Faster close, fewer touches, stronger controls.
Finance operations are rules, tolerances and evidence. That makes them a strong fit for verified automation, and a poor fit for anything that can't explain itself to an auditor. We redesign AP, AR, reconciliation and checks so agents do the matching and preparation, controls run on every transaction, and people approve what moves money.
Six finance workflows where the checking costs more than the work.
These are the workflows we would examine first in a finance team or shared service center. They run through the ERP, the bank, email inboxes, approval chains and many spreadsheets.
| Workflow | Manual today | AI-native version | Stays human | Verification check |
|---|---|---|---|---|
| AP invoice processing | Clerks key invoices from PDFs and emails, match them to purchase orders and goods receipts, code the GL account and chase approvers. | An agent reads the invoice (or the structured e-invoice), runs the three-way match, proposes coding and routes the approval. | Payment approval, non-PO invoices above a limit and price or quantity variances. | Three-way match within tolerance. A duplicate check on vendor, invoice number and amount. Bank details must equal the vendor master. |
| Bank and intercompany reconciliation | Statement lines are matched by hand in spreadsheets. Unmatched items roll forward month after month. | Rules match the clear cases. An agent proposes matches for the rest with its reasoning, and drafts the journal entries. | Write-offs, adjustments and journal approval. | Matched totals must equal statement balances. No journal posts without approval. |
| Collections and cash application | AR applies payments from remittance emails, sends reminders and handles disputes in the inbox. | An agent applies cash from remittance advice, sends reminders by customer segment and classifies disputes. | Payment plans, credit holds and write-offs. | Applied amount must equal the amount received. Invoices in dispute are excluded from reminders. |
| Vendor onboarding and KYC/KYB checks | Staff collect forms, tax IDs, registry extracts and bank letters, then run sanctions screening one vendor or customer at a time. | An agent collects and validates documents, checks VAT IDs, pulls registry and beneficial-owner data and runs screening. | The risk decision, approval of new bank details and every screening hit. | Required evidence must be complete for the risk tier. Screening hits are never cleared automatically. Bank detail changes need four eyes. |
| Month-end close | Accruals, prepayments, checklists and variance commentary are prepared by hand under deadline pressure. | An agent proposes accruals from open POs and contracts, tracks the close checklist and drafts the variance commentary. | Controller sign-off and every judgment-based estimate. | Journals must balance and stay within materiality limits, with supporting evidence attached. |
| Expense audit | A reviewer samples a share of claims against policy. The rest go through unchecked. | Every claim is checked against the policy rules and the receipt. Only flagged claims reach a reviewer. | Decisions on flagged claims and policy exceptions. | Policy rules are versioned. Each flag cites the rule it breaks. |
Six numbers that show whether it worked.
The Audit measures these before anything is built. The Pilot is measured against the same numbers.
| Metric | Definition |
|---|---|
| Touches per invoice | How many times a person handles an invoice from receipt to payment. |
| Straight-through rate | Share of invoices or bank lines matched and approved with no manual correction. |
| Days to close | Working days from period end to closed books. |
| Unapplied cash and DSO | Cash received but not yet applied, and days sales outstanding. |
| Payment error rate | Duplicate, wrong-amount or wrong-recipient payments per 10,000 payments. |
| Review coverage | Share of transactions checked against the rules. Sampling gives a low number. Full checking aims for 100%. |
What can go wrong in finance operations, and the control for each.
In finance, the failure that matters is money leaving to the wrong place, or a decision nobody can evidence. Agents never hold payment credentials. Every proposal, check and approval is written to an append-only audit trail.
| Failure mode | Control | How we'd detect it |
|---|---|---|
| Duplicate payment | Fuzzy duplicate check on vendor, invoice number, amount and date before approval. Payment runs come only from the approved queue. | Post-payment duplicate scan each run. |
| Payment to fraudulent bank details | Agents cannot change the vendor master. A bank detail change needs an independent call-back and a second approver. | Alert on every bank detail change and every first payment to new details. |
| Wrong GL coding or misstated accrual | Coding rules first, with a confidence score. Low confidence goes to a person. Controller signs off accruals. | Variance analysis against budget and prior periods. |
| Reminder sent to a customer in dispute | A dispute flag blocks dunning on that invoice. | Complaint log and a weekly check of reminders sent against open disputes. |
| Sanctions or KYC hit cleared without review | Screening hits always go to a person. Clearance needs a recorded reason. | Audit trail of every hit, decision and reviewer. |
| Segregation of duties broken | Agent roles are mapped to your SoD matrix. No agent both prepares and approves the same transaction. | Periodic role review against the SoD matrix. |
AP invoice processing, from diagnostic to pilot.
Illustrative example — not a client result. The numbers below are assumptions chosen to show the method and the arithmetic. Your Audit replaces them with measured values.
| 01 | Diagnostic | Suppose a mid-sized distributor's finance team picks "supplier invoice processing" on the General track. The Snapshot might show a High leverage signal with Medium confidence, friction at the Processing and Verification stages, and Level 2 (workflow augmentation) as the indicated AI depth. |
| 02 | Audit | Suppose the Audit measures 3,000 invoices a month at 7 minutes each from receipt to approval. That is 3,000 × 7 = 21,000 minutes, or 350 hours a month, about 2.2 FTE at 160 hours. It also finds that about 30% of invoices have a variance, a missing PO or a missing receipt. |
| 03 | Pilot | Suppose the 70% that match cleanly go straight through, with a person reviewing a 5% sample for 3 minutes each. The 30% exceptions take 9 minutes each because they are the harder cases. The Pilot measures whether that holds. |
| Line | Assumption | Hours per month |
|---|---|---|
| Baseline | 3,000 invoices × 7 min | 350.0 |
| Matched, after | 2,100 invoices, 105 sampled × 3 min | 5.3 |
| Exceptions, after | 900 invoices × 9 min | 135.0 |
| After, total | 5.3 + 135.0 | 140.3 |
| Difference | 350.0 − 140.3 (about 1.3 FTE) | 209.7 |
At an assumed loaded cost of €45 an hour, 209.7 hours is about €9,440 a month. Notice that most of the remaining time is exceptions. The Audit would also look at why those exceptions happen, such as missing goods receipts, because fixing the cause can matter more than automating the step. It would also count new costs: compute, ERP integration and review time. If the numbers don't hold, the recommendation is "don't proceed".
Where our proof comes from.
We'll be plain about this. Our production proof is in AdTech, not finance operations. We have not yet run these finance workflows for a client.
The closest relative is programmatic clearing and partner reconciliation, which we run in AdTech production: matching what was bid, what cleared and what was billed, with a verification check before anything commits. Our governance working paper maps runtime controls such as rule checks, human escalation and a tamper-evident audit trail to the EU AI Act. The AdTech page and case studies show the production work.
Our engineering team and delivery network have built related systems: multilingual agentic workflows for financial operations, in production with a large active user base, a multi-agent financial copilot for analysis and transaction routing, and hybrid search over complex financial documents. See Selected builds.
That is why finance engagements start with the Audit. It measures your own baseline and records your controls, including "what must never be automated", before anything is designed.
Questions finance leaders ask first.
Will our auditors accept this?
The design aims to give auditors more evidence than a manual process does: every proposal, check and approval is logged with who or what did it and why. We recommend involving your auditors early, and the Audit records the controls they rely on today.
Does an agent ever release a payment?
No. Agents prepare and check. Payment release stays with the people and systems you authorize, and agents never hold bank or payment credentials.
We already receive e-invoices. Does that help?
Yes. Structured e-invoices such as XRechnung or ZUGFeRD remove the reading step, so more of the effort goes into matching and exceptions, where the time actually goes.
Find out which finance workflow is worth redesigning.
Start with the free diagnostic. Finance teams use the Other / General Business track, or the track for your company's sector. Move to the Audit only if the Snapshot points to something worth measuring.
| 01 | Free AI Operational Diagnostic | 10 steps, 8–12 minutes. Snapshot on screen and as a PDF. |
| 02 | Operational Leverage Audit, from $1,500 | Fixed fee, one workflow, 10–15 working days after data access. Measured baseline, verification architecture, benefit model with stated assumptions, go / no-go. |
| 03 | Pilot implementation, from $6,000 | Fixed scope. The workflow built and run in your environment, with an audit trail, measured against the Audit baseline. Typically 4–8 weeks. |