AI adoption can change a bank’s capacity to process information. It does not, by itself, tell a strategy, finance, or risk team whether credit will become more available, more competitive, or riskier. That distinction matters when someone turns a productivity claim into a lending forecast.
The San Francisco Fed’s September 2026 Economic Letter finds associations between banks’ AI-related job-posting intensity and lending patterns, returns on assets, problem-loan shares, and the share of lending to small and medium enterprises. The authors explicitly caution that those relationships are not causal. That is a strong reason to use the research as a hypothesis generator—not as an approval to loosen standards, cut spreads, or predict a portfolio outcome.
This guide provides an AI adoption and bank lending scenario notebook. It converts a broad question—“what does AI efficiency mean for our lending conditions?”—into dated observations, rival explanations, falsification triggers, and an accountable human decision. It is not an underwriting model, a pricing engine, investment research, or lending advice.
The search intent: lending conditions, not AI loan approval
Most search results for AI and lending focus on credit scoring, origination automation, or generic governance. Those answer a different question. This notebook examines the path from a bank’s operating change to possible changes in its aggregate lending posture:
- Does a documented efficiency gain increase capacity, or merely move work elsewhere?
- Is a narrower spread or changed term visible in the market, and is competition a more plausible explanation?
- Are small-business, relationship-based, or complex-credit outcomes moving in a way that contradicts the initial story?
Keep this separate from an individual loan file. Our AI credit underwriting workflow is for assembling and challenging evidence before a credit decision. Our economic exposure card is for monitoring a borrower after a facility exists. This notebook is for a bank-level strategy or risk review of lending conditions; none of the three artifacts should silently substitute for the others.
Start with a causal path, not a productivity headline
The Federal Reserve’s January 2026 Senior Loan Officer Opinion Survey distinguishes lending standards, loan terms, and loan demand, and reports competition among factors banks cite when easing some commercial and industrial loan conditions. That makes it useful as a vocabulary check: a change in demand, a changed standard, and a changed spread are not interchangeable.
Use three scenarios. Each starts with a claim that could be wrong.
| Scenario | Working mechanism | Evidence that could support review | Plausible rival explanation | Falsification trigger |
|---|---|---|---|---|
| Capacity pass-through | AI reduces a bounded operating bottleneck, creating capacity to serve eligible borrowers | Dated workflow baseline, release record, queue and exception logs, policy-stable approvals | Staffing, seasonality, or a backlog cleanup created the change | Capacity rises but approved volume, turnaround, and exception mix do not move consistently after the stated change window |
| Competitive pass-through | Savings or faster service alter the bank’s market terms or distribution response | Dated product terms, approved competitor intelligence, win/loss notes, controlled offer versions | Broad funding-cost or demand changes explain the terms better | Similar movements occur where the AI-enabled workflow is not used, or are fully explained by market-rate changes |
| Risk-and-access trade-off | Better processing of hard information changes the mix of opportunities reached or declined | Segment definitions, review samples, override records, performance-monitoring plan | Portfolio mix, geography, acquisition, or policy changes altered the mix | A segment shift appears without a traceable process or policy link, or quality checks show unavailable/inconsistent inputs |
The Fed letter is particularly helpful because it discusses the different roles of hard and soft information and reports correlations that can point in more than one direction. A scenario notebook should preserve that ambiguity. Do not rewrite an association as “AI caused more lending,” “AI reduced risk,” or “AI improved access.”
Build the evidence row before discussing an outcome
One row represents one mechanism in one defined review window. Use real internal records only where they are authorized for the review; do not place customer-level data, nonpublic pricing, or personal information into a general-purpose model.
| Field | What to record | Why it matters |
|---|---|---|
| Claim ID and owner | A stable identifier, accountable business owner, and independent reviewer | Prevents an attractive narrative from becoming ownerless |
| Process boundary | Product, segment, geography, workflow step, and the change-release date | Stops bank-wide conclusions from a narrow pilot |
| Baseline and observation window | Predefined dates, units, inclusion rules, and data lineage | Makes before/after comparisons auditable |
| Mechanism and expected sign | Capacity, competition, or risk/access path; what should rise, fall, or remain uncertain | Forces a testable proposition |
| Counter-hypothesis | At least one non-AI driver and the evidence needed to assess it | Reduces post-hoc attribution |
| Evidence state | SUPPORTED, MIXED, INSUFFICIENT, or CONTRADICTED | Keeps uncertainty visible |
| Falsification and stop condition | Result that ends the claim or requires escalation | Gives the review a safe exit |
| Decision and retention | OBSERVE, DEEPER_REVIEW, PAUSE_CLAIM, or CLOSE; approver and retention location | Keeps the final judgment human and recoverable |
For the AI system itself, use the NIST AI Risk Management Framework as a voluntary risk-management reference. Its Govern, Map, Measure, Manage framing can help assign owners, define context, measure limitations, and decide how to respond. It does not create a causal estimate or replace applicable legal, model-risk, fair-lending, privacy, or supervisory requirements.
A bounded scenario-notebook prompt
Give the model structured inputs and force it to return gaps, rival explanations, and stop conditions. The model may organize evidence; a designated reviewer must validate sources and make any operational decision.
role: "Bank strategy and risk evidence analyst"
task: "Create a bank-level lending-conditions scenario notebook. Do not make a causal claim."
inputs:
scope:
product_or_segment: "<authorized aggregate segment>"
geography: "<scope>"
review_window: "<start and end dates>"
ai_change_boundary: "<workflow, release date, controls, owner>"
observations:
- source_id: "<approved record ID or public source URL>"
observation_date: "<YYYY-MM-DD>"
metric_definition: "<unit and inclusion rule>"
value_or_summary: "<authorized aggregate fact>"
possible_non_ai_drivers:
- "funding costs"
- "borrower demand"
- "policy, staffing, acquisition, or seasonality"
required_output:
- "three scenario rows: capacity, competition, risk/access"
- "claim, mechanism, expected sign, and counter-hypothesis for each row"
- "evidence gaps and data-lineage checks"
- "one falsification trigger and one stop/escalation condition per row"
- "evidence state only: SUPPORTED, MIXED, INSUFFICIENT, or CONTRADICTED"
- "human decision recommendation only: OBSERVE, DEEPER_REVIEW, PAUSE_CLAIM, or CLOSE"
constraints:
- "Do not infer borrower-level eligibility, pricing, or credit decisions."
- "Do not invent metrics, sources, causal effects, or regulatory conclusions."
- "Label every unsupported conclusion INSUFFICIENT."
- "Treat correlations in external research as hypotheses, not proof."For a financial-model assumption that needs cell-level source discipline, use the AI financial model audit trail. For a production release that needs requirements-to-test traceability, use the separate core banking change evidence pack. The notebook links their outputs only as cited evidence; it does not merge their decision rights.
Test the notebook before a live strategy review
Run a small, controlled test set. The goal is to catch attribution errors before a narrative reaches a committee, a product team, or an external audience.
| Test case | Input condition | Expected safe behavior |
|---|---|---|
| Efficiency claim with no baseline | A vendor or manager says processing is faster but supplies no dated definition | Return INSUFFICIENT; request baseline, scope, and source owner |
| Terms move with rates | Spreads change during a broader funding-cost movement | Surface the market-rate rival explanation; do not attribute to AI |
| Segment shift after policy change | Small-business share changes after an eligibility-policy revision | Mark policy as a competing mechanism and route to deeper review |
| Mixed results | Capacity improves while exceptions or quality findings rise | Return MIXED; preserve both observations and escalation condition |
| Sensitive detail supplied | Prompt includes borrower names, account data, or nonpublic customer information | Stop output and direct the user to an approved, access-controlled workflow |
Review the output for four failure modes: a causal verb without a design that can support it, a metric without a definition, a competitor claim without a dated source, and a recommendation that crosses into individual credit or pricing authority. Any one is enough to pause the row.
Measure review quality, not a promised lending result
The useful KPIs are about the quality of the review record, not a claimed return from AI adoption.
- Source-complete row rate: share of rows with a date, source ID, metric definition, and owner.
- Counter-hypothesis coverage: share of material claims with at least one named non-AI explanation and test.
- Falsification readiness: share of rows that name a measurable contradiction or stop condition.
- Human-resolution time: elapsed time from complete evidence to a recorded
OBSERVE,DEEPER_REVIEW,PAUSE_CLAIM, orCLOSEdecision. - Post-review correction rate: share of resolved rows later reclassified because source scope, timing, or attribution was wrong.
None of these measures shows that AI improved lending, profitability, borrower access, or investment performance. They show whether the institution can make and challenge a narrow claim with traceable evidence.
Keep lending authority and customer outcomes outside the prompt
An AI-generated notebook must not set a credit standard, change a borrower price, or determine a customer’s eligibility. Keep those decisions with authorized teams and governed systems. If the work touches a lending model, customer data, or an external communication, apply the institution’s model-risk, privacy, fair-lending, records, legal, and approval controls before using any output.
The practical win is modest but valuable: a team can say what it observed, what else might explain it, and what would prove its story wrong. That is a stronger starting point than turning an AI-adoption headline into a lending forecast.
Sources
- Federal Reserve Bank of San Francisco: How AI Adoption Might Affect Bank Lending, September 21, 2026.
- Federal Reserve Board: January 2026 Senior Loan Officer Opinion Survey on Bank Lending Practices.
- NIST AI Risk Management Framework.
About Enis
AI Engineer specializing in Machine Learning and LLMs. Combining Computer Engineering and Economics to build data-driven financial tools.