AI Expense Capture: OCR, Policy Flagging, and Approval Automation
Expense processing costs more than most finance teams realise. At 200 staff submitting monthly claims, a 5-day end-to-end cycle means someone is always chasing a receipt, waiting for an approval, or manually keying data into the ERP. AI handles all three stages. Here is what the setup looks like and what each tool is responsible for.
The Hidden Cost of Manual Expense Processing
Expense processing sits in a blind spot in most finance functions. It is not a strategic process, so it does not get the same attention as close or forecasting. But at scale, it consumes a significant amount of staff time across three groups: the employees submitting claims, the managers approving them, and the finance team checking, chasing, and posting them.
A typical mid-market business with 200 employees processing monthly expenses sees roughly 400 claims per month. If each claim takes 15 minutes to submit (finding the receipt, logging in, typing the details), 10 minutes for manager approval (reading the email, finding the policy, deciding), and 5 minutes for finance to review and post, that is 200 person-hours per month on expense administration. At average salary costs, the processing cost per claim is often higher than the claim itself.
The goal of AI expense automation is not to eliminate expense claims. It is to reduce that 200 hours to under 20, with fewer errors and a cleaner audit trail.
How We Built This With an Edinburgh Professional Services Firm
Our client is a professional services firm in Edinburgh with 120 staff. Before the project, their expense process ran entirely on paper and email. Staff photographed receipts and emailed them to their manager. Managers replied by email to approve. Finance printed the approvals, manually keyed the claim details into their accounting system, and filed the paper receipts. End-to-end: 5 days from submission to ledger posting. Rejected claims were sent back for correction, adding another 3-5 days.
We redesigned the process in three stages. First, receipt capture via Microsoft Lens on mobile, feeding into SharePoint with Copilot reading the receipt details. Second, policy checking via an automated rules engine in Power Automate, with Claude Projects handling the edge cases (client entertainment ambiguity, late submissions, foreign currency claims). Third, approval routing via Power Automate, with auto-escalation after 48 hours and direct posting to their accounting system on approval.
After: same-day processing for 90% of claims. Finance team now reviews only flagged exceptions, roughly 8% of total claims. End-to-end cycle for clean claims: under 4 hours. The other 10% that need human judgment now have better documentation than they had before, because the AI writes a structured flag reason rather than leaving it to the manager to figure out what to do.
Want to go deeper? Our AI for Finance Leaders course covers this in detail with practical templates and exercises.
The Three Components of AI Expense Automation
Expense automation has three distinct technical components. Each can be implemented independently, but the full benefit comes from connecting all three into a single workflow.
Component 1: Receipt Capture and OCR
The first step is getting the receipt data into a structured format without manual keying. OCR (optical character recognition) reads the receipt image and extracts the key fields: merchant name, date, amount, currency, and where possible, the category (meal, travel, accommodation, etc.).
Modern OCR is good enough that it handles most printed receipts cleanly. Where it struggles is handwritten receipts, low-quality phone photos, and non-English receipts. For those cases, Claude Vision or GPT-4V provides a significant accuracy improvement over traditional OCR, because it interprets context rather than just extracting text.
Component 2: Policy Checking and Flagging
Once the receipt is captured, the claim is checked against the expense policy. Rule-based checks handle the clear cases: meal over the per-person limit, hotel over the nightly rate, category not in policy. These can be configured in Power Automate or n8n without AI.
AI adds value on the edge cases. A rule cannot tell whether a restaurant bill is client entertainment (in policy) or a team meal that should have been pre-approved (out of policy). A rule cannot assess whether a late submission has a reasonable explanation or is a pattern. A rule cannot handle the nuance of a business trip where the hotel cost was high because the conference venue was expensive and alternatives were further away.
We load the expense policy into a Claude Project, then route edge cases to it with the claim details and context. Claude returns a structured assessment: whether it is likely in or out of policy, what additional information would help, and what the finance team should ask for if it escalates. The finance team makes the final call, but with a documented rationale rather than starting from scratch.
- Rule-based checks: Power Automate or n8n (fast, no AI cost on high volume)
- Edge case assessment: Claude Projects (policy-aware, documented output)
- Output: Clean / Flagged for information / Flagged for review / Rejected
Component 3: Approval Routing and ERP Posting
Approval routing is the component that currently lives in email for most organisations, which is why it is slow and creates audit gaps. Moving it to Power Automate or n8n fixes both problems: the approval request is structured and tracked, the escalation is automatic, and the approved claim posts to the ERP without manual intervention.
The workflow we build sends the approver a structured request: receipt image, extracted fields, policy check result, and any flags. The approver approves or rejects with one click from mobile or desktop. If no response within 48 hours, the workflow escalates to their line manager with a note. Approved claims post to the accounting system automatically. Finance only sees the exceptions.
- Microsoft 365: Power Automate with Teams approval cards
- Non-M365: n8n with email-based approvals and webhook responses
- ERP posting: Power Automate native connectors for D365; n8n API integration for SAP, Oracle, NetSuite
What to Automate First
The three components above can be implemented independently. If your organisation is starting from a paper and email process, the highest-impact first step is approval routing, not receipt capture. Getting approvals out of email and into a trackable workflow has immediate impact on processing time and provides a clean audit trail even before OCR is in place.
Once approval routing is running, add receipt capture. The combination of structured capture and tracked approval reduces the finance team's manual work by around 60-70%. Policy checking with AI is the third step, and most impactful for organisations with complex policies, multiple currencies, or travel-heavy workforces.
For organisations already using Microsoft 365, all three components can be built using tools already included in existing licences: Microsoft Lens, Power Automate, and Copilot. The build is typically 2-3 weeks of configuration and testing. This is different from the AP automation workflow described in our AP automation without new software guide but follows the same principle: use tools you already have before buying new ones.
Where to Start
Map your current expense cycle end-to-end before making any changes. Document every step from the employee photographing a receipt to the payment hitting their bank account. Count the number of manual touches and estimate the total time across all participants. This gives you the baseline to measure against and identifies which component to build first.
For a broader view of where expense automation fits within your procure-to-pay process, see our AI in procure-to-pay guide. Expense capture is one of five P2P workflows where AI reduces manual processing — and often the fastest to implement because it does not require ERP reconfiguration.
Our AI finance audit includes a review of your expense and AP workflows as part of a broader finance function assessment. If you want the full implementation handled, our AI consulting team designs and builds the workflow, including the Power Automate flows, Claude Projects policy checker, and ERP integration. For teams who want to build this capability in-house, our AI for Finance Leaders training covers workflow automation tools in a finance context.
Work with Prime AI Solutions
AI for Finance Leaders
8 modules covering FP&A, reporting, automation, and governance. No coding required. From £99.
View courseLive Workshop for Your Team
Half-day or full-day live sessions for finance teams. Tailored to your tools, workflows, and industry.
See workshopsAI Finance Audit
We map your finance function, identify what to automate, and give you a prioritised action plan.
Learn moreAI Consulting
We design and build the workflow, configure the tools, and train your team. 90-day typical engagement.
Learn more