← All issues

The AI FinTech Index Brief

October 31 to November 13, 2026 · Published November 15, 2026

The two weeks in one line

Know Your Agent went mainstream. LexisNexis Risk Solutions, BioCatch, Callsign, Feedzai and Sardine all shipped ways to spot the AI agents acting inside a customer session, and Forter started scoring orders placed through ChatGPT Wallet. Complead and ThetaRay showed agents working in shadow before they were allowed to act, Q2 had its assistant certified after outside testing, and Herald and Marloo started showing the source behind every answer they fill in.

This issue covers October 31 to November 13, plus twelve late October entries that reached the log after the last issue went out. The log recorded 123 entries across 123 vendors, the most this Brief has carried, so you may want a coffee. All but two were Verified against primary sources, and those two are recorded as Partially Verified. Product changes led with 79 and integration work followed with 16. The other 28 covered published results, security and compliance, oversight, models, pricing and one regulatory approval. Funding rounds, valuations, and awards are not logged, here or anywhere on this index.

The agent got a name tag

Last issue, Know Your Agent was a product name three vendors shared. This period the idea went mainstream, and the fraud stack spent two weeks learning to tell a customer from the customer’s robot.

LexisNexis Risk Solutions added Know Your Agent intelligence to its digital identity platform. It spots AI agents working inside a customer’s online sessions, including personal agents that never announce themselves. The signal flows into decisioning the institution already runs, so a bank can allow, authenticate, review or reject agent activity by policy. LexisNexis says many customers already use it, and a November 10 release extends it to government agencies.

BioCatch launched AiRE, short for Agentic AI Risk Essentials. It flags AI agents in digital banking sessions, whether they act for the account holder or for someone with worse intentions. Feedzai added agent detection to Digital Trust, weighing what an agent says about itself against device and behavioral signals. It also added cryptographic device binding, so a trusted phone can count as something the customer has.

Callsign went furthest on what happens next. A heavily automated sign in now gets one of three labels: a known account aggregator such as Plaid or MX, a suspected AI agent, or a bot. The bank sets the policy for each, and a suspected agent can be sent to the real customer for approval inside the bank’s own app. Callsign also keeps agent sessions out of the customer’s behavioral profile. A model that learns how you type should not start learning how your agent types.

Sardine now returns the agent by name, starting with Meta’s Muse and OpenAI’s ChatGPT agent, so a merchant can write one rule for browsing and another for checkout. Then the agents started paying. When OpenAI launched ChatGPT Wallet, Forter began identifying ChatGPT agent traffic on each wallet transaction and feeding it into its fraud decisions, using a signal set it says it agreed with OpenAI.

The human side of identity moved too. Veriff launched Presence Verification, a check with no ID document that confirms one live person is completing the session. A higher tier watches and listens for signs that someone is coaching the user, in 98 languages. Proof added session signals such as camera metadata to its deepfake detection, with reason codes a customer can write policy against. And Sift opened its network as a standalone API, so a business can look up a user’s chargeback history and cross network links without running the rest of Sift.

Our read

A year ago the agent question in fraud had one answer, which was to block it. Now there are at least four: a known aggregator, the customer’s own agent, somebody else’s agent, and a plain bot. That turns detection into a policy product. The bank writes the rules, and the vendor’s job is to make the label good enough to write rules on. The catch for anyone selling here is that Know Your Agent is now a category with LexisNexis in it, so the name alone no longer sets a product apart. The path back to the real customer, the way Callsign built it, is where products will start to look different.

The agent took a driving test

Last issue the agent got the last click. This period several vendors published how an agent earns it, and the answer looks a lot like a learner’s permit.

Complead added a shadow mode. A new AI agent works live cases, records the decision it would have made, and does nothing. A comparison screen shows how often it agreed with the analysts and what it missed, broken down by rule. Three checks inform go live: 100 cases decided by analysts, a 95 percent agreement target and zero missed true positives. Then a person flips the switch, and the flip is recorded.

ThetaRay published what the far side of that test looks like. At Travelex, its screening agent cut point of sale sanctions alert resolution from about 510 seconds to 9, and now closes 80 percent of alerts on its own across more than 20 countries. Before that it ran three months in production shadow on thousands of live alerts. It disagreed with senior investigators 18 times, and ThetaRay reports no case where it under screened.

Fingerprint applied the same idea to fraud models. Its new Fraud Score trains a separate model for each fraud type a customer tracks, using that customer’s own labeled outcomes. Fingerprint reviews each model, and nothing goes live until the customer approves it in the dashboard.

Some vendors had the test set by someone else. Q2’s assistant earned AIUC-1 certification after adversarial testing and a controls audit with Grant Thornton. The testing covered prompt injection, jailbreaks, unauthorized actions, data leakage, hallucination and unsafe tool calls, and Q2 says it will be retested every quarter. Baker Hill achieved ISO/IEC 42001 for its AI management system and named the certification body, Schellman, along with the audit and the product in scope. Last issue, two 42001 announcements differed mainly in whether anyone could check them. This one can be checked.

Others wrote down their own exam. AgentSmyth published the rules it uses to decide a research agent is ready to ship. Every run must clear nine minimum checks, every test case runs three times with the worst run deciding, and an unproven check never counts as a pass. ToltIQ published how it picks the models in its chat feature, with scores on 250 credit agreement questions and the models it leaves out and why. UPTIQ launched GovernAI, a console that registers agents built on any major platform, sets what each may and may not do, and scores and logs them. Its exam ready reports are listed as planned, not shipped.

And several simply kept a person in the seat and said so in the release. Fourthline’s Recurring Fraudsters check uses facial recognition to spot people previously flagged for fraud trying to onboard again, and a person reviews every match. Capsa’s Self-Improving Skills drafts better instructions for a firm’s agents, and nothing changes until an admin approves. WorkFusion’s screening agent now sends alerts with future or blacklisted dates to a person instead of resolving them.

Our read

Shadow mode may be the most useful idea in this issue for a buyer, because it turns “trust us” into a number your own analysts produced. Complead put the test in the product, ThetaRay put the result in a case study and AgentSmyth put the exam in writing. In each case the evidence comes before the agent acts, not after something goes wrong. The roadmap item more or less writes itself: ship the shadow run, show the agreement rate, and let the customer set the bar.

Every answer came with a receipt

The third pattern was small in each release and large in total. Vendors stopped asking people to trust a filled in answer and started showing where it came from.

Herald now puts a source button beside every answer its placement system fills in from submission documents and carrier quotes. Click it and the document opens beside the form, with the text it used highlighted in blue and the answer itself in amber. Hover, and Herald explains why it chose that answer.

Marloo rebuilt Forms along the same lines. It can fill up to ten client forms at once, and every field carries a citation and a status: cited, inferred, in conflict, or entered by the user. Inferred may be the most candid word to appear on a client form this year.

73 Strings links every figure it extracts from a portfolio company report back to its place in the source. Hebbia launched knowledge graphs that store each connection beside the source it came from, keep every version of a fact with its effective date, and say so when access rules limit an answer. ValidMind now drafts model documentation from test results and attached files, keeps the sources behind each accepted draft, and flags text for review when the evidence underneath it changes. That last part is the one a validator will like, because documentation usually goes stale without telling anyone.

The explanations also reached the decision itself. CRIF made counterfactual explanations visible in its ORCHESTRA platform. For a declined application, it shows realistic changes, such as a smaller loan or a different repayment period, that could change the outcome. Prove’s new Signals endpoint returns a risk level, at most one reason code and the feature values the model used. DigiFi’s application assistant explains why a decision passed or failed, and cannot change a record.

Oscilar made the most interesting architectural move. It disclosed a tabular foundation model for fraud, AML, credit and device risk that reads each customer’s own labeled history at decision time, instead of training a separate model per customer. It keeps SHAP attributions and reason codes, and it can be distilled into a standard XGBoost model, the kind a validator already knows how to review.

Our read

A year ago explainability was a slide. In these releases it is a button, a highlight, a status label or a reason code. That is a better product and a better sales asset, because a reviewer can click it during the demo. Oscilar’s release shows where this goes next for model teams. New architecture is welcome as long as it can still be shown in a form the validator already trusts. Novel inside, familiar at the review.

The agent picked up the phone

Most agents in this log wait to be asked. A few this period start the conversation.

Lorikeet launched Dreamweaver, which gives each customer their own agent with a goal. It wakes on a schedule or when something changes, then decides whether to call, text, chat or email, create a task for staff, or do nothing. Lorikeet says every message is checked against contact hours, limits and consent, and a holdout group measures whether any of it helps. Doing nothing is one of the agent’s options, which is more restraint than most inboxes see in a week.

InDebted launched Voice, an AI agent that makes collections calls, verifies identity, agrees payment plans within client limits and texts a payment link during the call. It runs a compliance check before every dial and stops dialing if the check cannot run. Hardship cases go to people, with a summary.

Prodigal made the identity check on a call faster. A small model now confirms which detail the consumer gave in about a tenth of a second, and it handles 28 to 45 percent of verification turns. Fundamento published results from a voice and WhatsApp deployment at insurer Axis Max Life, including a 61.6 percent renewal recovery rate.

The written channel moved the same way. Narmi now drafts a reply to a secure message the moment it arrives, with confidence tags on drafts that need a closer look. GReminders released an Outlook assistant in beta that drafts adviser replies with the client’s holdings and history, and only for emails that match a client in the CRM. Floatbot put its live agent assist inside NiCE CXone as a desktop app.

Our read

Outbound is where agent risk changes shape. An agent that answers can only be wrong to the person who asked. An agent that calls first can be wrong to someone who never asked, at the wrong hour, about a debt they are struggling with. InDebted’s stop rule and Lorikeet’s holdout group are the two details worth copying. One shows the agent stayed inside the rules. The other shows the call was worth making at all.

Market notes

The assistant integrations kept coming, with more care about what they are allowed to do. Middesk released a Claude plugin that can order verifications and asks for explicit confirmation before anything billable, which may be the most relatable guardrail in the log. AlphaSense launched an MCP server that outside agents can hand a research task to, with each user’s entitlements enforced. Unit21’s MCP server can now read alerts under existing permissions. Jump added a read only ChatGPT plugin, and Carta built a dashboard for limited partners inside its Claude plugin, with each investor seeing only their own funds.

The data kept moving toward the agent. S&P Global added corporate hierarchy tools to its kFinance library for agents, and Rogo added GLG expert call transcripts as a source. Avaloq launched Mosaic in the US with Agent Connect, which gives a firm’s AI tools governed access to investment data. Adclear opened its marketing compliance review to Claude, ChatGPT and in house assistants through MCP. Shortcut lets remote agents find the Excel workbooks a user has open, while still asking that user to approve every action. And Coris now serves every page of its API docs as Markdown, indexed for AI coding tools as much as for people.

“Agentic” got a price tag. Dynamic Planner’s Agentic License costs £59 per user per month, sold alongside a Review or API license. Riskified published flat plans for its Zendesk product at $55 and $109 a month, with a 30 day free trial. YCharts priced its new alternatives data add on at $600 per license per year. Argos Identity’s new Unit API charges only for calls that return a result. For anyone who has tried to find the AI on a quote, this is progress.

Screening is infrastructure now, and infrastructure needs a backup plan. Elliptic made Fast Track Monitoring generally available, a backup wallet and transaction screening service for firms whose main provider is unavailable, priced as a setup fee plus usage with no long contract. Chainalysis, after a security incident in which an outside actor reached some of its systems, now requires customers to register the IP addresses allowed to reach KYT, Address Screening and Reactor. Every user had to reset their password, and every earlier API key was invalidated. Deconflict made wallet sanctions screening available to all its customers, with a third answer besides Flagged and Clear: Unavailable, when the list data cannot be confirmed as current.

The rest of crypto compliance filled in the edges. AMLBot launched a Travel Rule suite that grades the completeness of transfer data against the EBA guidelines. Credora rated seven wrapped Bitcoin tokens, with annual default probabilities from 0.080 to 5.319 percent. Scorechain rebuilt its investigation graph so one case can span several blockchains, and Global Ledger now powers wallet screening inside Waml. TRM Labs launched a managed service that traces illicit funds and routes freeze requests to stablecoin issuers and exchanges for any verified law enforcement agency. And Broadridge let holders of tokenized US equities vote their proxies.

Government became a segment with a certificate attached. SentiLink and LeapXpert both announced FedRAMP Class C. V7 Go reached GovRAMP Core and sits on the public authorized list. Consilient listed its federated learning models on AWS Marketplace, including GovCloud, where only model updates leave the institution. LexisNexis extended Know Your Agent to agencies, and TRM Labs built its new service for them.

In lending, one vendor put money behind its number. Candor launched Income Clarity, which returns a qualifying income figure per source in under ten minutes and carries Candor’s own repurchase warranty on Conventional, FHA and VA loans. The last issue’s Index Answer found that almost no vendor says who answers when the AI is wrong. Here is one that does, in the currency mortgage lenders care about most.

Blend released a mobile app for loan officers that prepares actions such as pulling credit and sending disclosures, which run only after the officer approves. Ocrolus opened its loan file findings through an API. Lama AI wired in IRS tax transcripts through Tax Guard, recorded Partially Verified. MeridianLink added Zest Protect fraud scores to its origination flow, and Glia released a business prospect profile backed by cited evidence.

Insurance handed agents the portfolio. IntellectAI launched a portfolio steering app that shows where a book is drifting and lets underwriters reprice, refer or nonrenew one policy or thousands. The insurer sets, action by action, whether an agent handles the work or a person reviews it. Vertafore’s Change Request Agent compares renewal and endorsement requests with existing policy data for MGAs, and underwriters edit its summaries before they are saved. ZestyAI’s water risk model was accepted by regulators in Minnesota and Maryland, bringing it to 22 states. BriteCore added Verisk Lightspeed as an address provider, chosen per site without code.

The published results were unusually specific. Carrington Labs reported that Beforepay’s credit loss rate fell from 5.2 to 1.1 percent of amount lent, defined its metrics, and disclosed ten model releases, each tested before going live. Fraudio published fraud rates by its own score tiers across about 127 million transactions. Transactions it flagged that were approved anyway ran more than four times the fraud rate of unflagged ones, which is a polite way of saying the yellow light means something.

Three more put numbers on a named customer. Outmarket AI reported loss analyses at one brokerage falling from an hour to six minutes. Applied Systems reported commission reconciliation at Ames & Gough falling from two or three days to about half a day. Alkami reported more than half of Texoma Community Credit Union’s members adopting its platform within two weeks.

The small print on models is worth reading. Microblink now lets clients pin a previous release of its identity verification with a single request header, which is quietly a big deal for anyone validating it. Reality Defender’s audio model now analyzes speech it used to skip, and warns that scores for the same file can differ from before. F2 rebuilt its deal agent so skills, templates and memory live in its own harness rather than in any one model, and it routes work across several providers, open weight models included. ibl.ai turned on a sandbox virtual machine with public internet access by default for new agents, and tells tenants how to switch it off.

Agent controls got finer in a handful of small ways. Kore.ai can now stream the safe start of a reply while it checks the rest for personal data. Fiddler AI’s gateway now scans tool call arguments from earlier turns of a conversation. Linkurious’s Copilot can act on an investigation graph, limited to the actions an administrator allows. Advocate Technologies replaced Cancel with Stop, so a user can halt its case agent and redirect it without losing the work. Continuum now logs who created, edited and published each AI built client page. And Theta Lake began pulling AI incident detections from CrowdStrike into compliance review, with retention and legal hold for AI conversations. The chat with the assistant is now a business record.

Identity filled in the gaps. AiPrise launched Token Sharing, so two companies on the same transaction can share a customer’s verified KYC evidence with consent, sources and audit trail included. Trulioo’s business and person checks now run inside FrankieOne. Incode lets customers write rules on the individual reasons behind its fraud signals. iDenfy added bank card verification on its own, ADVANCE.AI verified Thai users through the government ThaID app, and Accertify brought its account protection scores into Ping Identity’s orchestration.

In wealth, Nitrogen’s agent can now plan and carry out a multi step job from one request, such as building a client profile, a portfolio and a proposal, with the adviser approving each action or several at once. Holistiplan checks a tax return against a library of strategies and estimates each one’s impact. Feathery connected to Goldman Sachs Custody Solutions for onboarding and transfers, and Entrilia now posts private credit loan activity from LoanStreet straight into fund accounting.

A few small ones will save someone an afternoon. Bretton AI’s agents now fill an institution’s own Word, PDF or Excel templates from a finished case. Strise sorts monitoring alerts by risk instead of time, and shows what is new in each sanctions or PEP hit. SEON exposed 123 velocity measures, such as cards per user per day, as rules and signals. Bits Technology took its no code compliance rules engine out of beta, with a full record of who changed what.

What the two weeks say about the space

A hundred and twenty three entries, and one idea running through them. The agent is now expected to show its papers. The fraud vendors check its identity at the door. The compliance vendors make it ride in shadow before it drives. The workflow vendors make it cite a source for every field. The servicing vendors make it check the rules before it dials.

For the people who build and market in this category, that changes what a launch looks like. The feature is the agent. The proof is everything around it, from the label and the shadow run to the citation, the stop rule and the certificate. The last issue said autonomy sells the demo and the boundary sells the renewal. This period vendors started shipping the evidence that gets a product from one to the other.

The most common proof in this log was a certificate, so the Index Answer below asks how many vendors actually let a buyer open one.

Index Answer

Which financial services AI vendors let a buyer open the certificate?

About one in 13. Of the 611 vendors the AI FinTech Index has graded on Security Certifications and Trust Center, 48 reach the top grade. That grade means the certifications are named with their type and offered as documents a buyer can retrieve, usually through a trust portal that opens without asking.

Another 274 name a recognized certification, such as SOC 2 Type II or ISO 27001, in their own material, but without the document itself, or with a scope or renewal question the buyer has to raise. A further 283 offer a footer line or certifications asserted without being named, and 6 publish nothing on security posture at all.

So most of the field says the right words, and very few hand over the paper. A logo on a security page is a claim. The certificate, with its scope and its dates, is the evidence. The distance between the two is roughly a week of emails in every procurement review.

This period’s log shows the new AI certifications arriving on the same road, a few years behind. Baker Hill named the body that certified its AI management system, the audit and the product in scope. Q2 named the firm behind its controls audit and committed to quarterly retests. V7 Go’s GovRAMP status sits on a public list under an ID anyone can look up, and SentiLink and LeapXpert announced FedRAMP Class C. The axis also covers how a vendor handles a security incident, and Chainalysis published what happened, what it changed and what every customer had to do next.

The practical point is short. If you hold a certification, put the document where a buyer can open it and say what it covers. The AI certifications will be judged the way the security ones already are, and the vendor that can be checked without an email reaches the shortlist first.

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

The AI FinTech Index Brief is published by AI FinTech Index, an independent reference for evaluating AI vendors in financial services. No vendor pays for inclusion, placement, or rating. Compare any indexed vendors by capability at Compare and read the evaluation standards at Methodology.

Contact us

Found a vendor we missed? Have feedback on the index? We’d love to hear from you.

AI FinTech Index

The AI FinTech Index is an independent index that tracks changes to AI vendors in financial services. It holds 611 vendors across banking, lending, insurance, wealth, capital markets and financial crime compliance, each graded on the same 15 capability axes from public sources. No vendor pays for inclusion, placement, or rating.

Index Status
Last index update
November 13, 2026
The AI FinTech Index is an editorial reference, not a regulatory body. Vendor data is verified against published sources and public regulatory filings. Figures labeled “Estimated” have not been confirmed by the vendor. See the Methodology page for evaluation standards and limitations.
© 2026 AI FinTech Index
3801 N Capital of Texas Hwy, Ste E240 · Austin, TX 78746