How to spread financial statements for business lending with an AI agent

Have the agent extract every line, map it to your template and prove the spread ties back to the accounts. It reads the notes for items that change the picture, proposes any normalising adjustments separately, and calculates the ratios. A credit analyst approves each adjustment and signs off the spread. Nothing moves into a credit memo until the tie-out passes.

What is spreading, and who owns it?

Spreading means entering a borrower's financial statements into the bank's standard format, so that different businesses, with different accounting layouts, can be compared on the same lines and the same ratios. A revenue line in one set of accounts and a turnover line in another must land in the same row. So must a director's loan, whether it is called "other creditors" or "amounts due to related parties".

Credit analysts usually own spreading, or a central spreading team does it for them. Credit risk owns the template and the mapping rules. It is slow, careful work, and errors flow into every ratio and every lending decision that follows. The steps below show where an agent can take the load and where a person must look.

What are the steps?

1. Agent, depth 1: identifies the statements. Entity, period, whether audited, management or draft, accounting basis, currency, and units. Units are where many spreading errors start.
2. Agent, depth 1: extracts every line. Income statement, balance sheet, cash flow statement, and the notes, each line cited to its page.
3. Agent, depth 1: maps lines to the template. Using your mapping rules. Lines the rules do not cover clearly are listed, not guessed.
4. Agent, depth 1: ties out. The spread's totals agree to the statements, the balance sheet balances, subtotals add up, and the cash flow reconciles to the change in cash.
5. Agent, depth 1: reads the notes. Leases, contingent liabilities, related party balances, post balance sheet events, going concern language, restatements of prior periods.
6. Agent, depth 2: proposes adjustments. Non-recurring items, owners' remuneration normalised to market rates, one-off gains. Each adjustment is separate from the reported spread, with its reason.
7. Agent, depth 1: calculates ratios. On the reported spread and, separately, on the adjusted one.
8. Person, credit analyst: reviews and signs off. Approves or rejects each adjustment, resolves unmapped lines, and signs the spread.

Which controls apply, and what evidence does each leave?

Tie-out as a gate. A spread that does not tie out cannot be signed. Evidence: the tie-out results filed with the spread.

Mapping rules versioned. The rule set used is recorded on the spread. When credit risk changes a mapping, you can find every spread built under the old one.

Reported and adjusted kept apart. Adjustments never overwrite reported figures. The analyst and the approver see both, and the memo states which ratios are adjusted.

Citations on every line. Any figure can be traced to its page. Evidence: the spread itself.

Provenance. Audited accounts are declared by the borrower and audited. Management accounts are declared and unaudited. The spread shows which, so nobody treats draft numbers as final. Whether an agent can prepare financial statements is a different question; here it reads them, never writes them.

What should you measure?

• Spreads passing tie-out first time.
• Analyst corrections per spread, by template line. Shows which mappings are weak.
• Adjustments rejected by analysts. Many rejections suggest the agent is too willing to normalise.
• Turnaround from statements received to spread signed.
• Spreads found wrong after sign-off, and where the error started.

Where does it usually go wrong?

Units. Accounts in thousands spread as units, or the other way round. The ratios still look plausible, which is why the error survives. Record units explicitly and test them against known figures such as share capital.

Restated prior periods. The new accounts restate the prior period's figures. The agent must use the restated comparatives and flag the restatement, not keep the old spread.

Consolidated versus company-only. Group accounts often contain both. Lending to a subsidiary while spreading the group's figures overstates the borrower's strength.

Odd-length periods. A change of accounting date produces a period longer or shorter than twelve months. Growth rates and cover ratios mislead unless the agent flags the period length, and any annualising is agreed with the analyst.

Hidden debt. Director loans sitting in "other creditors", leases in the notes, invoice finance netted off receivables. The notes step exists for these.

Generous adjustments. Every year's "one-off" costs are still costs if they recur. The agent should show how often an item has appeared before calling it non-recurring.

For sole traders and small partnerships, cash flow often comes from bank statements rather than accounts, and verifying income from bank statements takes over. Some of those borrowers are natural persons, and creditworthiness assessment of natural persons is listed as high-risk in Annex III of the EU AI Act. Check which borrowers fall inside before you treat spreading as purely commercial work. The general bounds for bank agents are in AI agents in banking.