The AI FinTech Index Brief

The ten days in one line

The vendors built the disagree button. Built, WorkFusion and Kore.ai shipped the record of what happens when a person overrules the agent, Persona and Spektr made the audit trail say who decided and on what number, and SEON wrote the terms sheet for its connector before anyone asked for one. Veriff stopped asking for the document and started listening for a coach in the room, seven vendors published results with the method attached, and one meeting notes vendor closed its doors in the same week a rival shipped two integrations.

This issue covers October 11 to October 20. The log recorded 76 entries across 76 vendors, 70 Verified against the vendor’s own materials and six Partially Verified from third party reporting. Product changes led with 38, integration work followed with 12, seven entries were published performance evidence, six changed security or privacy terms, five changed a model or the architecture around it, three changed how a person oversees an agent, two were product sunsets, two were pricing or packaging, and one was a license. Funding rounds, valuations, and awards are not logged, here or anywhere on this index.

One change of address for this issue. It is written for the people who build and position products in this category rather than the people who buy them, so the note at the end of each section says what belongs on a roadmap instead of what to ask a vendor.

The disagree button got a form

Two issues ago the approve button went to work. Last issue it barely counted as news. This period the vendors built the other button, the one a person presses when the agent is wrong, and gave it a form to fill in.

Built’s CLA Draw Agent now lets a draw processor or approver override a blocking finding by choosing a decision and a reason, with an optional comment. The original finding stays visible beside the override rather than turning green, and each override carries who made it, when and why. Policies the agent judged not to apply to a draw are listed with its reasoning, so a reviewer can see they were weighed rather than skipped.

That is a user interface decision with a compliance consequence. A finding that turns green when overruled tells an examiner the loan passed. A finding that stays red with a signature beside it tells the examiner a person decided to proceed anyway, which is the true story and the one the lender wants on file.

WorkFusion made the no stick. Tara 4.1.1, its payment sanctions screening agent, now suspends a learned decision pattern the moment an analyst escalates a matching hit, instead of only declining to clear hits while escalations continue. The pattern resumes after a set number of consecutive analyst releases, and new report columns trace each reapplied decision back to the historical alerts behind it. A human disagreement now changes the agent’s habits, not only today’s alert.

Kore.ai shipped the dial. Agent Platform 1.8.0 gives Autoloop, the loop that proposes improvements to an agent from its own evaluations, three modes: Autopilot applies changes on its own, Copilot applies them after a person approves, and Advisor only suggests. The mode can be set at the start of a loop or changed mid run. Autopilot being on the menu is the news for a governance team. Which mode is the default is the news for a product manager.

Persona and Spektr worked on the record itself. Persona’s new Review Matches endpoint writes screening review decisions back into the report history, and its webhooks now carry a reviewer token that says whether a person, a Workflow run or an API key made the call. Three possible answers to who decided, which is two more than most audit trails offer. When a Spektr monitoring rule fires, the case now stores the measured value behind each condition with its threshold and a Met or Not met badge, and the AI case summary quotes those stored numbers, so the narrative cannot cite a figure the rule did not fire on. The story is handcuffed to the number, which is where a SAR narrative wants to be.

The drafting agents kept arriving with the review step attached. Hawk’s Regulatory Reporting Agent turns a closed case into a draft suspicious activity report inside Case Manager, keeps the four eyes workflow, tracks the filing deadline, and checks every field and narrative sentence against the case passage it came from, leaving out what it cannot verify. Hawk says the design eliminates hallucinations, which is Hawk’s claim to test. FIS made its AML Investigations Agent, built with Anthropic and run on Claude models in FIS’s private cloud, available to any US bank including banks that are not FIS clients. No case closes without explicit analyst confirmation, and native integration covers the Horizon and Systematics cores today, with IBS next.

Vertafore took its Policy Intake and Policy Check agents to general availability in AMS360, writing to the agency management system only after a person approves each change and keeping the review history. nCino’s Mortgage Advisor for Loan Officers creates tasks and sends reminders once the loan officer says yes. WTW’s RiskAgility assistant can plan, build, test and document a life insurance model migration for actuarial review. Aurionpro’s Integro LOS 2.0 generates lending guardrails from an uploaded credit policy and routes exceptions through a deviation flow with an audit trail. Rulebase’s new routines run on a schedule and log every run as Acted or Quiet with a version number, and edits stay in draft until someone publishes.

Our read

The approve button is table stakes now. Three things separate the entries in this section, and all three are product decisions rather than model decisions: what the override looks like on screen, whether a human no changes the agent or only overrules it this once, and who sets how much autonomy the agent gets. Built’s red finding, WorkFusion’s suspended pattern and Kore.ai’s three modes are each one of those answers made concrete. None of them is a feature a demo audience will applaud. Every one of them is a feature an examiner will ask to see.

For the roadmap

If your agent can block, hold or decline anything, the override is a feature and not an edge case. Ship it with a reason code, a timestamp, a name and a view where the original finding stays visible. Then decide whether a human no retrains the agent or only overrules it once, and say which in the release notes, because WorkFusion just turned that into a comparison point.

The connector came with a terms sheet

Three issues ago the data vendors became connectors inside the frontier assistants. This period the connectors grew paperwork, and in two cases the paperwork was the product.

SEON published data handling and security terms for its MCP connector, the one that lets an AI assistant query a SEON account. SEON keeps no copy of data passing through, does not train on it, and treats the connected AI provider as the customer’s own recipient rather than a SEON sub processor. Each connection signs in as a named user over OAuth 2.1 and inherits that user’s permissions, every AI query lands in the Activity Log, and controls to withhold specific data points are not yet available. Recorded Partially Verified. That last clause is a roadmap item wearing a disclosure, and it is the most honest sentence in this issue.

Oscilar made its MCP Server generally available, so Claude, Claude Code or any MCP compatible assistant can take a written credit or fraud policy, turn it into a draft workflow, test it and work cases. The assistant runs under the user’s existing roles, permissions and audit trail, and any workflow change it makes is saved as a draft that a person on the team deploys. The assistant drafts. A human ships. Both the server and Oscilar AI, the in product assistant that explains decisions rule by rule, are included at no additional cost.

Sumsub sold the filter rather than the data. Its KYA Module for MCP sits between a business’s own AI agents and the sources they connect to, screens each request, strips personal and other sensitive data before it reaches the agent, keeps an agent activity log, and can demand a live human verification when an agent tries a flagged action. It works first with Amazon Bedrock AgentCore Gateway. Fiddler’s release 26.23 does the same job one layer down for Portkey traffic, redacting identifiers and secrets inline and blocking the call if redaction fails.

Proof sold the signature. x401 is an open protocol, now in production, that lets a website ask an AI agent to prove who it is acting for and what that person approved. An agent identity wallet holds the person’s verified legal identity and lets them approve, on their own device, exactly what the agent may do, such as authorizing a wire or signing a payment mandate. It works with Meta’s Muse, ChatGPT, Grok, Claude and other MCP capable assistants. Forter sold the bouncer. Its Agentic Commerce Engine decides on scraper and AI agent traffic in real time inside Fraud Management, separating shopping assistants from abuse bots, and an AI Activity Dashboard shows a merchant which agents are buying before the merchant writes a policy for them.

Put Sumsub and Forter side by side. One is a bouncer for your own agents. The other is a bouncer for everyone else’s. Both shipped in the same five days, which tells you the category now expects agents on both sides of the door.

The smaller connector entries show the sequence. FinTech Studios took its MCP server from read only in September to version 1.1.0 with four write tools, and the package notes say the AI client confirms each write before it runs. Ravelin shipped a read only documentation MCP server with no authentication, so a coding assistant can validate an API payload without a browser tab. RavenPack put an official Bigdata.com connector inside Microsoft 365 Copilot Chat and Copilot Cowork with a source on every answer. Hebbia connected to every Octus product a firm subscribes to through one MCP connector, S&P Global put Visible Alpha’s broker consensus into its AI Data Portal for MCP applications, and BlueFlame now runs inside Datasite’s Grata Intelligence workspace, where each user’s existing data room permissions limit what the agent can retrieve. Bloomberg’s ASKB went generally available and can now answer over three months of a client’s own Instant Bloomberg chats, with each answer cited to the chat or file it came from.

Our read

Ten of this period’s 76 entries involve MCP or agent access, so the connector is no longer a launch. It is the appendix with the terms in it. What separates the entries is what the connector may do, in order: read, write behind a confirmation, draft for a human to deploy. Then whose credentials it carries, and then what the paperwork says about retention, training and who the AI provider is to the customer. SEON wrote that page before procurement asked for it, which is the point: in this market the terms sheet is a feature, and the vendor who writes it first gets to set what the next vendor is compared against.

For the roadmap

Ship the connector read only, then add writes behind a confirmation, then let it draft what a person deploys. Write the terms at step one: retention, training, logging, and whether the customer’s chosen AI provider is your sub processor or theirs. If the answer to any of those is not yet, say so in the document the way SEON did, because the alternative is a security questionnaire finding it for you.

The identity check changed the subject

Last issue the document lost the benefit of the doubt. This period the identity vendors stopped arguing with the document and asked a different question.

Veriff published Presence Verification, which uses a passive selfie video and device signals from its native SDKs to confirm that a real, live, single person is completing the session, with no identity document requested. A higher tier adds Coercion Detection, which analyzes the session’s audio and video for signs that a third party is guiding the user. It needs microphone access and is limited to adults. No other entry this period checks whether the customer has company, and in money mule and push payment fraud the company is usually the problem.

Incode put TrustGraph into production as a module for Flows and Workflows. It checks whether the identity data in a session has appeared before, in the customer’s own history or across Incode’s network, returns the signal during the session, and links connected sessions in a Fraud Ring Explorer. Incognia launched Real Time Scam Prevention for banks, which flags a likely scam payment before money moves by looking for evidence of a real world relationship between sender and recipient, drawn from apartment level location intelligence and device recognition, pooled across participating banks without any of them seeing another’s location data. Incognia reports catching about 70 percent of scam transactions at an unnamed bank with more than 80 million users, with no method published.

None of those three questions is answered by a better PDF detector. Is the person alone. Has this identity been seen before. Does the sender know the recipient. The document checks did keep shipping underneath: GBG took its Biometric Onboarding module to production with face match, liveness and an injection attack variant that uses an encrypted selfie, and HyperVerge’s HyperKYC 3.5.0 added injection checks to document capture rather than only the selfie, along with fewer false rejections on Android. They read as the entry fee now.

Two deprecations pointed the same direction. Oz Forensics’ SDK 10.4.0 deprecated on device and hybrid liveness on iOS and Android and set their removal for 1 January 2027, telling customers to route captures to a server. IDVerse 1.7.0 set data retention to one year by default and three years at most, with anything longer available only on request after legal review. Verification is moving off the handset and out of the archive in the same month.

On the business side, Trulioo launched Business Risk Horizon, which combines its KYB data with agentic research into a company’s operations, traces ownership through layers of entities to find sanctioned owners, and attaches a plain English explanation and the evidence to every score. Shufti expanded its Bank Account Verification API from 48 to 109 countries and added an IBAN check that validates structure and check digits, returns the bank behind it, and says in the response whether the result confirms ownership or only the IBAN. Middesk removed the per identifier scores from Business Risk responses, so anything parsing them breaks, and iDenfy let card verification run inside its mobile SDKs with card images going straight to its PCI environment rather than through the client app.

Our read

Document fraud detection is becoming a feature of the older product. The newer product asks a question about the session, the network or the relationship, and sells the answer as a signal the bank’s own rules consume. That is a positioning change as much as an engineering one. A vendor whose pitch still leads with tampering detection is describing what every entry in this section already does. Shufti’s small change is worth a second look for the same reason: a response that says what it did not confirm is a product feature, and a cheap one.

For the roadmap

Add one signal that does not need the document, whether presence, prior appearance or relationship, and name it in the release note as such. Pick a retention default you can say out loud in a sales call. And if your SDK still runs the model on the handset, read the Oz Forensics entry twice, because the server is winning that argument and the deprecation notice is the kind of release note a competitor reads first.

The case study grew a methods section

Seven entries this period were published performance evidence, against one in the previous two weeks. The count is the smaller story. The shape of the evidence changed.

Narmi published its first measured results for AI DecisionAssist, the account opening review agent that recommends a decision for staff to approve or override. The comparison was applications reviewed with and without the tool at six institutions over their first six weeks live: median time to decision fell from about 8.5 hours to 4.4, hands on review time fell from 8 to 12 minutes to under 2, and 90 percent of decisions at all six now finish within three days. The institutions are unnamed and Narmi did the measuring, and the method is stated, which is the part most case studies leave out.

Primer tested its own filing system and reported the cost of the test to the cent. Its retrieval evaluation ran company root, the file layout it ships with its research agent, against three other layouts and two tool setups over 480 attempts by GPT-6 Sol. With enhanced search, company root surfaced every required piece of evidence in 60 of 60 attempts at about 9.3 US cents and 41 seconds each, and complete retrieval across all layouts rose from 85.4 to 97.1 percent. Primer then wrote the limit itself: the result shows the evidence reached the agent, not that its sixty answers were right.

Caudena priced an investigation at 32 cents and then said what the price does not buy. Its benchmark ran 11 model configurations through agent workflows on its Prism tracing tools across Bitcoin, Ethereum and TRON cases, with a 25 minute and $3 cap per attempt, scored on graph recall, graph precision and required report findings over 69 attempts. Gemini 3.8 Flash at medium reasoning scored 100 at an average of 3 minutes 40 seconds and about $0.32 per investigation, tied on quality with Claude Opus 5.5 and GLM-5.3. Caudena notes there is no comparison with human investigators and that a perfect score does not prove every narrative claim was supported.

The other four were conventional and still useful. Facctum processed all 1,653 listings in the EU’s 22nd Russia sanctions package, adopted on 12 October across three regulations, in 4 hours and 53 minutes, without saying when the clock started. Bdeo published first production results from S&G Response in UK motor claims: 69 percent of claims processed within 24 hours through self service, 98 percent routed automatically to the right repair pathway, and a savings estimate of £250 to £375 per claim that is an estimate. Temenos reported 16,385 transactions per second in its own composable core benchmark on Azure, recorded Partially Verified, and Sympera relayed Republic Bank of Chicago’s 30 percent rise in new business relationships across 20 branches, the bank’s own figure with no baseline.

The model moved too, and this time the vendors said so in the release notes. Boost.ai’s platform version 14 moves its core classification features onto Presis, its own native model, replacing the GPT-4o and GPT-4.1 models those features used before. Orchestration, out of scope detection, the hallucination guardrail and AI review switch over on update, and customers can no longer pick a different model for them. Boost.ai cites lower token cost and latency. For a bank’s model inventory it is a supply chain change dressed as an upgrade, and the honest release note is what makes it manageable.

SBS launched SBS AI Foundation for core banking, which runs on a private cloud, a managed cloud or fully on premises so data never leaves the bank, on NVIDIA’s stack with NIM microservices, NeMo Guardrails and open Nemotron models, and is live at Scottish Building Society. Kount expanded its Email Verification Service model to take in login, account opening and payment events, so First Seen Date moves earlier for many emails, and Kount advises customers using those attributes in outside logic to review it. Darwinium’s 15.3 reweighted the inputs behind device signatures after finding that iOS version numbers had been breaking the link every time a user upgraded. AlternativeSoft published a LASSO factor model with all 25 factors listed, recorded Partially Verified, and Axyon put SHAP attribution on every daily signal so a manager can see what pushed a stock up the ranking.

Our read

The old proof point was a percentage and a logo. The new one is a method, a cost, a time, and a sentence about what the result does not show. Primer and Caudena wrote that sentence themselves, which used to be the analyst’s job, and a case study without one now reads like the older product. The model entries make the same point from the other side: Boost.ai, Kount and Darwinium each changed something underneath and described the change before the customer found it, which is cheaper than the alternative.

For the roadmap

Publish the method with the number: the baseline, the sample, the window, the cost per run and the limit. If the number is the customer’s own, say so. If you change the model underneath, put it in the release notes before the upgrade, with what the customer loses as well as what they gain. Boost.ai’s note says bring your own model is gone for those features. That sentence will cost a few deals and save a great many renewal conversations.

Market notes

Mortgage had the week’s largest quiet change. ICE Mortgage Technology’s Encompass 26.2 October Service Pack supports multiple credit score models across Encompass, Consumer Connect, TPO Connect and ICE PPE, so a lender can process a report carrying FICO Classic and VantageScore 4.0, price both at once, and have the scores and the borrower disclosure follow whichever model the rate lock chose, with FICO 10T ready to switch on once approved. Fannie Mae and Freddie Mac have accepted VantageScore 4.0 on conventional loans since 9 September, and this is the origination system catching up, with configuration and business rule work ICE documents separately. Ocrolus gave every mortgage customer its Detect document fraud product at no added cost, with an authenticity score and reason codes on each bank statement, pay stub and tax form, which changes the cost comparison against standalone fraud tools and is a packaging decision worth studying. nCino bundled six AI tools into a Loan Officer Productivity Suite with early adopters named and general availability not stated. Mortgage Capital Trading connected its secondary marketing platform to Dark Matter’s Empower, writing loan sale commitments back automatically, and the 3 minutes per loan in the headline is MCT’s estimate. Aloan published a public reference for its Commercial API with an OpenAPI document, 30 operations and a sandbox, recorded Partially Verified, and DigiFi added AppOne to its indirect lending aggregators beside Dealertrack.

The notetaker lane had a death and two births in the same week. Pulse360 replaced its homepage with a notice that it has shut down, apologizing to the advisors who used it, its investors and its team, with its product and pricing pages now returning errors and no word on export, retention or deletion of stored meeting notes. Recorded Partially Verified, the date taken from the page’s last modified record. In the same five days Jump released an FMG Suite integration that surfaces compliance approved content from what was discussed in a meeting, for the advisor to review before sending, and GReminders added a connector to SS&C Black Diamond that pulls households, holdings and performance into its meeting briefs, which also means portfolio data now flows into GReminders. A shutdown notice is a product release too, and this one shipped without the export feature.

Elsewhere in wealth and research, Dynamic Planner connected to Plannr so client data and risk results flow both ways, and Shortcut 1.4.2 let its Excel agent run on a schedule and write reported actuals back into the model without anyone starting the run, which raises the question of what it changes on its own and how that is logged. Wealth Dynamix published a fair use policy for CLMi with the limits in numbers: 6,000 API calls per sliding five minutes, 50 concurrent, 1 TB of ingested storage, and 35 monthly active portal users per authorized user, recorded Partially Verified. Carta received CSSF authorization as a Specialised PSF, adding Luxembourg fund administration and transfer agency to the same platform it runs elsewhere, and MyComplianceOffice extended personal trading controls to employee activity on prediction markets and widened communications surveillance to Telegram and Slack.

In financial crime, Abrigo launched Customer Hub, which puts customer, account, transaction, alert, case and risk information in one investigation workspace so AML and fraud teams work from the same record, for hosted customers only and rolling out in waves. Unit21 shipped a Corelation KeyStone integration for credit unions that syncs member and transaction data about every five minutes and lets an analyst block a card in the core from inside a case. TRM Labs launched Ardence, an AI investigation platform drawing on more than 60 partner data providers, with investigators directing the work and every finding linked to its source. Bits Procedures can now pause for an outside provider and resume on their own, with a Waiting status, and rolling monitoring windows go down to the minute. Bretton changed its related entity check to return no result when a screening provider has merged two similarly named organizations, rather than a false mismatch, which means a new outcome for rule logic to handle. eGain joined the Jack Henry integration network and grounded its agents in more than 100 named regulators and industry sources.

The plumbing entries were small and each one closes a security questionnaire question. StandardC added single sign on through OpenID Connect with Entra ID, Okta, Google Workspace and Auth0, and once every user has moved an institution can ask StandardC to retire passwords entirely. SS&C’s WorkHQ now routes AI Gateway traffic to its UK production deployment through a UK load balancer instead of a US one, and deletes form attachments after a fixed 90 days. Guidewire Cloud added cross region AWS PrivateLink, with up to three weeks for a change to reach each environment. ibl.ai gave every agent its own privacy policy setting, so an intake agent can be forced to keep no history while a research agent keeps it, and one endpoint reports which tier decided and whether it is locked. Charm published a privacy policy for its browser extension, which reads case and account screens inside bank tools on an allowlist the customer sets, and says plainly that filtering identifiers out of AI inputs is a best efforts commitment rather than a guarantee.

In insurance, WTW’s RiskAgility Financial Modeler gained a semi autonomous assistant that can take an unfamiliar migration specification through planning, building, testing and documentation for actuarial review, with no time savings claimed. Vertafore’s Policy Intake and Policy Check went generally available in AMS360 with the approval step covered above. Finastra launched StandbyX so regional US banks can automate standby letters of credit without its full trade finance platform, with no AI capability described, so it widens coverage rather than depth. Eltropy made Branded Texting generally available on RCS, giving credit unions a carrier verified name and logo on member texts, with the protection lost on the SMS fallback. Wink’s face and palm checkout went into PAX’s PxRetailer, ready on every camera enabled PAX terminal already in the field.

What the ten days say about the space

Seventy six entries, and the common thread is that the paperwork became the product. The override came with a reason code. The connector came with a terms sheet. The verification came with a retention default and a line saying what it did not confirm. The case study came with a method and a limitation the vendor wrote itself.

For the people who build and market in this category, that is uncomfortable and useful in the same breath. Uncomfortable because the caveat is now part of the claim and a release note is now a compliance document. Useful because the differentiators got more concrete than they have been in a while: the override record, the permission model, the signal that does not need the document, the method behind the number. None of those demo well. All of them survive a procurement review, and in this market the procurement review is the demo that counts.

The last issue ended with financial services building its agent infrastructure the way it built payments, every message checked against a rule and every rule written down. This period the vendors wrote down what happens when a person disagrees with the rule. That is the harder document, and the one that will be read.

Index Answer

Which financial services AI vendors publish results a reader can check?

Far more than publish a price, and fewer than the launch posts suggest. Of the 610 vendors the AI FinTech Index has graded on Operational and Outcome Evidence, 231 reach the top grade, which means a named customer, hard performance figures and enough method to test them. Another 275 publish real figures without a name attached, whether aggregates across a customer base or audited scale disclosures from a listed company. The remaining 104 publish logos, unnamed case studies or claims without a number, and not one vendor in the index publishes nothing at all.

Set that against the same 610 vendors on Commercial Transparency, where 37 reach the top grade and 501 sit in the bottom band. The field documents its outcomes far more readily than it documents its prices, for the same reason it names the core it plugs into: a case study gets the meeting and a price list is thought to lose it.

The top band is where this period’s entries push on the bar. The grade stops at enough method to test the figure. Narmi stated the comparison, Primer and Caudena published a cost per attempt and a sentence on what the result does not show, and Facctum published a time without saying when the clock started. Three of the seven evidence entries this period were written the way a validator would want to read them, which is a higher share than the band itself requires.

The practical move for a product team is to read its own case study the way a model risk reviewer would. Is there a baseline, a sample, a window, a cost and a limit. A named logo with a percentage reaches the band this index can grade. The vendors ahead of the band this period are the ones who wrote the paragraph an analyst used to write for them, and they will set what the next case study is compared against.

The full grading method and what separates each band are on the capability framework page.