AI agents in lending prepare the loan file so a credit officer can decide faster: they sort and read documents, analyse bank statements and tax returns, spread financials, draft the appraisal note or credit memo, check the proposal against policy, and price the risk, draft the sanction letter, track conditions after sanction and prepare the disbursal. The officer still sanctions, approves deviations, accepts valuations and releases funds. The gain is in preparation time and consistency, not in taking the decision away from people.
Where the time goes on a lending desk
Ask a credit team where its hours go and the answer is rarely the decision itself. It is chasing missing documents, re-keying figures from statements into a template, writing the same sections of a note for every file, checking each proposal against a policy that runs to dozens of pages, and following up conditions after sanction. Each step is careful work. Most of it is preparation.
That is the work agents are suited to. They read quickly, show their workings and do not get bored on the fortieth file of the week. They are also easy to check, because every figure they produce can carry the formula and the source behind it.
Retail and MSME loans: from application to disbursement
On a retail and MSME desk, the work runs as a chain. In Tuvel’s retail credit operations team, each link has its own agent:
- Loan File Orchestrator runs each file from application to disbursement and routes every decision to the right authority.
- Document Intake and QC sorts and reads KYC, income, business and property documents, and drafts the list of what is missing.
- Cash Flow and Statement Analyser reads statements, GST and tax returns, cross-checks income and flags risky patterns, with workings shown.
- Credit Appraisal Note Drafter drafts the appraisal note from the file, the scorecard and the policy, with deviations listed.
- Policy and Deviation Checker checks each proposal against credit policy and pricing, and names who may approve each deviation.
- Legal and Technical Report Reviewer reads legal and valuation reports against the bank’s norms and drafts the queries that close each gap.
- Sanction to Disbursement Tracker and Post-disbursement Tracker follow conditions, deferred documents and insurance, and draft the chasers.
- Disbursal Execution Preparer prepares the disbursal instruction once a person has cleared every condition, and puts it on an approval card. It posts only the exact instruction the disbursement authority approved.
The orchestrator matters as much as the specialists. It knows where each file stands, hands the next step to the right agent, and makes sure nothing reaches a credit officer half-prepared.
Corporate and mid-market credit: the credit memo
Corporate credit is slower and more bespoke, and the memo is the heart of it. The corporate and MSME credit team splits the memo into the parts analysts already recognise:
- Financial Spreading spreads statements into the bank’s template and computes the ratios, with formulas shown.
- Borrower Research and News builds a cited profile: group structure, promoters, filings, ratings and adverse news.
- Corporate Credit Memo Drafter drafts the memo for working capital and term loans, with proposed terms marked for decision.
- MSME GST Limit Analyst sizes working-capital limits from GST returns and bank statements, with the calculation shown.
- Loan Document and Sanction Terms Checker checks loan and security documents against the sanction letter and lists every exception.
- Covenant and Early Warning Monitor recomputes covenants and drawing power from periodic submissions and raises early-warning alerts.
- Credit Scenario and Pricing runs base and stress scenarios, scores the borrower on the bank’s own model and proposes risk-adjusted pricing on the bank’s grid.
- Sanction Letter and Agreement Drafter drafts the sanction letter and the loan and security agreements from the bank’s templates, and flags each non-standard clause for legal.
- Collateral Registration Tracker keeps the register of security creation, registration and perfection, and drafts the loan system update for a maker.
- Early Warning Orchestrator combines covenant signals, account conduct, market data and news into one risk score per borrower, and proposes mitigation for the monitoring officer.
- Corporate Workout Analyst sets out restructuring options for a stressed account within the bank’s policy, with recovery scenarios and draft amendments.
A good memo agent does not invent a view. It assembles the evidence, sets out the terms the analyst proposed, marks every figure with where it came from, and leaves the recommendation and the terms for people to settle.
What stays with people
On Tuvel’s lending desk, these decisions always stay with a person, approved on a card in the desk’s room:
- Sanctioning, declining or renewing a limit.
- Approving a deviation from policy.
- Approving a rating or a price.
- Clearing a title or accepting a valuation.
- Waiving a document or a condition.
- Issuing a sanction letter.
- Releasing funds, on the exact instruction.
- Setting an account’s early-warning status.
- Agreeing a restructuring.
Lending agents start at level 1 or 2. At level 1 an agent reads and reports, and anything it would write waits for a person. At level 2 it drafts on its own, and changes to records other people rely on still wait. A person decides each step up. We explain the levels in how banks govern AI agents.
What makes a lending agent trustworthy
- Workings shown. Every ratio, limit and income figure carries its formula and source document, so a reviewer checks the method, not just the answer.
- Money maths in code. Calculations are done by ordinary code. The language model explains the numbers; it does not produce them.
- Policy as a source. The bank’s credit policy, pricing grid and delegation of powers are loaded into the knowledge base every agent reads before it answers.
- Granted access. Each agent sees only the systems and customer details it was granted, through connectors scoped to that agent.
- A record per file. Every step, hand-off and approval is recorded, so audit can replay how a file reached its decision.
How it connects to your loan system
Agents sit alongside your loan origination system, not in place of it. Your LOS, core banking, bureau and document systems connect through an OpenAPI description or an MCP server, granted to one agent at a time and revoked in one click. Connectors are built for each bank as part of its pilot Per deployment. See integrations.
What others offer
Lending is where many vendors start. nCino offers an Analyst agent for credit and underwriting, Oracle has pre-built agents for corporate credit and lending, and TCS BaNCS AI Compass includes credit underwriting agents. Our buyer’s list describes each from its own announcements. The difference to look for is whether lending agents share one control plane with the rest of the bank’s agents, or arrive as a separate island.
Starting a lending pilot
A typical first pilot takes one product line, such as MSME working capital or home loans, and runs the agents on recent files side by side with the team. Measures are simple: time per file, share of notes accepted without edits, and every action within its level on the record. Six to eight weeks is enough to see the pattern. See how a pilot works.