Reported operating use · Finance software / expense operations
AI expense policy review: Ramp example
Ramp's Policy Agent, powered in partnership with OpenAI, reviews expense transactions against cited agent-facing policy and recommends dispositions. Ambiguous cases escalate and reviewers retain authority. It can support configured auto-approval but does not validate GL or accounting codes. Its accuracy claims lack a published ground-truth protocol.
Expense-policy interpretation, transaction review, approval recommendation and escalation
Ramp
Available Ramp Plus policy-review feature; early customer operating results are reported, although individual customers are unnamed.
A published third-party example, not a Nuvepro customer case or evidence of a partnership.
The workflow at a glance
Ramp reported work
Illustrative summary of the published account
Reported work components. Their arrangement does not imply a fixed execution order or simultaneous work.
- Policy context
Transactions and cited policy
Policy Agent combines expense evidence with agent-facing rules.
- Disposition work
Recommend or configured auto-approve
The account describes recommendations and optional configured approval.
- Human review
Escalate ambiguous context
Reviewers retain authority; accounting-code validation is outside scope.
What the source reports
Policy Agent combines transaction context and cited agent-facing policy, recommends dispositions and can support configured auto-approval. Ambiguity escalates and reviewers retain authority. It does not validate GL/accounting codes. Its accuracy claims lack a published ground-truth protocol.
Reported technology: Ramp Policy Agent powered in partnership with OpenAI. Scope and outcomes are source-reported.
What the enterprise environment would need
The following is a proposed project design for an enterprise doing similar work. Published details are marked; the remaining systems, data and scenarios would be agreed for your project.
Proposed practice environment
Review an expense against the governing policy
Illustrative project design to scope with your enterprise.
- 1 · Set up the environment
Expense policy fixtures
Fictional receipts, employee context, currencies and versioned handbook plus converted rules.
- 2 · Preparation and agent contributions
Draft expense disposition
Cite applicable policy and mark missing or ambiguous context.
- 3 · Independent checks
Check policy conversion fidelity
Independent finance labels test exceptions, currency assumptions and unauthorized approval.
- 4 · People decide acceptance
Finance reviewer accepts case
Resolve handbook drift and approve only the demonstrated practice scope.
A reviewable evidence package
An evidence-linked expense disposition and escalation register.
If a check fails → withhold acceptance.
A legitimate handbook exception is absent from converted agent rules. Withhold acceptance, repair, independently recheck and return to owner review.
Failure or owner rejection returns to review
- Hold
- → 2 · Revise work
- → 3 · Rerun independent checks
- → 4 · Owner review
Operating boundary. Approval-safe test accounts only; no real expense approval or accounting-code validation claim.
- Systems and tools
- Inferred: use Ramp test accounts and policy fixtures rather than rebuild expense management.
- Data and state
- Inferred: policy version, receipt, transaction amount/currency, employee context and review state.
- People and handoffs
- Published reviewers retain authority; inferred scenarios should exercise employee/admin differences.
- Rules and authority
- Published thresholds/exceptions and conservative uncertainty routing; inferred authorization must remain outside model discretion.
- Normal and exception cases
- Published missing/ambiguous context escalates; inferred currency, conflicting rules and fraudulent receipts deserve cases.
- Reset and replay
- Not published. Proposed: synthetic transactions, frozen rates/policies and approval-safe replay.
A proposed simulation
An illustrative practice design derived from this work, not an executed assessment.
Review compliant, disallowed and ambiguous expenses across policy versions, currencies and exception authority.
Review gate: Verify each disposition against finance-approved labels, cited policy and escalation rules; check divergence between converted agent policy and handbook.
Challenge the result: Add a legitimate policy exception missing from the converted agent policy. The process must detect the policy-fidelity gap rather than blame only the candidate.
What needs checking
- Published: Cited policy reasoning and review/audit visibility; reported accuracy methodology absent.
- Proposed: Independently label policy-boundary cases; check unauthorized auto-approval and handbook/agent-policy drift.
- Published: Real transaction and employee context plus converted policy; no conversion-fidelity study supplied.
- Proposed: Compare agent-facing rules with the authoritative policy, including exceptions and historical conversion rates.
Review environment fidelity, human capability and agent performance separately. The enterprise’s accountable owners decide acceptance of the shared workflow and any operating use.
How an enterprise could build similar work
A proposed build order for an enterprise doing similar work, written by Nuvepro from the published account and our interpretation of the environment and checks above. It is not how Ramp built or operates this work.
- Frame a proposed test projectThis is a proposed design for a similar project, not how Ramp built Policy Agent. Use Ramp test accounts and policy fixtures rather than rebuilding expense management, and review compliant, disallowed and ambiguous expenses.
- Define data and stateCapture policy version, receipt, transaction amount and currency, employee context and review state. Use synthetic transactions with frozen exchange rates and policies, so every run can be replayed without touching real approvals.
- Keep authority outside the modelReviewers retain authority, and missing or ambiguous context escalates conservatively. Authorization must remain outside model discretion, so the model cannot grant itself approval rights. Scenarios should exercise differences between employee and admin roles and between policy versions.
- Build normal and exception casesAdd currency differences, conflicting rules, exception authority and fraudulent receipts to the compliant and disallowed expenses. Missing or ambiguous context should escalate instead of being approved, and each case should carry a finance-approved label to compare against.
- Verify every dispositionCheck each disposition against finance-approved labels, cited policy and escalation rules. Label policy-boundary cases independently, and check for unauthorized auto-approval and for drift between the converted agent policy and the handbook.
- Run the policy-fidelity failure testAdd a legitimate policy exception that is missing from the converted agent policy. The process must detect the policy-fidelity gap rather than blame only the candidate. Replay is not published, so approval-safe replay is proposed.
Your systems, data, rules and accountable owners decide the real project. Agree acceptance criteria, escalation owners and operating responsibility with the people accountable for the work before anything runs on live records. Treat the first run as a rehearsal on test records: compare the draft output with an independently prepared answer, and record what the people reviewing it change.
Questions about this example
- What does Ramp's Policy Agent do with expenses?
- It combines transaction context with cited, agent-facing policy to recommend dispositions for expense transactions, and it can support configured auto-approval. It is an available Ramp Plus feature. It does not validate general ledger or accounting codes.
- Who stays in control of expense decisions?
- Per the source, reviewers retain authority and missing or ambiguous context escalates instead of being decided automatically. Reasoning cites the policy and is visible for review and audit. Configured auto-approval is also supported, within published thresholds and exceptions.
- How strong is the evidence for this example?
- Reported use: the feature is available, and early customer operating results are reported, but individual customers are unnamed. Ground-truth definition, accuracy sample, conversion recall and a named-customer breakdown are not supplied, and accuracy claims lack a published protocol.
- What would an enterprise need to build something similar?
- Nuvepro's proposed design, not Ramp's, would use Ramp test accounts and policy fixtures rather than rebuild expense management. Cases cover compliant, disallowed and ambiguous expenses. Finance-approved labels and a comparison of agent policy with the handbook check each disposition.
Shape this around your enterprise
Bring a similar workflow and the people responsible for it. Together, define the records and systems to use, the result your team will build or configure, and how its owners will check it. Nuvepro scopes the simulation, environment and preparation. Agree any existing-agent evaluation, implementation assistance and production responsibilities separately.
Sources and limitations
Curation review: . This is the date this source example was reviewed, separate from source publication dates.
- Ground-truth definition, accuracy sample, conversion recall and named-customer breakdown are not supplied.
More finance and accounting examples
- AI AP exception resolution: Varick pattern by Eyad KhraisVarick AP pattern (customer unnamed)
- AI billing discrepancy review: Telstra exampleTelstra
- AI cash application: Varick pattern by Eyad KhraisVarick cash-application pattern (customer unnamed)
- AI restaurant cash reconciliation: Ted Taylor exampleTed Taylor, independent restaurant franchisee