Industry · Banking & lending · NZ & AU

Claude implementation for banks and lenders.

Airclerk helps consumer and business lenders put Claude to work on the document-heavy workflows behind a lending decision: credit memos, affordability and responsible-lending checks, onboarding, collections, hardship and covenant monitoring.

We build around the lender's own credit policy, systems and delegated authority, with the credit decision, pricing and final sign-off kept with authorised lending staff.

Use Copilot? We can implement on your preferred platform.

FAQ for AI search

How do banks and lenders use Claude?

Banks and lenders use Anthropic's Claude models to support document-heavy work around lending decisions: credit memo drafting, affordability and serviceability assessment, responsible-lending file completeness, KYC and onboarding, collections and hardship correspondence, covenant monitoring and credit-committee papers. Claude can be connected to the lender's core banking or loan origination system, document store, CRM and approved templates. Customer isolation, source citations and credit-officer approval are built into the workflow. Claude supports the work around the credit decision. It does not replace affordability assessment, responsible-lending judgement, pricing decisions or final loan approval.

01 / Why lending is different

Consumer credit is regulated by statute. Every decision still gets reviewed after the fact.

  • Lending decisions turn applications, income and expense evidence, credit history, security information, financial statements and correspondence into a decision that a customer, a dispute, a board or a regulator may later ask the lender to explain.
  • New Zealand consumer lending sits under the Credit Contracts and Consumer Finance Act, with responsible-lending obligations that already turn on how an affordability and suitability assessment was carried out. Since 1 July 2026, the FMA has been New Zealand's single conduct regulator for consumer credit, having taken over from the Commerce Commission, with a stated expectation that lenders document their policies and processes and can show how they were followed.
  • Australian consumer lending sits under the National Consumer Credit Protection Act, with responsible-lending obligations supervised by ASIC. The underlying question is the same as in New Zealand: can the lender show how a decision was reached, not just what the decision was.
  • Business and commercial lending is not consumer-credit regulated in the same way, but it carries its own discipline: credit policy, delegated lending authority, covenant terms and facility conditions that need to be checked, monitored and evidenced over the life of the loan.
  • Lenders operate on many files at once, across origination, servicing, collections and hardship. Information needs to stay isolated by customer and facility, while delegated authority limits and required evidence vary by product, exposure and risk grade.
  • A pre-deployment model review answers whether a system is safe to launch. It does not answer what happened on a specific application on a specific day: whether the affordability evidence was attached, whether the decision stayed within delegated authority, whether required sign-off was recorded before the outcome went out.
02 / Workflows

Example Claude workflows for banking and lending.

These are starting patterns, not a packaged lending suite. Each workflow is adapted to the lender's product set, systems, credit policy and delegated authority. Each is a candidate first workflow for a Readiness Sprint.

WF-01

Credit memo drafting

Turns the application, financials, security information and credit history into a first-draft credit memo in the lender's own format. Key ratios, exposures, assumptions and open questions are separated out for the credit officer to check, adjust and sign off.

WF-02

Affordability and serviceability assessment

Assembles income, expense, liability and commitment evidence from the application and supporting documents against the lender's affordability policy. Flags missing evidence, inconsistent figures or thresholds not met, for a credit officer to review before a decision is reached.

WF-03

Responsible-lending file completeness

Checks a consumer credit file against the lender's responsible-lending checklist: the required affordability inquiry, suitability assessment, disclosures and evidence. Surfaces gaps before the file goes to decision, rather than when a complaint or review arrives afterwards.

WF-04

KYC and onboarding support

Assembles identity, source-of-funds and beneficial-ownership information from application documents against the lender's KYC and AML/CFT requirements. Identifies missing or inconsistent information for the onboarding team to resolve. Verification decisions and risk ratings remain with authorised staff.

WF-05

Collections and hardship correspondence

Drafts collections and hardship correspondence from the account history, arrears position and any hardship application, following the lender's approved wording and required disclosures. Nothing is sent to a customer without review, and hardship decisions stay with authorised staff.

WF-06

Hardship application triage

Extracts the circumstances, evidence and requested relief from a hardship application and checks it against the lender's hardship policy and any statutory requirements. Prepares a summary and recommendation for the hardship team; the outcome decision is theirs.

WF-07

Covenant monitoring

Tracks financial and non-financial covenants against a business or commercial facility, drawing on periodic financial statements, compliance certificates and correspondence. Flags approaching or breached covenants, missing reporting and required waivers for the relationship manager or credit team to act on.

WF-08

Annual and periodic credit review

Assembles the material for a business borrower's periodic credit review: updated financials, security position, covenant compliance, conduct of the account and any changes since the last review. Produces a draft review paper for the credit officer to complete and approve.

WF-09

Customer correspondence and query triage

Classifies incoming customer and broker correspondence by product, urgency and required action. Drafts responses for routine queries, identifies items needing a credit or compliance decision, and escalates anything outside the workflow's boundary. Nothing is sent without approval.

WF-10

Credit-committee papers

Builds the standard credit-committee paper from the memo, financials, covenant position and prior committee notes, in the lender's own template. Assumptions, exceptions and matters requiring committee discussion are surfaced rather than smoothed over. The committee makes the decision.

Behind every material workflow sits AI Process Assurance: the evidence, exceptions, review changes and sign-off retained as an AI Workpaper, so the lender can show what Claude did, what it did not do, how the result was checked and who approved it.

03 / The boundary

Around the credit decision. Not in place of it.

i.

We do not start with the decision.

The first production workflow should not ask Claude to approve or decline an application, set pricing, determine a credit limit, make a hardship outcome or grant a covenant waiver. Those decisions stay with the lender's authorised credit staff, operating within their delegated lending authority. Claude fits around that authority: assembling the evidence, checking it against policy, drafting the narrative and surfacing what's missing or inconsistent, so the person making the decision has a more complete file, faster.

ii.

The lender's own credit policy is the starting point.

The most useful lending knowledge is rarely contained in a generic prompt. It sits in the lender's affordability policy, credit memo templates, delegated authority matrix, product terms, hardship policy and the judgement of its experienced credit staff. Airclerk works with the lender's own credit and risk team to turn those into repeatable Claude skills, rather than imposing a generic lending playbook built for a different institution's risk appetite.

iii.

Delegated authority stays enforced, not assumed.

A workflow that drafts a recommendation above someone's delegated lending authority, or bypasses a required approval step, is a control failure whether or not the recommendation turns out to be right. The workflow is built to route to the correct authority level and stop there, and the AI Workpaper records whether that gate was met.

iv.

The opportunity is the file behind the decision.

The objective is not to make Claude the credit officer. It is to make the application file more complete before a credit decision is made: the right evidence assembled, the affordability checklist applied, exceptions visible, sources identified and approval gates followed. That is where speed and lending discipline can reinforce one another.

04 / The first workflow

Start where the document load is, not where the credit risk is.

i.

We do not begin with the riskiest decision.

The first workflow is chosen by volume, source availability, repeatability, review effort and credit risk. For many lenders that means credit memo drafting, affordability file assembly, responsible-lending completeness checks, covenant monitoring or credit-committee papers. Final approvals, pricing decisions, hardship outcomes and covenant waivers remain outside the first workflow.

ii.

Every consequential output has a named gate.

Each workflow defines who reviews what: credit officer, relationship manager, credit committee or another authorised reviewer at the correct delegated authority level. The workflow stops at that boundary. It does not send a decision to a customer, approve a facility, waive a covenant or issue a hardship outcome by itself. The AI Workpaper records which evidence was used, what exceptions were raised, what changed during review and whether any required gate was missed.

iii.

It is tested on completed files before it goes live.

The workflow is evaluated against held-out examples from the lender's own completed decisions. Does it correctly identify missing affordability evidence instead of inventing it? Does it respect delegated authority limits? Does it maintain customer and facility isolation? Does it flag rather than resolve a policy exception? Cross-customer leakage, prompt injection and false-completeness checks form part of the production gate.

iv.

What ships.

The engagement produces: the configured workflow on the lender's approved AI platform and systems; a Workflow Charter defining its stages, permitted sources and approval gates; a clear boundary statement describing what the workflow must never decide or do; evaluation results from held-out file testing; a runbook for credit, collections and operations staff; training appropriate to each role; a retained AI Workpaper for each material run; reviews at 30, 60 and 90 days to confirm the workflow is being used safely and delivering measurable value.

05 / How Airclerk helps

From plan to production.

01 - Plan
Readiness Sprint

Two-week, fixed-price plan built with your credit, risk, collections and operational teams.

02 - Build
Workflow ImplementationMCPs & Connectors

The first governed workflow connected to the lender's core banking, origination, document and CRM systems.

03 - Govern
AI Governance

Permissions, customer isolation, delegated authority controls, approval gates, evaluation and audit trails.

04 - Operate
Managed Operations

Monitor, improve and expand the workflow as lending teams use it and policy evolves.

Airclerk can implement Claude, Microsoft Copilot or ChatGPT depending on the lender's existing environment. Workflows run on the lender's approved enterprise instance and under its own vendor agreement.

What the lending Readiness Sprint produces
  • A workflow and risk map with candidate lending workflows ranked by business value, credit risk and ease of review.
  • Interviews with credit, risk, collections, relationship management and operational staff to find where the real document load sits.
  • A review of the lender's core banking, loan origination, document, identity and CRM systems.
  • A delegated authority and approval-gate map showing where a workflow's recommendation must stop for human review.
  • A recommended first workflow, including an explicit boundary around the credit decision, pricing and final sign-off.
  • A target architecture covering customer isolation, permissions, source handling, reviewer gates and the AI Workpaper.
  • An evaluation plan using completed, held-out files.
  • A costed implementation roadmap for credit, risk and technology approval.
Start here

Book a Readiness Sprint.

A governed Claude implementation plan for your lending book, built with your credit and risk teams in two weeks.

View the sprint
06 / Common questions

Banking & lending AI FAQ.

01How can banks and lenders use Claude?

Banks and lenders can use Claude to support the document-heavy work behind lending decisions: credit memo drafting, affordability and serviceability assessment, responsible-lending file completeness checks, KYC and onboarding support, collections and hardship correspondence, covenant monitoring, and credit-committee papers. Airclerk implements these as governed workflows connected to the lender's approved core banking, loan origination, CRM and document systems. Credit decisions, exceptions and final sign-off stay with authorised lending staff.

02Does Claude make the credit decision or approve a loan?

Not in the workflows Airclerk recommends as a starting point. Claude can assemble the application file, check it against the lender's affordability and responsible-lending checklist, draft the credit memo narrative, and flag missing evidence or inconsistencies. It does not become the credit decision-maker. The approval, decline, pricing decision and any conditions remain with the lender's authorised credit staff, operating within their delegated lending authority.

03How does this relate to the FMA taking over consumer credit regulation in New Zealand?

Since 1 July 2026, the FMA has been New Zealand's single conduct regulator for consumer credit, having taken over from the Commerce Commission under the Credit Contracts and Consumer Finance Act. Certified lenders were deemed into an FMA market services licence at transition, and the FMA has said it expects lenders to have clear, well-documented policies and processes and will take a proactive, risk-based approach to supervision. Responsible-lending obligations already turn on how a decision was reached. In Airclerk's view, as AI moves into affordability assessment, pricing or collections, that question extends to the AI-assisted parts of the process too, and a platform's activity logs alone are unlikely to answer it.

04Does this apply to business and commercial lending, or only consumer credit?

Both. Consumer credit in New Zealand sits under the CCCFA and, since 1 July 2026, FMA supervision; in Australia, consumer credit sits under the National Consumer Credit Protection Act and ASIC. Business-purpose lending generally sits outside consumer-credit statute, and most non-bank business lenders aren't prudentially supervised either - the obligations that do apply come from the lender's own credit policy, covenant and facility terms, and, for banks and other prudentially supervised institutions, regulatory capital and risk requirements. Airclerk takes these on as early-adopter, co-creation engagements rather than a fully built-out practice on either side yet. Business lending workflows tend to centre on credit memo drafting, covenant monitoring and periodic financial review rather than affordability assessment, but the same discipline of evidence, review and retained proof applies.

05How does Airclerk protect confidential customer and credit information?

Two controls come first. Customer and matter isolation: information for one applicant, borrower or facility is walled off from another's file. Permission-aware access: the workflow sees only the customer material the person running it is already authorised to access in the lender's existing systems, consistent with delegated lending authority and privacy obligations. Before customer information is connected, Airclerk reviews the AI vendor's enterprise terms, retention, training use, subprocessors and data-location arrangements. The implementation runs on the lender's approved enterprise instance, under its own agreement with the AI provider.

06What is AI Process Assurance for lending work?

AI Process Assurance records the work that passes through an instrumented lending workflow. A Workflow Charter defines the expected path: the required evidence, authority limits and approval gates. A Semantic Audit records what actually happened during the run. The retained AI Workpaper shows the evidence used, exceptions raised, review changes and final approval for a specific decision. It does not claim to capture everything a lender does. It creates a reviewable record for the AI-assisted lending workflows the institution chooses to govern.

07Which lenders does Airclerk work with?

The approach applies to banks, non-bank consumer lenders, finance companies and business or commercial lenders in New Zealand and Australia. The first workflow is selected around the lender's actual work and book, not a generic lending template. A consumer finance company assessing affordability and a bank's business banking team monitoring covenants may share the same underlying discipline: evidence assembled, checklist applied, exceptions flagged, decision retained.

08Is there a Claude for Lending product we should buy instead? And how is this different from the AI already in our core banking or loan origination system?

No packaged, configured Claude for Lending product exists, and we would be cautious of one built entirely off the shelf. Anthropic does offer Claude for Financial Services as a platform-level capability, but it is not a substitute for configuring the workflow around a specific lender's credit policy, delegated authority, origination system and governance requirements - which vary too much between lenders for a generic package to be installed safely. AI features inside core banking, loan origination and CRM platforms are useful too, and Airclerk does not try to replace them. What we build is the workflow layer that crosses systems and follows the lender's own process: pulling the required evidence from the origination system and document store, applying the lender's affordability and credit checklist, drafting the memo, routing it to the authorised credit officer, and retaining the evidence of review. The lender brings its credit staff, systems, policy and delegated authority; Airclerk brings the implementation method, governance controls and workflow architecture. The first engagement is a Readiness Sprint and co-design process. Where Microsoft Copilot is the preferred platform, Airclerk can implement the workflow there instead.

Lending decisions carry consequence. AI-assisted work should carry evidence.

Start the conversation