Finance agents

Three finance jobs that fit an agent

1. Invoice and payment matching

Input: supplier invoices (PDF, photos on WhatsApp), bank statements, EcoCash merchant reports, Paynow transaction exports. The agent extracts the fields, proposes matches (invoice ↔ payment ↔ purchase order), flags exceptions with a reason (amount differs by the EcoCash fee; duplicate invoice number; currency mismatch USD/ZiG) and produces a reconciliation list. A person approves each posting in the accounting system.

2. Payment reminders and promise-to-pay

For recurring bills — school fees, rentals, subscriptions — the agent sends a utility template (US$0.004 to +263), takes a promise-to-pay or a Paynow payment, and logs the outcome. It never threatens, never discusses disputes, and hands off any pushback. Consent and tone rules are guardrails in code; see the collections row in the database.

3. FDMS-ready invoicing

ZIMRA's Fiscalisation Data Management System requires VAT-registered operators to transmit invoices — including buyer details, mandatory from 31 May 2025 — through fiscal devices or, for those able to build an API integration, a virtual fiscal device. An agent can assemble a compliant invoice from a WhatsApp order (buyer TIN, VAT number, line items) and stage it for submission; the submission itself should be a gated, logged step, because a fiscal invoice cannot be quietly un-sent.

Why the human stays in the loop

Two reasons. Errors in finance compound silently — a mis-matched payment is not noticed until month-end. And the Cyber and Data Protection Act's rule on automated decisions with significant effect on a person applies squarely to credit holds, dunning and account suspensions: those need consent or legal authorisation. Keeping a person on the posting step satisfies both.

Cost, and the batch lever

Documents are expensive tokens: a scanned invoice is thousands of tokens as an image. Two levers matter. First, none of this is real-time — the Batch API halves token prices for jobs that can wait hours, so nightly reconciliation runs at half the list price. Second, extract once and store the structured result; do not re-read the PDF every time the agent needs a field. Three thousand documents a month sits in a band of roughly US$60–300 depending on document quality and model tier (illustrative; basis on the costs page).

Controls checklist

• Read-only credentials to bank, EcoCash and Paynow exports; write only to a staging table.
• Every proposed posting carries the evidence it used (page, line, amount) so the approver can verify in seconds.
• Immutable audit log of proposals, approvals and rejections — also your evidence for a POTRAZ query or an auditor.
• Currency explicit on every line (USD, ZiG) with the rate and date used if converted; the interbank rate is published by the Reserve Bank and should be the tool's source, not the model's memory.
• Evaluation on a golden set of last quarter's real reconciliations, including the ones that went wrong.

Integrations

Accounting packages (Sage, Xero, QuickBooks, Pastel), Paynow reports and the ZIMRA FDMS API are described on the integrations page. For the payments tool contract see payments; for FDMS see ZIMRA.

Sources (accessed 2026-09-14): ZIMRA "Fiscalisation explained" — zimra.co.zw; FDMS buyer-detail deadline 31 May 2025 — RTC Suite; Batch API 50% discount — platform.claude.com; automated decision rule — DLA Piper Africa.