Skip to main content
    Kriv AI

    01, Financial Services AI Platforms

    Databricks AI for Financial Services

    What building and governing AI on Databricks actually looks like for wealth managers, banks, and insurers: portfolio analytics and advisor tooling, MCP-connected agents, and model risk documentation mapped to SR 26-2.

    Databricks financial services AI means building and governing machine learning and generative AI workloads for banks, wealth managers, and insurers on the Databricks Data Intelligence Platform: a unified lakehouse spanning data engineering, MLOps, and Mosaic AI agent tooling, layered with model risk and compliance controls organized around SR 26-2, the current interagency guidance on model risk management.

    databricks ai financial

    What Databricks AI for Financial Services Actually Means

    Databricks financial services AI means building and governing machine learning and generative AI workloads for banks, wealth managers, and insurers on the Databricks Data Intelligence Platform: a unified lakehouse spanning data engineering, MLOps, and Mosaic AI agent tooling, layered with model risk and compliance controls organized around SR 26-2, the current interagency guidance on model risk management.

    Databricks itself is infrastructure: Unity Catalog for data and AI governance, Delta Lake for the underlying storage layer, MLflow for experiment tracking and the model registry, and Mosaic AI for building, serving, and evaluating models and agents. None of that is finance-specific out of the box. It becomes financial services AI when the workloads running on top of it are wealth management personalization, credit and fraud models, treasury forecasting, or agentic workflows that touch account data. One worked example of an agentic workflow on account data is agentic AML alert triage on Databricks.

    Financial institutions gravitate to Databricks because a single platform can hold structured core-banking and portfolio data alongside unstructured content, such as advisor call notes, compliance filings, and client correspondence, under one lineage and access-control model. That matters more in a regulated environment than raw model accuracy, because an examiner or a model risk committee may ask where a number came from, and a traceable answer is better than a rebuilt pipeline assembled for the review.

    The gap Kriv works in is what sits between a working Databricks deployment and an AI program a model risk committee or a bank examiner can review: model inventory discipline, validation evidence for models that were not built the way traditional statistical models were, and a documented record of who, and what agent, can call which tool and dataset.

    where wealth management

    Where Wealth Management Firms Actually Use AI on Databricks

    Wealth management firms running AI on Databricks tend to start in four places: portfolio and household-level analytics, advisor productivity tooling built on meeting and email content, suitability and KYC document extraction, and custodian data reconciliation using Delta Sharing. None of these require replacing the core portfolio accounting or CRM system Databricks sits alongside.

    Portfolio and household analytics: combining custodian feeds, market data, and CRM records in Unity Catalog to surface next-best-action signals for advisors, model portfolio drift against target allocations, and household-level risk exposure across accounts that live in separate source systems.

    Advisor productivity: using Mosaic AI to summarize client meetings, draft follow-up notes, and pull relevant portfolio context into a single advisor workspace, with every generated note traceable back to the source records it was built from rather than presented as an unattributed summary.

    Suitability and KYC document extraction: turning scanned account applications, investment policy statements, and beneficiary forms into structured fields a compliance team can query and audit, instead of leaving that information locked inside PDFs and scanned images.

    Custodian and clearing-firm reconciliation: Delta Sharing lets a firm receive custodian or clearing-firm data through a governed, open protocol rather than a proprietary point-to-point feed built separately for every counterparty, which matters for a wealth manager working across multiple custodians.

    Kriv AI's Databricks wealth and portfolio accelerator packages the data model and access-control pattern behind the first two use cases, so the underlying Unity Catalog structure does not have to be designed from a blank page.

    mcp powered financial

    MCP-Powered Financial AI Workflows on Databricks

    MCP, the Model Context Protocol Anthropic introduced in November 2024 as an open standard, gives an AI agent a governed way to call external tools and data sources instead of relying on a fixed set of hand-wired integrations. Databricks supports MCP in two ways relevant here: Unity Catalog functions can be reached through a managed MCP endpoint (in Public Preview, and a workspace admin must enable it), and custom MCP servers can be hosted on Databricks Apps.

    In a financial workflow, that looks like an agent that, on a request from an advisor or an analyst, calls a Unity Catalog function to pull a client's current holdings, calls a second function to check a concentration or suitability rule, and drafts a summary, with each call logged against the same access controls that already govern direct table queries.

    The practical shift MCP introduces is scope, not raw capability: an agent that previously only read from one approved dataset can now be connected to many tools at once, and can chain calls across them in a single request. That is useful for automating multi-step workflows, and it is also the reason MCP tool permissioning needs to be treated as seriously as data-table permissioning, not as a lighter-weight engineering configuration decided outside the model risk process.

    Kriv AI treats MCP tool definitions as part of the model and agent risk inventory: each tool a financial services agent can call gets documented, scoped to the narrowest data access it actually needs, and put through the same validation and change-management process as a traditional model, rather than left as a configuration detail nobody outside the build team can see.

    context

    How Does MCP Access Control Work on Databricks?

    Databricks documentation says Databricks-provided and registered MCP servers use Unity Catalog permissions, and that Unity Gateway can be used to control access, manage credentials and monitor usage for the servers it governs. The Unity Catalog functions endpoint is not one of those: it relies on permissions on the underlying resources. Servers can be Databricks-provided (for example the Genie One server), external servers added through the MCPs page, or custom servers built with Databricks Apps templates. Each server and tool an agent can reach is one more route to financial data. We record every one in the model and agent inventory, with the narrowest data access it needs and the person who approved it, so a committee can see the list without asking the build team.

    governance layer databricks

    The Governance Layer Databricks Doesn't Ship With

    Databricks provides the technical building blocks, an MLflow model registry, Unity Catalog lineage, and monitoring dashboards, but it does not provide a model risk management program. That program, and the documentation an examiner or a model risk committee may ask to see, is separate work that sits on top of the platform.

    SR 26-2, issued jointly by the Federal Reserve, the OCC, and the FDIC on April 17, 2026, is the current interagency supervisory guidance on model risk management. It supersedes and replaces SR 11-7, the 2011 guidance, and retains effective challenge as part of sound model risk management. The guidance says its principles apply to traditional statistical and quantitative models, such as the gradient-boosted risk scores a bank might run on Databricks, and that generative AI and agentic AI models, such as fine-tuned language models, are not within its scope.

    SR 26-2, the April 2026 interagency guidance that supersedes and replaces SR 11-7, says generative AI and agentic AI models are not within its scope. That does not mean those tools go ungoverned: the guidance says a banking organization's own risk management and governance practices should guide the controls for tools it does not cover. The voluntary NIST AI Risk Management Framework is one reference a firm can use for that.

    For wealth management and broker-dealer functions specifically, FINRA's Regulatory Notice 24-09 is explicit that existing supervisory, recordkeeping, and communications rules apply in full when a firm uses generative AI. There is no separate, lighter rulebook for an AI-assisted advisor workflow just because the output was drafted by a model.

    In practice, the deliverable Kriv builds alongside a Databricks financial services implementation is a model and agent inventory that maps each system as in or out of SR 26-2 scope, documents validation evidence appropriate to how it was built, whether that is a traditional statistical model, a fine-tuned LLM, or an MCP-connected agent, and sets a monitoring cadence a committee can actually review on a recurring schedule.

    context

    Which Banks Does SR 26-2 Apply To?

    SR 26-2 is the Federal Reserve letter, dated April 17, 2026, that transmits the agencies' revised model risk management guidance. The Federal Reserve says the letter is expected to be most relevant to banking organizations with over $30 billion in total assets that it regulates. It supersedes and replaces SR 11-7 (April 4, 2011) and SR 21-8, the 2021 statement on model risk management for bank systems that support Bank Secrecy Act and anti-money laundering compliance. The guidance covers traditional statistical and quantitative models and non-generative, non-agentic AI models. Generative AI and agentic AI models are outside its scope. Whether it applies to your institution depends on your regulator and size, and we confirm that in the first assessment.

    kriv ai databricks

    What a Kriv AI Databricks Financial Services Engagement Includes

    Engagements are scoped to the institution's actual Databricks footprint and regulatory posture rather than sold as a fixed package. Four phases recur across most of this work, in roughly this order.

    1. 1. Platform and workload assessment

      Review of the current Databricks environment, Unity Catalog structure, MLflow registry contents, Mosaic AI agents in production or planned, against the financial workloads it actually supports, to establish what needs formal governance versus what is still exploratory.

    2. 2. Model and agent risk inventory

      Every model, fine-tuned LLM, and MCP-connected agent is catalogued and mapped as in or out of SR 26-2 scope, with gaps flagged against validation, documentation, and effective-challenge principles.

    3. 3. Governance control implementation

      Unity Catalog permissions, MCP tool allowlisting, and change-management workflows are built or tightened so access controls match the risk tier of each model and tool, instead of a single default applied everywhere.

    4. 4. Monitoring and audit readiness

      Drift monitoring, a validation refresh cadence, and inventory documentation are set up so a model risk committee or examiner review starts from records that already exist, instead of ones assembled the week before the exam.

    engagement

    How engagements are scoped and priced

    Every engagement is scoped and priced after a short discovery call.

    Straight answers

    Frequently asked questions about Databricks AI for Financial Services

    Does Databricks provide AI model risk management for financial services out of the box?

    No. Databricks provides the platform, Unity Catalog for governance and lineage, MLflow for the model registry, Mosaic AI for building and serving models and agents, but it does not provide a model risk management program. Mapping models and agents as in or out of SR 26-2 scope, documenting validation evidence, and setting a monitoring cadence is separate work that sits on top of the platform.

    How do wealth management firms use AI on Databricks?

    The most common starting points are portfolio and household-level analytics built from custodian and CRM data in Unity Catalog, advisor productivity tools that summarize client meetings and draft follow-ups, suitability and KYC document extraction, and custodian data reconciliation using Delta Sharing rather than point-to-point feeds.

    What is an MCP-powered financial AI workflow?

    MCP, the Model Context Protocol, is an open standard, introduced by Anthropic in November 2024, that lets an AI agent call external tools and data sources through a common interface instead of custom point-to-point integrations. Databricks offers Unity Catalog functions through a managed MCP endpoint (Public Preview) and can host custom MCP servers on Databricks Apps, so a financial agent can pull holdings data, check a compliance rule, and draft output in one governed sequence.

    Does SR 26-2 apply to generative AI and MCP-connected agents built on Databricks?

    SR 26-2, the April 2026 interagency guidance that supersedes and replaces SR 11-7, says generative AI and agentic AI models are not within its scope. That does not mean those tools go ungoverned: the guidance says a banking organization's own risk management and governance practices should guide the controls for tools it does not cover, firms can also use the voluntary NIST AI Risk Management Framework as a reference, and Kriv treats MCP tool access the same way it treats a model in the inventory regardless of whether it falls within SR 26-2 scope.

    Which banks does SR 26-2 apply to?

    SR 26-2 is the Federal Reserve letter of April 17, 2026 that issued revised interagency model risk management guidance. The Federal Reserve says it is expected to be most relevant to banking organizations with over $30 billion in total assets that it regulates. Whether it applies to a given bank, wealth manager or insurer depends on that institution's regulator and size. We confirm that scope in the first assessment.

    What does Kriv AI do differently from Databricks's own professional services?

    Databricks's own services teams build and tune the platform. Kriv's work is the governance layer around it: model and agent inventory mapped as in or out of SR 26-2 scope, validation appropriate to how each model or agent was actually built, and MCP tool permissioning documented the same way a traditional model's access controls would be.

    How much does a Databricks financial services AI governance engagement cost?

    Every engagement is scoped and priced after a short discovery call.

    Do you work only with banks, or with wealth managers and insurers too?

    All three. The Databricks patterns overlap heavily: Unity Catalog governance, MLflow validation, and MCP tool permissioning apply the same way whether the workload is a bank's credit model, a wealth manager's advisor tooling, or an insurer's underwriting or claims model. For how pilots are gated before they reach production on Databricks, see pilot-to-production patterns on Databricks for financial institutions.

    pilot-to-production patterns on Databricks for financial institutions

    Talk to the team that would do the work

    Bring your requirements to a working session with the person who'll actually deliver.

    Book a Discovery Call