The AI FinTech Index Brief

The week in one line

Alloy let risk teams build their own attributes without filing a support ticket, Guidewire made core customisation upgrade itself, and ICE put the Fannie Mae income calculator inside the origination system. The work being removed this week was not the buying decision. It was the implementation project that always followed it.

This issue covers August 30 to September 5: 18 entries across 18 vendors, eight Verified against the vendor’s own materials and ten Partially Verified from third party reporting. Funding rounds, valuations, and awards are not logged, here or anywhere on this index.

The integration stopped being a project

Alloy launched the Alloy Marketplace, merging its Partner Center and Policy Library into one hub of more than 270 partner integrations with pre built decisioning snippets, a guided merge flow and conflict detection.

The larger half of that release is the Attribute Tool reaching general availability. Customers can now browse, build, test and modify custom attributes from partner API responses themselves, with a changelog per integration. Needing a support ticket to create an attribute is the kind of dependency that turns a two day integration into a two month one, and removing it changes how quickly a risk team can test a new data vendor.

Guidewire did the same thing one layer down with an Extension Layer and Extension Manager for the Guidewire Cloud Platform, isolating InsuranceSuite customisation into managed and unmanaged modules so extension upgrades apply automatically with backward compatibility validation.

The thing being deleted there is the manual three way merge. Extension upgrades are the reason InsuranceSuite deployments fall behind release versions, and falling behind is how an insurer arrives at a security patch it cannot take without opening a project.

ICE Mortgage Technology made Fannie Mae’s Income Calculator generally available inside Encompass and the ICE Income Analyzer, so self employed income is calculated from extracted tax return data without leaving the origination system. It sits on the same screen as last year’s Freddie Mac AIM integration.

Income calculation is where underwriting rework concentrates, and every portal switch is an opportunity to transcribe a figure wrong. Running the agency’s own calculator in place moves the representation and warranty question from the lender’s spreadsheet to the agency’s logic.

Candor launched Candor DI for touchless indexing, classification and extraction across loan files, and the word carrying the weight in the announcement is deterministic. Candor states the cross source reconciliation between paystubs, tax returns, credit reports and the URLA is rule based even where the extraction is not.

That distinction is the whole product. Reconciling income across four documents with a probabilistic model produces an answer a lender cannot defend in an audit. And Jump’s August release added SCIM provisioning through Okta and Microsoft Entra, which is the least interesting feature in this issue and the one an auditor will ask about first, because manual deprovisioning is the control that gets missed.

Our read

Every vendor in this section already had the integration. What shipped was the removal of the work that came after it: the support ticket, the merge, the second portal, the reconciliation spreadsheet, the leavers list. That work has always been carried by the buyer and has never appeared on a licence line, which is why it rarely gets compared during selection and always gets discovered during implementation.

Buyer question

For any platform in a live evaluation, ask who creates a new attribute or field, how long it takes, and what it costs. Then ask what happens to your customisations when the vendor ships its next release. The answers separate two products that look identical in a demo.

The data provider stopped waiting to be logged into

In the August 30 issue the pattern was vendors opening their products to agents they do not control. This week the same move arrived from the other side of the table, at the firms that sell the data.

Marloo agreed to bring Morningstar investment data and research into its adviser platform, and pulled it from the Morningstar MCP Server rather than a batch feed, so fund and model portfolio data can be queried in natural language. Licensing a recognised provider answers the question every compliance officer asks, which is where the numbers came from. Sourcing it over MCP makes the currency of those numbers a live dependency rather than a scheduled one.

MioTech signed definitive agreements to combine with the green finance and ESG business of CCX Group, and the first stated technical output of the combination is a Model Context Protocol service exposing the group’s sustainability data and evaluation methodologies to client systems.

A merged data business whose opening artefact is a protocol endpoint has decided it expects to be queried rather than logged into. For ESG reporting that is the correct read of where the work actually happens, which is inside somebody else’s spreadsheet at the end of a quarter.

The compliance platforms moved the same way. CUBE added three agentic coworkers to RegPlatform: one scoring regulatory updates for relevance, one for natural language querying of regulations, one consolidating enforcement actions across jurisdictions.

Relevance scoring is the load bearing feature there. Regulatory monitoring has never had trouble finding updates, it has had trouble telling a compliance team which of several hundred this month actually bind it, and a score nobody can explain is a score nobody can rely on the day they skip an update.

Spektr released version 3.0 with real time transaction monitoring, a Policy Engine that turns written compliance policies into executable rules, and Case AI, which gathers evidence and recommends investigation outcomes with citations and confidence levels. Onboarding and monitoring have historically been bought from different vendors and reconciled by hand, which is why so many alerts arrive without the customer context needed to close them.

FP Alpha released Tax Projector Scenario Insights, modelling Roth conversions, asset sales, income changes and state relocations, with an agent that builds and adjusts the scenarios conversationally. Most adviser tax tooling reads a filed return, which is a record of decisions already made. Modelling forward is more useful and more dangerous, because a projected saving delivered in conversation is very easy to repeat to a client as advice.

Our read

The query surface is moving off the portal and into whatever assistant the buyer already uses, and this week it was the data owners doing the moving rather than the application vendors. That is a bigger shift, because a data business that is queried rather than visited loses the session, the audit log and the interface it used to control. Two of these four shipped citations and confidence levels alongside the answer, which is the only version of this that survives a regulator asking where a number came from.

Buyer question

When a vendor exposes data through MCP or an agent, ask what the response carries besides the answer: the source, the as of date, the confidence, and whether your own logs record the query. A conversational answer with no provenance is a citation you will have to reconstruct later, under time pressure, in front of somebody who is not amused.

Vendors sold a segment rather than a capability

FIS launched an Embedded Banking Platform letting US banks place their own accounts, card issuing, receivables, payables and expense management inside third party business software, through APIs, SDKs, widgets or white labelled applications, with pilot banks including Cogent Bank, Commercial Bank of California and M&T Bank.

The structural point is that the accounts stay on the bank’s own balance sheet rather than on a third party virtual ledger, which is precisely what separates this from the sponsor bank model regulators spent two years examining. Note the date though: accounts and payments are stated as a fourth quarter plan, not a shipped capability.

Sumsub released a Brazil localised onboarding path for betting platforms that retrieves identity data and a document image from local databases, so a player supplies a CPF number and a liveness selfie instead of photographing a document, with retrieval requiring explicit consent under the LGPD.

Brazil’s regulated betting market is new enough that requirements are still settling, and buying a localised path is a bet that operators would rather buy the interpretation than maintain it. Removing the document upload also removes the operator’s own copy of the capture, so confirm what the retained audit record actually contains before a regulator asks rather than after.

WealthAi launched WealthAi for Advisors aimed squarely at independent advisers and smaller firms, consolidating notetaking, client management, document generation and compliance monitoring around a Client File workflow and designed to connect to existing practice management systems rather than replace them. The reported 60 percent reduction in routine client administration is the vendor’s own beta observation and should be read as one.

Verisk released a US Data Center Exposure Database covering more than 2,500 facilities with rooftop level geocoding, construction and capacity detail, building footprints and a 90 metre disaggregation grid. Data centres became a concentrated, high value and poorly modelled exposure class in a very short space of time, and accumulation risk around them is currently estimated rather than measured.

Two more in the same shape. Fenergo added the LexisNexis WorldCompliance Data Plus watchlist alongside the legacy dataset, citing richer structured data and more frequent sanctions updates, and update frequency is the entire argument given that intraday screening obligations are not satisfied by a daily refresh. Identomat shipped Mercury, a verification rebuild spanning interface, liveness, document capture and cross device handoff.

Our read

A general capability sold into a specific segment is a product. A specific segment sold as a capability is a maintenance commitment, and the failure mode for jurisdiction and segment products is not the launch, it is the eighteenth month, when the regulator moves and the localised path has to move with it. Ask how the Brazil path gets updated and who decides, because that answer is the product you are actually buying.

Buyer question

For any segment specific or jurisdiction specific product, ask when it was last updated for a rule change, how the vendor learned about that change, and how long the update took from publication to production. A vendor that cannot answer the third question is selling an interpretation with no maintenance plan behind it.

Market notes

Nasdaq completed its acquisition of Dasseti, first announced in July, and is folding due diligence and monitoring into Nasdaq eVestment so RFP, questionnaire and database workflows sit in one environment. Anyone with Dasseti in procurement is now buying from Nasdaq, with the pricing and contracting posture that implies.

ACI Worldwide agreed to acquire Cranium Ventures and its SYNAP microservices card switching framework, to be folded into ACI Connetic for Cards, expected to close in the third quarter. Card switching is one of the last parts of the payments stack still commonly running on architecture designed decades ago, and buying the microservices version rather than writing it is a statement about how long writing it would have taken.

With the MioTech and CCX combination above, that is three corporate events in one week and all three fold a capability into a platform the buyer was already using. The consolidation pattern this log has now recorded for three straight issues holds: the acquirer is usually a vendor already inside the stack, so a shortlist built around one product ends up running through another company’s contract.

What the week says about the space

Put the three sections together and they describe the same movement at different depths. The application vendors removed the implementation work that sat behind their integrations. The data providers moved their answers out of the portal and into somebody else’s assistant. The platform vendors stopped selling a general capability and started selling a jurisdiction, a segment or an exposure class.

What connects them is that the differentiator has moved past the feature and into the cost of living with it. Integration was the moat for a decade in this market. This week four vendors treated it as an expense to be deleted, which is what happens to a moat once everyone has crossed it.

Index Answer

Which financial services AI vendors document how they connect to the core?

Most of them, which makes this the exception in this index rather than the rule. Of the 490 vendors the AI FinTech Index has graded, 151 name the systems of record they integrate with in a way a buyer can verify, in a marketplace listing or in public API documentation.

Another 266 name a system or document a public API with the depth left open. Only 73 claim integration through standards or connectors without naming anything, and not one vendor publish no integration evidence at all. On most axes the bottom bands are where the population sits. Here they are nearly empty.

Now set that against the same 490 vendors on two other questions. On commercial transparency, 8 publish enough about price to reach the top grade and 429 sit in the bottom band. On governance and bias disclosure, 6 reach the top grade.

Read together those three numbers say something plain about disclosure in financial services AI. It is not a cultural trait and it is not evenly distributed by risk. It follows sales necessity. A vendor that cannot name the core it plugs into never reaches the demo, so it names the core. A vendor that will not publish a price still gets the meeting, so it does not publish the price.

The practical consequence for a buyer is that the axis where almost everyone grades well is the axis where the grade tells you least. Naming the core is now table stakes, and this week’s entries are vendors competing on what happens after the connector exists: who can create an attribute, whether an upgrade applies itself, whether the agency calculation runs in place or in another tab.

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