AI for Financial Reporting in Banking

By the time product control opens the daily explain, the risk system has published its own view of the same portfolio, and the two disagree by an amount too large to ignore and too small for anyone to have an answer ready. Somebody will spend the morning finding out why, and the answer will be definitional: two systems valuing on different bases, at different moments, over populations that do not quite match. Neither number was wrong. The explanation simply had to be rebuilt from nothing, as it was last month and will be next month.

Producing numbers is not the constraint in a bank. The constraint is explaining why two true numbers differ, on a timetable, to people who will ask again next cycle.

One boundary governs everything below. A properly built agent does not set an accounting judgement, does not post an unreviewed journal and does not sign a return. It prepares, reconciles, drafts and evidences.

The close is a relay, not a task list

The close is a sequence of handoffs. Sub ledgers and product systems feed the general ledger on a timetable, each downstream team sits idle until the upstream one finishes, and the same feed is late most cycles.

An agent watches the relay rather than the tasks. It checks each expected feed against its arrival window, reconciles delivered control totals to what the source says it sent, and tells the downstream team what has landed and what is missing. The quiet output is the pattern: which handoff slips, by how much, and which team absorbs the delay.

One balance sheet, several true versions of it

A bank does not hold one set of numbers. It holds a general ledger, the product systems that fed it, a risk engine with its own view of the same exposures, and several returns slicing that balance sheet on different classification rules. All of them can be correct at once.

An agent can hold those versions side by side and reconcile them by entity, product, counterparty type, currency and maturity band, reporting where a difference follows a known definitional rule and where it does not. The unexplained residual is the point. The durable asset is the written bridge, which once explicit and versioned stops leaving with the person who held it.

Substantiation and the evidence that lives in inboxes

Every account on the balance sheet is attested by somebody who has to state what is in it and why it moved. The evidence sits in personal spreadsheets and mail threads, so the attestation becomes a signature on a list.

An agent assembles the substantiation file before the owner is asked for it: the balance, its composition, the ageing of open items, the movement since the last attestation with the entries responsible, and the supporting documents located rather than requested. The exceptions are the useful part: accounts whose composition has changed shape without explanation, items ageing past the point your policy allows, accounts with a stale owner or none at all.

Where finance and risk disagree by definition

The same portfolio serves two purposes, so it is measured two ways. Scope, valuation basis, netting, timing and classification all differ. Most of the gap is intended, and almost none of it is documented in a form that survives to the next cycle.

An agent builds the bridge from data rather than recollection. It reconciles the populations first, because a difference in what is included explains more than people expect, then attributes the remainder to named causes and quantifies each. It carries forward the explanation accepted last cycle and tests whether it still holds. What is left is the small unexplained amount, which is where an experienced person should be spending attention and rarely is.

Product control and the explain before the desk opens

Daily profit and loss attribution has an unforgiving deadline. The actual result has to be reconciled to what the risk measures predicted, and the residual explained by new business, market moves, fees, amendments and static data corrections, before trading begins.

An agent can assemble that overnight, attributing the movement across the factors your methodology defines and drafting the explain with the trades cited. Where it cannot attribute part of the residual it marks that explicitly, rather than smoothing it into a category that already carries a balance. The controller then spends the morning on the unattributed part, which is occasionally the part that matters.

Return production and the mapping nobody has reopened

The prudential and statistical returns your supervisor asks for take largely the same source data and map it into different shapes on different classification rules. That mapping logic often lives in a document older than the current product set, read differently each time.

An agent is only worth deploying here if the mapping stops being prose and becomes tested, versioned, executable rules with an owner. Then it produces the return, runs the validations before submission rather than after rejection, and reconciles it to the ledger and to the other returns that share a figure with it. The most valuable exception it raises is the unmapped one: a product booked under a legacy code, landing in a classification nobody chose.

Lineage from the reported line back to the transaction

The question always comes, usually late, usually from someone entitled to a fast answer. Where did this figure come from. An agent working inside the close records lineage as a by product: the source system, the query, the parameters, the extraction time, the record identifiers and every transformation between the transaction and the reported line.

That turns an archaeology exercise into a retrieval, and it imposes an honest constraint. Where the systems do not record a link, the agent must say so rather than construct a plausible path between two figures. Inferred lineage is worse than none, because it will be believed.

Provisioning inputs and the disclosure around them

The impairment model your market requires draws inputs from lending systems, collateral records, customer data and economic assumptions, and its outputs feed disclosure tables that readers examine closely.

An agent works the preparation layer. It reconciles the population feeding the model back to the ledger, produces the movement analysis between periods, attributing the change in the provision to new lending, repayments, transfers between classification stages and changed assumptions, and assembles the evidence behind any overlay so the basis is documented at the time. Scenario selection and weighting, and whether an overlay is still warranted, have named owners.

Late adjustments and the resubmission problem

Adjustments posted after the cut are the least documented part of the close and the most likely to cause trouble later. A journal lands, a figure that has already gone into a pack or a return moves, and the question becomes what else consumed that number.

An agent can hold a register of post cut adjustments with the basis for each, then trace the affected figure through everything downstream: the returns, packs, disclosures and schedules that quoted it. The decision about whether to resubmit is then taken with the consequences visible. Across cycles it also shows which adjustments recur, and one posted every quarter for the same reason is an upstream defect with a manual workaround.

Commentary, packs and the requests that follow

Movement commentary written from the ledger

Commentary is usually written last and from memory, which is why it so often restates the movement instead of explaining it. An agent computes the movement at the level your pack requires, identifies the portfolios and transactions driving it, separates timing from substance, and drafts the explanation with the entries cited. The accountant edits, adding the context that exists in no system.

The board and investor pack

The pack presents the same numbers in a different register, for an audience that reads the narrative more carefully than the schedules. An agent populates it from the reported ledger rather than from retyped figures, then runs the check that catches the most embarrassing errors: every figure quoted in the text reconciled to the table beside it.

The assurance request cycle

Requests arrive across mailboxes in inconsistent formats, and the same schedule gets produced repeatedly in different shapes. An agent maps each request to the report that satisfies it, produces the schedule with its source, population and extraction date stated, and tracks status by item and owner.

Proving it against cycles you have already closed

Do not accept a demonstration. Demonstrations use clean periods.

Run the agent in parallel across periods you have already closed and signed, including a quarter end and one messy cycle with a system change or restatement. Compare its bridges, explains and drafted commentary against what your team concluded, and score three things separately: was the number right, was the explanation right, and did it decline to conclude where your people escalated. The third score predicts whether this survives.

Where AI fails here

An agent cannot make an accounting judgement. Classification of a novel instrument, whether a provision is adequate, how a restructuring should be treated: it can lay out the considerations, and it should refer rather than improvise.

It cannot invent lineage the systems never recorded. If the link between a reported line and its transaction was lost in a migration, no reasoning recovers it, and an agent producing a confident path is doing damage.

A plausible but wrong explanation of a variance is more dangerous than no explanation, because it reads like diligence and stops the enquiry. Insist every drafted sentence cites the entries behind it, and treat untraceable output as a defect.

None of this repairs data that was wrong when it entered the ledger. An agent reconciles bad data neatly and explains movements that exist only because something was booked to the wrong place.

Accountability does not transfer either. The attestation, the return and the accounts carry the names of people with authority and liability. What changes is the state those people are in when they sign.

Where to start

Start with the finance to risk bridge or with account substantiation. Both are contained, both have a ground truth checkable within a single cycle, and neither requires the agent to write into a system of record. Take one portfolio or one account population, run it against a period you have already closed, and measure the disagreement honestly, separating cases where the agent was wrong from cases where it found something your team had missed.

If you want an honest view of which parts of your reporting cycle an agent could carry and which it should not touch, describe your close to Adopt AI Partners at hello@adoptaipartners.com.