AI for Underwriting in Banking
The analyst has the statutory accounts open in one window and the management figures in another, and the two do not agree. The director's loan account has moved, a property has been revalued, and the auditor's note explains one of those but not the other. Further down the file is an email from the relationship manager saying the owner already covered all this on a call. The credit paper is due this week, committee reads the pack the evening before, and the spreading is not finished.
Credit underwriting in a bank is rarely held up by the decision. It is held up by assembly: getting numbers from several sources into one comparable shape, deciding which adjustments are real, evidencing what the grade rests on, and writing it down so a committee can follow the reasoning.
One boundary governs everything below. A well built agent assembles, drafts, evidences and flags. It does not sanction, does not set the risk grade and does not price. Those sit with mandate holders under your delegated authority.
Spreading the accounts into a shape you can compare
A commercial file arrives as a pile of formats: filed accounts as a scanned document, management figures exported from whichever package the borrower runs, an aged debtor listing, bank statements, and a forecast nobody else can follow.
An agent reads all of it and spreads it into your standard template, with the same line items and the same treatment of leases and intercompany balances for every borrower. Where a source is ambiguous, say one line covering both sales and other income, it maps to the line your policy specifies and records the ambiguity rather than choosing silently. Every cell points back to the page it came from, so the reviewer checks exceptions instead of rekeying.
Adjustments and add backs are judgement, not arithmetic
Spreading is mechanical. Normalising is not. Owner remuneration, a one off legal settlement, a disposal gain, related party rent set at something other than market: each changes the earnings figure your policy tests against, and each is arguable.
An agent proposes here and never decides. It surfaces candidate adjustments with the evidence attached, the note in the accounts, the prior treatment on the same borrower, and the policy rule it believes applies. Across the book, an add back treated one way in one region and differently in another becomes visible, which is a portfolio finding.
What the borrower says against what the accounts show
Every file carries a narrative: call notes, the borrower's commentary, the business plan. It is often the most useful material in the pack and almost never tested line by line against the numbers.
An agent holds the two side by side. The plan attributes margin improvement to a new contract, while the figures show the improvement arriving before that contract started. The commentary calls the overdraft peak seasonal, while the statements show it peaking at a different point each year. None of this is a decision. It is a list of questions for the relationship manager to answer before committee asks them, which is the value of asking early.
Serviceability and stress cases that do not depend on who built the model
Serviceability is the heart of the file and where files diverge most. One analyst tests coverage on adjusted earnings, another on operating cash flow, a third includes hire purchase obligations the second left out. All three may be defensible. Read side by side, they are not comparable.
An agent computes the coverage measures your policy defines, from the spread it has already built, and shows the workings. It then runs the required stress cases, the same ones in the same order on every file: a rate rise, a revenue fall, a margin squeeze, the loss of the largest customer. A committee seeing identical stresses on every paper compares risks instead of arguing about method. The same holds at much higher volume in consumer lending, where an affordability assessment should give the same answer whoever ran it.
Risk grading inputs, assembled and evidenced
Most banks do not want an agent choosing a grade, and they are right. What slows grading is not the model but the inputs: ratios, the sector view, the management assessment, payment behaviour, group structure, and the qualitative overlays somebody must justify in prose.
An agent assembles those, populates the scorecard inputs from the spread and the servicing data, and flags where an input is missing, stale or contradicted elsewhere. Explainability here is practical rather than philosophical: if the agent cannot show which document and which figure produced an input, that input should not reach the scorecard.
Security pledged against security perfected
A credit paper says what security is taken. The recovery position depends on what is actually perfected, a different question and usually asked too late.
An agent reads the security schedule, the valuations, the charge documents and the registry evidence, then reports the gaps: a debenture agreed but never registered, a guarantee signed by only some of the named parties, a valuation prepared on a basis your policy does not accept. It also flags the quiet erosion nobody is assigned to watch, such as valuations ageing past your limit, insurance lapsing on charged property, or an asset sold while still carried as security.
Writing the credit paper so the reasoning is visible
The credit paper exists so a committee can follow how a recommendation was reached. Most of it is assembly: borrower background, group structure, the facility requested, the financial summary, security, covenants, policy exceptions.
An agent drafts everything except the recommendation, pulling each section from evidence it has already assembled and citing a source for every figure. The analyst then spends their time on the argument, not on reformatting tables. Papers arrive in the same structure, and the sections people skip when tired get written properly because the agent does not get bored.
Sanction conditions and the conditions precedent nobody tracks
Approval is not the end of the file. Sanction carries conditions: documents before drawdown, undertakings after it, security to be perfected inside a window. In many banks these live in free text inside the approval, which means they live nowhere.
An agent extracts every condition into a tracked list with an owner, a due point and a stated evidence requirement, watches for the evidence arriving, matches it against what was asked for, and reports what is outstanding by facility rather than by whoever remembers. This is unglamorous and often the largest single gap an audit finds in a lending book.
Annual review and the signals already in your servicing data
Annual review is where credit teams lose most of their capacity, and the work is largely repetition. An agent produces the first draft by updating the previous paper against new evidence: a fresh spread, changes in group structure, movements in security, conduct over the period, conditions still outstanding. Anything unchanged is marked as unchanged rather than repeated silently, which is the failure mode of manual rollover.
Between reviews, what predicts trouble is usually already inside the bank, spread across systems nobody reads together. Credit turnover falling while the overdraft sits at its limit. Payments to suppliers stretching. Returned direct debits. A filing deadline missed at the registry. An agent can watch these across the portfolio and produce a short ranked list with the evidence behind every flag, so a human reads the right file while options remain.
The declined application and the explanation owed
Declines get the least attention and carry real risk. An applicant is entitled to know why the answer was no, in terms they can follow, and the file must show that the reason given was the reason applied.
An agent drafts that explanation from the decision record, in plain language, naming the specific policy test that was not met rather than a generic category. It checks the drafted reason against the evidence and flags where the two do not line up, which is precisely the mismatch a complaint will find. Across the book it can also surface declines clustering in a way your policy never intended.
Where AI fails here
Bad spreading becomes fluent prose. A misread management account, a comparative column mapped to the wrong period, a currency assumption nobody stated, and the agent produces a credit paper that reads beautifully around the wrong numbers. Tidy output persuades more than messy output, which is the danger.
It has no scepticism. Management figures are unaudited, sometimes optimistic and occasionally dishonest, and an agent normalises whatever it is given. Borrower fraud in smaller business lending often looks like unusually neat numbers arriving on time, the opposite of what an automated reader flags.
Anchoring is real and hard to design out. Once the agent proposes a set of add backs, reviewers argue with the proposal instead of forming their own view. Judging which earnings are sustainable is where lending is won and lost, and it degrades once it becomes a review task.
Unusual files are where it is weakest and most confident. A young business, a complex group, a restructuring, a sector turning for reasons that have not reached the accounts yet: these need judgement most and resemble your existing book least.
Missing evidence reads as clean evidence. If a registry search returns nothing, if a valuation was never uploaded, if a guarantee sits unsigned in a drawer, the file reads as though there is no issue rather than as though nothing was found.
It also inherits your past inconsistencies, reproducing how you have been treating a sector, a region or a customer type, including the parts you would not defend if you looked deliberately. And none of it moves accountability: sanction, grading, pricing and the treatment of a borrower in difficulty stay with named people under your governance.
Which part of the credit workflow to hand over first
Start where volume is high, evidence is good and a human already reviews the output: spreading and proposed normalisations on annual reviews.
Then test it honestly. Run the agent across files you have already sanctioned and compare its spread, its proposed adjustments and its draft paper with what your analysts produced at the time. Where the two differ, ask which was right. That tells you more than any vendor demonstration.
Expect model risk and audit to want the intended use written narrowly, the configuration change controlled, and every input and output retained so a paper can be reconstructed. Scope is your strongest argument here.
If you want a plain view of which parts of your credit underwriting workflow an agent could carry and which it should not touch, describe the process to Adopt AI Partners and we will tell you straight.