Learning point 1
What is agentic AI and how is it different from what most banks have deployed so far?
Agentic AI refers to AI systems that can take autonomous sequences of actions to achieve a goal - planning, executing, and adapting without requiring human input at each step. Most banks have deployed non-agentic AI: predictive models (credit scorecards, propensity models) that take inputs and produce a score, or chatbots that answer questions. Agentic AI goes further: it acts, not just advises.
The distinction matters because it changes the value proposition dramatically. A predictive credit model tells your analyst this applicant has a 72% probability of default. An agentic credit system analyzes the application, pulls the bureau report, reads the bank statement, checks policy eligibility, calculates the recommended limit, writes the credit note, and submits it to the sanctioning authority - with the analyst only reviewing and approving the final recommendation. The agent does the work; the model only provides a score.
Agentic AI also introduces new risks that predictive models do not - specifically, the risk of the agent taking wrong actions, not just making wrong predictions. This is why evaluation, guardrails, and HITL design are so much more important for agentic systems than for traditional ML models, and why the regulatory and risk governance framework must be more robust.
In the US, agentic AI sits in a regulatory grey zone - SR 26-2 explicitly excluded it from scope in April 2026, meaning banks deploying agentic systems must self-govern using SR 26-2's principles applied by analogy until separate agency guidance is issued.
Learning point 2
How should an NBFC think about building an internal AI capability versus buying from a vendor?
The build-versus-buy decision for AI agents in lending hinges on four factors: time to value, total cost of ownership, domain expertise depth, and strategic differentiation. For most lenders, the honest answer is: buy the platform and domain-specific agent architecture, but invest in building internal AI literacy and customization capability.
Building AI agents from scratch requires a team with expertise in LLM engineering, prompt design, agent architecture, evaluation methodology, integration engineering, and MLOps. This team costs $1.5M-$3M/year in the US market for senior AI engineering talent. Experienced senior AI engineers are expensive in the US market, and the team must also understand model risk, fair lending, adverse action, bank integrations, and production operations. For a lender whose core competence is credit and collections - not AI engineering - this is a significant and slow-to-produce investment.
Buying from a specialized vertical AI vendor like LendingIQ gives you agents that have already been designed for lending workflows, calibrated against US tri-bureau data, tested against SR 26-2 model documentation expectations and ECOA adverse action requirements, and integrated with US core banking systems including FIS, Jack Henry, Fiserv, nCino, Blend, and Encompass. You go to production in weeks instead of years, and the vendor's continuous improvement benefits your deployment automatically.
The strategic case for some internal capability remains: your institution's credit philosophy, risk appetite, and customer segment create nuances that no vendor can fully anticipate. The optimal model is a vendor-provided foundation with institutional customization layers - you configure the agent's credit policy parameters, escalation thresholds, and product-specific rules, while the vendor maintains the underlying engineering.
Learning point 3
What will AI agents be able to do in banking five years from now that they cannot do today?
Over the next five years, AI agents in lending will evolve across three dimensions: capability depth, integration breadth, and trust.
On capability depth, agents will develop reliable long-term reasoning - the ability to analyze a borrower's complete credit history across multiple institutions, identify multi-year patterns, and produce credit assessments that consider the economic cycle context of individual decisions. Current agents are limited to the data in a single session; future agents will operate across longitudinal data architectures.
On integration breadth, the CFPB 1033 open banking rule will give US borrowers consented access to their financial data across institutions - combined with Plaid/MX/Finicity infrastructure, this will transform credit assessment from a point-in-time bureau pull to a continuous, cash-flow-aware evaluation.
On trust, as SR 26-2-style risk-based supervision matures and agency-specific AI guidance is issued, expected across 2026-2028, regulator comfort with autonomous credit AI will increase incrementally - likely starting with prime borrowers and bureau-rich segments before extending to thin-file lending.
Spanish-language lending for the Hispanic market - agents fluent in Spanish-language documents including income certifications, ITINs, and foreign asset declarations - can expand credit access to a roughly $2T addressable underserved market.