Approval-Driven Automation for B2B Processes Above 10k EUR
Last week an EU distributor's RPA bot approved a €47,000 supplier invoice with a wrong IBAN. The bank cleared it. Recovery took 11 days.
Approval-Driven Automation for B2B Processes Above 10k EUR
Last week an EU distributor’s RPA bot approved a €47,000 supplier invoice with a wrong IBAN. The bank cleared it. Recovery took 11 days. That is the failure mode approval-driven automation exists to prevent: an agent proposes; a human commits; the ledger only moves when both agree.
The propose-then-act pattern splits any high-stakes B2B workflow into two enforced phases. Phase one, the agent reads context, drafts an action, cites its evidence, and writes a signed proposal to a durable queue. Phase two, a human (or a second machine authority for sub-threshold cases) reviews the proposal, checks the citations, and issues an approval token. Only the token unlocks the side-effectful call. No token, no wire transfer, no ANAF submission, no contract signature.
Where the threshold sits
The 10k EUR boundary is not arbitrary. It aligns with reporting thresholds in Directive (EU) 2015/849 (AMLD) and matches the internal-control materiality most audited RO mid-market firms already use. Below the threshold, agents may act with post-hoc logging. Above it, propose-then-act is mandatory. The ENISA Threat Landscape 2024 documents supply-chain fraud as a top-five vector for organisations in this bracket, so the split maps to a live risk.
Architecture
Every proposal is an immutable record. A minimal shape we deploy in production:
proposal_id: prop_2026-08-30_0f4a
action: sepa_transfer
amount_eur: 47320.00
counterparty:
name: "Meridian Logistics SRL"
iban: "RO49AAAA1B31007593840000"
vat_id: "RO12345678"
evidence:
- invoice_pdf: sha256:8b2c...
- anaf_vies_check: valid_2026-08-29
- po_match: PO-2026-4412
risk_signals:
new_iban_for_supplier: true
amount_above_median: 3.2x
approver_required: role:cfo
expires_at: 2026-08-31T17:00:00ZThe new_iban_for_supplier flag is what would have caught the €47k mistake. The ANSPDCP guidance on automated decision-making under GDPR Art. 22 makes human review not just a control but a legal expectation whenever the decision produces significant effects. For firms in scope of NIS2 (Directive (EU) 2022/2555), the same audit ledger doubles as evidence for Art. 21 risk-management obligations.
Where teams get this wrong
Two failures dominate. First, treating the approver queue as a formality — humans rubber-stamp within seconds because the UI hides the risk signals. The fix is forcing the approver to acknowledge each red flag before the sign button unlocks. Second, letting the agent both propose AND control the queue; that collapses the separation of duties. The queue must live in a system the agent cannot write to except through the proposal API. Related failure patterns are catalogued in our LLM agent failure taxonomy and the operational playbook we ship with SOC triage engines uses the same token-gated model.
CAI Technology deploys propose-then-act as the default for any client workflow where a single mistake exceeds a month of the team’s salary budget. The pattern is boring, which is exactly why it holds under audit.