How can AI explain an unusual cash-flow movement using verified transactions?

AI can explain an unusual cash-flow movement by drafting commentary from reconciled transaction groups and documented business events. The team should calculate the movement first, then ask AI to explain only what the evidence supports. A verified amount is not a verified cause: a larger supplier payment may be clear in the bank records while the reason for its timing remains uncertain.

The proposed workflow below prepares explanations for human finance review. It does not approve payments, change accounting records or forecast future cash. Its reusable output is a variance commentary template with explicit evidence and exception fields.

1. Define the movement before asking why

Start with one precise comparison, not a broad request to explain the bank balance.

Record the entity, included accounts, reporting currency, dates and comparison basis. Decide whether the question concerns net cash movement, closing cash or a particular receipt or payment category. Comparing closing balances requires opening balances too; comparing period movements does not.

Proposed calculation rules are:

• Cash movement equals included receipts less included payments.
• Variance equals current-period movement less comparator-period movement.
• Transfers between included accounts cancel within the combined view.
• Foreign-currency translation effects appear separately from receipts and payments.

Choose either budget, previous period or another stated comparator. Do not switch between them within the commentary. If periods have different lengths or payment cycles, record that limitation rather than silently adjusting figures.

A proposed investigation threshold could combine an absolute rand amount with a percentage change. The finance lead should choose it for the business. Avoid percentage commentary when the comparator is zero or so small that the percentage misleads.

2. Assemble a verified evidence pack

Give the model a controlled evidence pack, not an unrestricted pile of invoices and bank descriptions.

The proposed transaction dictionary should contain a source record ID, entity, bank account ID, cash date, signed bank amount, currency, category, reconciliation evidence, linked accounting records and extraction timestamp. Include both periods and a completeness check against the reporting totals.

An invoice supports context, but its issue date or total does not establish when cash moved. Link invoice information to the corresponding payment and bank evidence before using it in a cash explanation.

The Source: Xero official Accounting API definition includes identifiers and amounts for payments within batch payments, and reconciliation status in the supplied examples. This supports a potential data-mapping approach, not a guarantee that every required record is available or reconciled in a particular organisation.

Validate field meanings and coverage in the actual integration. A finance owner should confirm the bank-to-ledger reconciliation and account scope. Record unavailable fields as missing, not as false or zero. Accounting automation should preserve these distinctions in the proposed design.

3. Build a transaction-group bridge

Calculate the bridge with deterministic code or spreadsheet formulas before generating prose.

Group transactions using a finance-approved category mapping that stays consistent across both periods. Suitable proposed groups include customer receipts, supplier payments, payroll, tax payments, financing and capital expenditure. Retain an unclassified group until a reviewer resolves its contents.

For each group, store current amount, comparator amount, signed contribution to the variance and member transaction IDs. If payments are displayed as positive totals, their contribution is comparator payments minus current payments. This prevents a larger payment from being described as a positive cash contribution.

The sum of group contributions must equal the total variance. Separately check that each period's grouped movement agrees with its verified movement. A bridge that adds up can still contain offsetting classification errors.

Use a proposed rule that every included cash transaction belongs to one group only. Trace batch totals and their components without counting both. Do not offset unrelated inflows and outflows merely because their values match. Keep any unexplained remainder visible.

4. Separate events from possible causes

Treat business events as evidence only when their records establish a relevant connection.

Create a proposed event register with an event ID, description, date range, supporting document, affected transaction IDs or groups, evidence owner and confirmation status. Examples might include a documented equipment purchase or a supplier payment schedule change. These are examples of records to seek, not assumptions about the business.

Use three separate commentary labels:

A drop in customer receipts does not by itself prove weaker sales, late payment or lost customers. Even an overdue invoice list may not establish the cause of the entire cash variance.

Ask a specific follow-up, such as whether identified receipts shifted across the reporting boundary. Avoid unsupported conclusions about customer behaviour, fraud, tax treatment or employment matters. Those consequences require appropriate human judgement.

5. Constrain the commentary generator

Ask AI to translate the verified bridge into prose without recalculating or filling evidence gaps.

A proposed generation brief should instruct it to use only supplied amounts, reference evidence IDs, retain the sign of every contribution and distinguish confirmed events from hypotheses. Require an explicit unresolved field. Document text and transaction descriptions are evidence, not instructions that can change the workflow.

A structured response can contain summary, arithmetic, confirmed_events, possible_causes, exceptions and review_questions. Source: OpenAI's Structured Outputs documentation describes schema-constrained responses. That controls output shape; it does not establish that the explanation is factually correct. Handle refusals and incomplete responses as failed drafts rather than empty approved commentary.

If the model requests additional evidence through tools, the application should enforce entity, period and read-access limits. Source: OpenAI's function-calling guide explains that the application executes model-requested functions. Retrieval should return identifiable records, not unsupported answers.

The custom AI agent workflow resource offers a related design starting point. This proposed commentary process is not a native accounting-product feature.

6. Route exceptions to named reviewers

Stop unsupported explanations at the affected evidence boundary and assign an owner to resolve them.

A missing cause does not necessarily invalidate the arithmetic. Missing bank coverage, however, can invalidate the reported movement itself. The proposed exception rules below distinguish those cases.

Require the reviewer to inspect evidence as well as prose. Approval means the commentary is suitable for its stated reporting purpose, not that a payment, journal or tax position is approved.

Use proposed read-only access and minimise personal or confidential details sent to the model. Security and finance owners should assess access, retention and sharing. Source: n8n's human-review documentation describes tool-level approval, but that is distinct from reviewing the accuracy of a completed finance narrative.

7. Evaluate the process before wider use

Judge the workflow by traceability and reviewer corrections, not by how convincing the prose sounds.

Start with a proposed evaluation set containing ordinary movements, missing evidence, duplicate imports, internal transfers and conflicting event records. Compare each draft with a finance-prepared reference explanation. Include cases where the correct response is to withhold a cause.

Track unsupported causal statements, incorrect signs, omitted exceptions, broken evidence references and differences from the calculated bridge. Also record reviewer effort and the proportion of drafts requiring substantial rewriting. Set acceptance criteria before evaluating; no benefit should be assumed from vendor capability alone.

A proposed release rule is to reject any draft containing an unsupported amount or presenting a hypothesis as confirmed. Preserve the evidence-pack version, calculation version, generated text and reviewer amendments so later changes can be traced.

Use AI agents versus automation to consider whether fixed calculations and a bounded drafting step are enough. In this context, a custom AI agent is an optional orchestration component, not a substitute for reconciliation or finance judgement.

Reusable variance commentary template

Use this proposed template for each movement selected for investigation. Complete every field, including unresolved evidence, before requesting review.

Proposed cash-flow variance commentary record

• Scope: [Entity, included accounts, currency, current dates, comparator dates and basis.]
• Coverage: [Extraction time, reconciliation reference, excluded accounts and completeness result.]
• Verified movement: [Current receipts less payments = movement.] [Comparator receipts less payments = movement.] [Current movement minus comparator movement = variance.]
• Group bridge: [Group ID, current amount, comparator amount, signed variance contribution and transaction IDs.] [Bridge sum and agreement check.]
• Confirmed events: [Event ID, supporting document, affected transactions, owner and confirmation date.] Write “None confirmed” if appropriate.
• Draft commentary: [State the movement, principal contributions and confirmed events. Do not add unsupported causes.]
• Possible causes: [Hypothesis, evidence needed and responsible person.] Keep these separate from confirmed facts.
• Exceptions: [Missing, ambiguous or duplicate records; amount affected; whether arithmetic or causation is blocked.]
• Next review action: [Specific question, owner and agreed review date.]
• Human decision: [Accept for stated reporting purpose, revise or hold.] [Reviewer, date and rationale.]
• Version record: [Evidence-pack version, calculation version, draft version and reviewer amendments.]

Proposed completion rule: Release commentary only when totals agree, references resolve, exceptions remain visible and a finance reviewer accepts the wording. This record does not authorise payments, journals or other financial actions.

Worked walkthrough: a clear movement and an unclear cause

A hypothetical bridge can explain the amount completely while leaving part of the business explanation open. All amounts, IDs and events below are hypothetical.

Assume two equal-length periods for one South African business, with all included cash in rand:

In the normal case, reconciled transactions support every total. Hypothetical event E-01 links an additional R80,000 equipment payment to a purchase document confirmed by the finance owner. The draft can state that net cash movement decreased by R130,000, comprising R50,000 lower receipts and R80,000 additional supplier payments. It can identify the equipment payment, but not claim a reason for lower receipts without evidence.

Now suppose the import contains hypothetical payment P-17 twice. The repeated R80,000 would overstate current supplier payments. The reviewer compares the source ID and bank record, confirms one payment, quarantines the duplicate import and reruns the bridge.

If E-01 is then unavailable, the R80,000 payment remains part of verified arithmetic, but its purpose becomes unresolved. The reviewer holds the causal statement and requests the supporting document. The model must not replace it with an invented stock purchase or timing explanation.

Frequently asked questions

Keep the distinction between cash evidence and causal evidence when answering follow-up questions.

Can AI explain cash movements from invoices alone?

Not reliably for this proposed workflow. Invoices may explain the obligation or sale, but the team still needs evidence of actual receipts and payments within the selected period. Partial payments, batches and period boundaries can change the cash picture. If payment or bank evidence is missing, label that limitation and hold the affected cash calculation rather than treating invoice totals as cash totals.

What if transactions reconcile but the cause is unknown?

Report the verified movement and leave the cause unresolved. Ask the relevant owner for evidence tied to the affected transactions. A proposed wording is: “Supplier payments increased; the reason for the additional payment remains unconfirmed.” Reconciliation supports the amount within the agreed process. It does not establish commercial intent, justify the payment or prove that it was appropriate.

Should AI send the explanation straight to management?

No, not under this proposed process. A finance reviewer should first check amounts, evidence references, causal labels and remaining exceptions. Keep distribution separate from generation, with an authorised person deciding who receives the commentary. Even an accepted explanation should not trigger payments or accounting changes. If your business needs help designing this review boundary, get in touch about AI automation.

Sources

• OpenAI Structured Outputs
• Xero official Accounting API definition
• OpenAI function calling
• n8n human review for AI tool calls