financial data aggregator or document extraction service is named anywhere, despite agents necessarily consuming borrower financial statements, credit histories and collateral documentation from somewhere. Whether those files pass to an external model provider during analysis is unstated, which is the fourth party question a bank examiner asks directly. Announcement coverage notes the importance of data infrastructure quality for reliable deployment in regulated environments without identifying whose infrastructure is involved.

Integration into existing workflows is asserted and nothing is named. No loan origination system, core banking platform, spreading tool, document repository or credit bureau connection appears anywhere, and no developer documentation or application programming interface reference was located.

That matters disproportionately for this buyer, because a community bank or credit union cannot fund a long integration project and needs to know what connecting an agent to its existing origination and document systems actually involves. Contrast interface-ai, which sells to the same segment and names core banking, loan origination, customer relationship and knowledge systems as prebuilt.

No hosting model, cloud provider, region selection, residency commitment or private deployment option was located. The footprint is United States centric so cross border transfer is not the live question, but a supervised institution handing borrower financial statements and credit files to a third party still needs to know where they rest and under what controls, and nothing published answers it.

No pricing, packaging or basis of charge is published and the only indication of the model is a third party description of subscription based services with customised implementations. For a product sold as analyst capacity the unit of charge is the substantive question, since per seat, per deal and per institution pricing produce very different economics for a community bank with a handful of credit staff. Nothing addresses it.

Banks, credit unions and private lenders are all named, and the platform is stated to work across the full commercial credit spectrum and the whole loan lifecycle rather than one step. The segment focus is the strategy and it is a deliberate one: regional and community institutions that cannot fill credit analyst positions, which the chief executive frames as thousands of unfilled roles at any given time.

That is the same underserved buyer thesis that interface-ai runs in the conversational lane. The limits are that coverage is United States centric with no international footprint evidenced, and confined to commercial credit rather than spanning retail or consumer lending.