AI financial reporting that keeps figures sourced and decisions human

What AI financial reporting should and should not do

The simplest rule is a division of labour. Calculation, posting and closing are deterministic: the same inputs always give the same totals, and the code that produces them can be tested. Interpretation, drafting and follow-up are where AI adds value: reading many sources at once, noticing what changed and explaining it in plain language. When those two roles blur, a report stops being defensible

Proposal, not output

In a well-governed setup every piece of AI-generated content is labelled as a proposal. A drafted variance comment, a suggested reclassification or a summary of open exceptions waits for a named person to accept, edit or reject it. Nothing the AI writes changes a balance, and nothing it writes is published as if a person had written and checked it

That framing matters for AI in financial reporting because the audience of a report rarely sees how it was produced. A board member reading a monthly pack should be able to tell which sentences are reviewed commentary and which figures are confirmed, without having to ask

Where AI genuinely helps in the reporting cycle

AI is most useful where the work is reading, comparing and writing, and where a person still checks the result. Four areas stand out

• Drafting commentary: turning confirmed movements into a first draft of management commentary that a controller edits rather than writes from scratch
• Detecting variances and exceptions: comparing actuals with budget, the prior period and expected patterns, and flagging items that cross agreed thresholds or look unusual
• Assembling morning and monthly briefs: collecting cash positions, open reconciliation items, pending approvals and overdue tasks into one short brief
• Explaining movements: tracing a change in a balance back to the entries, counterparties and documents behind it, so the reader sees why and not only how much

A worked example: explaining a cash movement

Suppose operating cash in one entity fell sharply overnight. The deterministic layer has already posted every event as a balanced entry and reconciled the bank source. An AI agent then reads those entries and drafts its explanation: two supplier payments released after approval account for most of the fall, a customer receipt expected yesterday has not arrived, and one bank line is still unmatched. Each statement links to the entries and the reconciliation item it relies on. The treasury lead reviews the draft, corrects the wording on the late receipt and approves it for the morning brief

Notice what the agent did not do. It did not calculate the closing balance, it did not match the open bank line by guessing, and it did not chase the customer on its own authority. It proposed, explained and pointed at evidence

What must stay deterministic and under human approval

Some parts of financial reporting cannot be delegated to a probabilistic model, however capable. Totals, exchange rates, fees, postings and period close belong to tested, deterministic code, and the ledger remains the reference system for every balance

• Figures: every number in a report is computed from ledger entries by defined rules, never typed or generated by a model
• Postings: every balance change is an event with a balanced double entry, and a mistake is corrected with a reversing entry rather than edited or deleted
• Closes: a period is closed by the close process, and reopening a closed period is treated as a sensitive action
• Sensitive changes: adjustments, opening balances, post-hoc rate changes and transfers above limits need two separate people, one who records and one who approves

If you want to see why an append-only record is the right foundation, our guide to the immutable double-entry ledger and its audit trail covers the mechanics in depth

The practical consequence is that AI can recommend an adjustment, but the adjustment itself is entered by a person, approved by a different person and posted by the deterministic engine. The AI recommendation becomes part of the evidence, not part of the books

Source evidence and data status beside every figure

Source evidence is what turns an AI-drafted report from plausible into defensible. The principle is simple: no figure appears without a source and a time. A reader should be able to open any number and see which ledger, which reconciliation and which moment it reflects

Four statuses worth showing

• Confirmed: reconciled against its independent source and approved
• Computed: derived by deterministic rules from confirmed inputs
• Unreconciled: recorded, but not yet matched to its source
• Unknown: the data is missing or late, and the report says so openly

The last status is where many reports fail. When a bank feed is late or a subsidiary has not submitted, the easy shortcut is to show zero. A zero looks like a fact and quietly corrupts totals. The safer rule is that missing data is never silently treated as zero; it stays marked unknown, and the report shows the position as of the latest valid data, with that time visible

In a multi-entity group this discipline compounds. Consolidated reporting with each figure's status beside it lets a group CFO see that one company is confirmed to last night while another is still unreconciled, instead of receiving a single total that hides the difference

An approval workflow for AI proposals

Human approval works best when it is a workflow with a record, not an informal glance. A practical sequence looks like this

1. The deterministic close and reconciliation run first and publish figures with their status
2. The AI agent reads those figures and the open exception list, then drafts commentary, flags variances and proposes follow-up actions, each labelled as a proposal with its sources
3. A named reviewer accepts, edits or rejects each proposal, and edits are kept alongside the original draft
4. Anything that would change the books is routed to the normal recording and approval path, with the recorder and approver as two different people
5. The approved report is published as a numbered version, and a later correction becomes a new version rather than a rewrite

Every step leaves an audit trail: who or what proposed, which sources were cited, who approved, what was changed and when. That trail is what lets an auditor or a board member reconstruct how a sentence in the monthly pack came to be written

Clear authority boundaries make this workable at scale. In an AI command center with specialist agents, observing, analysing, proposing and drafting are free, while execution needs permission and anything touching money needs management approval

Governance checklist for AI-drafted reports

Before AI-drafted content reaches management or the board, check that each of these statements is true

• Every figure is produced by deterministic, tested calculation from the ledger
• Every AI-generated sentence is visibly labelled as a proposal until a person approves it
• Every claim in commentary links to the entries, reconciliation items or documents it relies on
• Every figure shows its source, its time and its status, and missing data is shown as unknown
• Each agent has a written authority table stating what it may observe, draft and propose, and what it must never do
• Recording and approval of sensitive changes are done by two different people
• Published reports are versioned, and corrections are issued as new versions
• The record of proposals, edits and approvals is retained and exportable with the rest of your data

Treat the checklist as a release gate. If one item fails, the report can still go out, but the affected section should state what is missing rather than paper over it

Pitfalls that undermine trust in AI reporting

Hallucinated numbers

A language model asked to summarise can produce a figure that appears nowhere in the source data, or round and combine numbers in ways that look right but are not. The cure is structural: the model quotes figures from the deterministic layer by reference, and any number in AI text that cannot be traced to a source is rejected in review

Unlabelled AI output

When drafted commentary is pasted into a report without a label, readers assume a person wrote and checked it. That misplaces accountability. Keep the proposal label until approval, and record who approved and what they changed

Missing sources and stale data

Commentary that explains a movement without citing entries is opinion. A brief built on last week's feed without saying so is misleading. Both are avoided when the reporting cycle starts from a clean nightly close and source reconciliation, so the AI works on figures that already carry their status and age

How to introduce AI into reporting without disrupting the close

Start narrow. Pick one report, such as the daily cash brief, define its sources and statuses, and let the agent draft while the team keeps producing the report the current way. Comparing the two over a period shows where the drafts save time and where review effort goes. If cash is the first report you automate, our guide to treasury daily close and reconciliation automation explains the groundwork

At Arfinex this follows the same delivery path as any financial change: an introduction and scope review, design and build in a sandbox with synthetic or anonymized data, a parallel run alongside your current method, and go-live only with your approval. Regulated activities stay with licensed providers, and local accounting and tax treatment is confirmed by your own accountant or adviser

Faster reports, same accountability

AI financial reporting is worth adopting when it shortens the path from data to understanding without moving accountability. Let AI draft, detect and explain; keep figures, postings and closes deterministic; show the source and status of every number; and record every proposal and approval. That combination gives you reports that are quicker to produce and easier to defend