AI Credit Underwriting: Lessons From DBS's 70-Agent Workflow

AI can reduce the assembly work behind a corporate credit memo. It should not quietly become the credit committee.

That distinction is the useful lesson in DBS’s recent agentic-AI rollout. The bank says specialized agents handle more than 70 tasks to prepare a review-grade first draft of a credit memo for corporate and credit-risk bankers. Following a 150-user pilot, DBS says it expanded the workflow to about 1,500 employees. The bank describes the result as a starting point for deeper assessment, not an automated approval decision. DBS’s announcement is worth reading closely for that boundary.

For a US finance team, the transferable idea is not “deploy 70 agents.” It is to split credit work into evidence-gathering, drafting, challenge, and approval—then keep authority with accountable people. Reporting on DBS’s broader AI program has likewise described stops and gates intended to preserve human cognition in the credit-memo workflow. Forbes reported that approach before the rollout announcement.

This article explains how to use that pattern without treating a polished memo as a reliable credit decision.

A credit memo is not a credit decision

A corporate credit memo combines financial statements, management information, collateral details, industry context, covenants, prior approvals, and exceptions. Some work is repetitive: locating a disclosed ratio, comparing a quarter with the prior period, assembling a timeline, or formatting a checklist. Some work is irreducibly judgmental: deciding whether an exception is acceptable, weighing weak evidence, interpreting a deteriorating relationship, and approving an exposure.

An AI system can help with the first category. It can extract, organize, flag, and draft. It should not be allowed to convert uncertain inputs into a final recommendation without an accountable reviewer.

That is also consistent with the NIST AI Risk Management Framework, which calls for defined and documented human-oversight processes, independent assessment, and repeatable testing and validation. A practical underwriting workflow should make those controls visible in the work product itself.

The four-stage workflow worth borrowing

The safest useful pattern is not one large prompt. It is a sequence with a clear owner and stop condition at every stage.

Stage	What AI may do	Required control	Human owner
1. Intake	Inventory supplied documents and identify missing items	Approved-source allowlist; document IDs and dates recorded	Relationship manager or analyst
2. Draft	Extract facts, calculate only with approved tools, and create a cited memo draft	Every material statement links to a source page, cell, or document ID	Credit analyst
3. Challenge	Search for contradictions, stale data, covenant gaps, and unsupported assumptions	Separate challenge prompt; unresolved items become exceptions	Independent credit-risk reviewer
4. Approval	Present the evidence, exceptions, and proposed terms	Final decision remains outside the model; approval binds to the exact facility and terms	Authorized credit approver

The key is that the challenge stage is not a cosmetic second opinion. A second model can repeat the first model’s mistake. The challenge should force a different task: compare the draft against the original evidence, search for conflicting facts, and identify what the evidence does not establish.

For broader controls on source handling, validation, and approval boundaries, see our AI financial analysis control checklist.

What an AI-assisted credit memo should contain

A reviewer should be able to distinguish source facts from model-generated language in seconds. Require a memo draft to use sections such as:

1. Evidence register: document ID, source owner, reporting period, retrieval time, and exact citation for each material figure.
2. Borrower snapshot: a plain-language description limited to supplied evidence. Missing facts stay marked as missing.
3. Financial trend table: reported revenue, profitability, leverage, liquidity, and cash-flow figures with units and periods. Calculate formulas in a controlled spreadsheet or service, not in free-form prose.
4. Risk and mitigant matrix: each risk linked to a source, a confidence level, and a proposed follow-up—not a fabricated reassurance.
5. Exceptions and open questions: covenant waivers, stale statements, data conflicts, or incomplete collateral information routed to a named owner.
6. Decision record: reviewer comments and the eventual approval decision, stored separately from the generated draft.

This design protects against a common failure mode: a memo that reads as if it were comprehensive while silently filling gaps with plausible language.

A prompt chain for drafting, challenging, and escalating

The following is a starting point for an internal, approved environment. It is not a prompt for approving, pricing, or denying credit.

Prompt 1: Build the evidence register

You are assisting a corporate credit analyst. Use only the approved documents listed below. Create an evidence register with: document ID, date, reporting period, source page or cell, extracted fact, unit, and confidence. Do not infer missing figures. If two sources conflict, preserve both values and flag the conflict. Return MISSING, CONFLICT, or OUT_OF_SCOPE where appropriate.

Prompt 2: Draft the analytical narrative

Using only the evidence register, draft a credit-memo narrative under these headings: business profile, financial performance, liquidity, leverage, key risks, mitigants, and open questions. Cite every material claim with its document ID and location. Do not recommend approval, denial, pricing, facility size, or covenant terms. State the evidence limitation whenever a conclusion depends on incomplete or conflicting data.

Prompt 3: Produce the counter-evidence review

Act as an independent credit-risk challenger. Review the draft against the evidence register. List: (1) every unsupported claim, (2) contradictory or stale source data, (3) missing downside scenarios, (4) facts that could change the credit view, and (5) questions that require a human decision. Do not rewrite the memo. Return a table with source references and severity.

Prompt 4: Prepare the reviewer packet

Create a reviewer packet containing the cited draft, counter-evidence table, unresolved exceptions, and a checklist of required human confirmations. Label this as a decision-support draft. Do not state or imply that the system approved a credit decision.

The wording matters. “Do not infer” and “do not recommend” are not magic safeguards, but they establish the correct job boundary and make violations easier to test.

Controls that should not be delegated

Some underwriting controls must remain deterministic or human-owned:

• Data entitlement: The system should access only approved borrower and internal data. Keep credentials outside the model and log each retrieval.
• Numerical validation: Recalculate totals, ratios, and covenant tests with controlled formulas. A model may explain a result; it should not be the only calculator.
• Policy rules: Eligibility, limits, authority matrices, and required documentation belong in a rules service or workflow, not in a prompt.
• Exception authority: A model can identify an exception. Only an authorized person should accept it.
• Adverse-action and customer communications: Route any customer-facing outcome through the institution’s applicable legal, compliance, and fair-lending review process.

Before connecting an agent to a live system, apply the least-privilege pattern: read-only access for research, narrow tool permissions for any workflow action, server-side validation for every argument, and an audit trail tied to the case ID. The controls in Financial Prompt Engineering 101 provide a useful baseline for defining scope, limits, and safeguards.

How to test the workflow before production

Do not validate an underwriting agent with a handful of clean historical cases. Build a test set that includes:

• complete and incomplete credit files;
• contradictory statements from different reporting periods;
• changed accounting classifications;
• missing collateral documentation;
• unusual but legitimate borrower conditions;
• stale market or industry documents; and
• documents containing misleading instructions aimed at the model.

Measure more than draft speed. Track citation accuracy, numerical reconciliation failures, unsupported-claim rate, missed exceptions, escalation rate, reviewer edits, and the percentage of outputs that correctly stop when evidence is missing. Re-run the tests whenever the model, prompt, retrieval process, data source, or credit policy changes.

Start with a narrow pilot

A sensible first pilot does not automate the highest-stakes decision. Choose one memo subsection—such as evidence indexing or prior-period variance extraction—and keep the output read-only. Define acceptance criteria before the pilot starts:

• every extracted figure has a verifiable citation;
• calculations reconcile outside the model;
• the workflow identifies known missing-data cases;
• a reviewer can reject the result without friction; and
• no output can approve, decline, price, or execute a facility.

If the system cannot pass those checks on representative cases, it is not ready for a broader role. Faster drafting is only valuable when it gives the credit team more time to challenge the evidence and make the actual decision well.

The practical takeaway

DBS’s public example is compelling because it places agentic AI in the preparation layer of credit work. That is where teams can gain time without confusing automation with authority. A strong credit workflow should make it easier to see the source, find the exception, challenge the draft, and record the human decision.

Use AI to make a reviewable first draft. Keep responsibility for the credit decision with the people who can explain, defend, and own it.

This article is for general information and operational-design discussion. It is not investment, lending, legal, compliance, or risk-management advice.