Stop reading thought-leadership about AI in finance. Here are seven concrete, copy-pasteable workflows that founders, ops leads, and accountants are running today with MCP for business expenses — using ExpenseMonkey and Claude.
If you've read our plain-English MCP explainer, you already know MCP for business expenses means an AI assistant can read, query, and act on your live expense data through a standard, permissioned interface. What follows is the practical layer — what to actually do with that, starting Monday morning.
Every workflow below is built on the ExpenseMonkey Claude integration via the Model Context Protocol. Each one includes the exact prompt to try, what it does under the hood, and the realistic time saving you should expect.
Workflow 1 — Conversational expense lookups
The simplest, highest-frequency win: ask questions of your live expense data instead of building a report. This is the workflow that quietly retires half your one-off dashboards.
> Show me every expense over $500 in May 2026, grouped by category, with the top 3 vendors per category.
Under the hood: Claude calls ExpenseMonkey's MCP search_expenses tool with a filter, then aggregates the result. No SQL, no Excel pivot, no waiting on the finance team.
Time saved: 10–20 minutes per ad-hoc question. Across a 10-person company, this alone often pays for the integration.
Pro tip: Pin frequently-asked questions as Claude Projects so your CEO can self-serve without Slacking finance. Claude Projects docs.
Workflow 2 — One-shot receipt processing
The old way: open ExpenseMonkey, click "Add Expense," fill in vendor, date, category, amount, attach the receipt, submit. The MCP way: drop the receipt into Claude, say "process this," done.
> Here are five receipts from my Lisbon trip. Create expense records in ExpenseMonkey, category = "Travel — International," and flag any over our €120 per diem cap.
Under the hood: Claude OCRs each image, extracts vendor / date / total / VAT / currency, calls create_expense via MCP for each one, then runs a policy check against your cap. You get a confirmation table with anything that needs review.
Time saved: 30–45 seconds per receipt → 2–3 seconds. For a heavy traveller, this is ~30 minutes per trip.
Pro tip: For paper-only receipts, take a phone photo first. Modern vision models read crumpled, faded, low-light receipts surprisingly well — but never substitute for a clear scan when amounts are over your audit threshold.
Workflow 3 — Bulk categorisation cleanup
Every growing company has the same archaeology problem: a backlog of "Uncategorised" expenses that nobody has the heart to clean up. MCP turns that from a half-day project into a coffee-length task.
> Find all uncategorised expenses from the last 90 days. For each, propose a category based on similar past entries from the same vendor or amount range. Return a table; I'll approve in bulk.
Under the hood: Claude queries your uncategorised set, looks up the historical categorisation pattern per vendor, and proposes a best match with confidence. You get a single table to scan; one click in ExpenseMonkey applies the changes via MCP.
Time saved: What used to be a 4-hour quarterly cleanup is typically under 30 minutes — and the AI's classification accuracy on familiar vendors is often higher than a human's, because it doesn't get bored.
Realistic expectation: Expect ~85–95% accuracy on first pass. The 5–15% you reject teach the model your edge cases. By the third run, your team is reviewing rather than categorising.
Workflow 4 — Real-time policy enforcement
Policy enforcement traditionally happens after the fact — during close, when nothing can be done about it. MCP shifts it left.
> Every Monday, check all expenses submitted in the previous week against our policy doc. Flag any that exceed limits, are missing receipts, or look like personal spend. Post the list to #finance in Slack.
Under the hood: A scheduled Claude task pulls the week's expenses via MCP, cross-references them against your policy document (also exposed through an MCP server or a Claude Project), and writes the exceptions list to Slack via Slack's MCP server. Slack and ExpenseMonkey never talk directly — Claude is the glue.
Time saved: Catches issues 2–4 weeks earlier than a typical close-time review. The dollar impact is usually larger than the time saving: a policy breach caught at submission can still be discussed; one caught at audit is just an audit finding.
Workflow 5 — Card reconciliation autopilot
Reconciliation is the canonical "boring middle." Match this card charge to that receipt. Match a hundred of those. This is what MCP-powered AI is genuinely great at.
> Reconcile this week's company card charges against expenses in ExpenseMonkey. Match by amount + date (±2 days) + vendor name. List unmatched charges (no receipt) and unmatched expenses (no card charge). Group by cardholder.
Under the hood: Claude pulls card transactions (from your bank feed or a card MCP server) and expense records (from ExpenseMonkey's MCP server), runs a fuzzy match, and gives you a clean exceptions list. No more eyeballing two spreadsheets side by side.
Time saved: A weekly reconciliation that used to take 60–90 minutes drops to 5–10 minutes of reviewing exceptions. For most finance teams this is the single biggest weekly time return.
Workflow 6 — Cross-tool reporting that doesn't lie
The most under-appreciated MCP benefit: the AI doesn't see ExpenseMonkey in isolation. With other MCP servers connected — Slack, Gmail, Google Calendar, your project tool — you can ask questions that span systems.
> For each project tagged #client-acme in our project tool, pull the total expenses from ExpenseMonkey and compare to the project budget. Highlight any project over 80% of budget.
> Build a travel report for the leadership offsite: cross-reference calendar events tagged "Offsite Q2," expense records dated within that window, and the travel policy. Output a one-page summary with anomalies called out.
Under the hood: Claude is the only place that needs to know how all your tools fit together. Each MCP server speaks for its own data. The reasoning step — joining a calendar event to an expense to a policy — is the AI's job, not yours.
Time saved: Highly variable. Some reports that took half a day to assemble are now sub-minute. Some reports that used to be impossible are now trivial. The latter is the more interesting category.
Workflow 7 — Month-end close prep
If you only adopt one workflow, make it this one. The close is where small ambient pains become a single, large, end-of-month migraine — and where MCP earns its keep.
> It's the 28th. Prep our close pack: summary of total spend by category vs last month, list of expenses still pending approval, list of expenses awaiting receipts, list of likely accruals (recurring vendor missing this month's bill), and any policy exceptions from the last 30 days.
Under the hood: A single MCP-powered query produces a draft of every artifact your accountant usually chases for. You review and refine instead of assembling from scratch.
Time saved: Most close prep is information retrieval, not judgement. Reclaiming the retrieval half typically gives a small finance team a full extra day per month.
Bonus: Pair this with a recurring scheduled Claude task on the 28th of each month, and the draft close pack lands in your inbox before you've poured your first coffee. Real-world ExpenseMonkey customers using this pattern report close cycles compressing from 5 days to 2.
Setup checklist (10 minutes)
- Get an ExpenseMonkey account. If you don't already have one, start at expensemonkey.io. The free tier covers solo users; the Team plan unlocks MCP and multi-user permissions.
- Generate a Claude / MCP connection token. Settings → Integrations → Claude. Scope it to read-only for your first week.
- Install Claude Desktop. Available from claude.ai/download. Web Claude works too if you prefer browser-only.
- Add ExpenseMonkey to Claude's MCP servers. Paste the token. ExpenseMonkey will appear in Claude's tool picker.
- Run a sanity-check prompt. Try Workflow 1 with a question you know the answer to. If the numbers match your dashboard, you're good.
- Promote to write access. Once you've stress-tested read mode, flip the token to allow writes. Confirm the per-call consent prompts behave the way you want.
- Onboard your team. Share the seven workflows above as a Loom or in your wiki. Most teams adopt MCP fastest when there's a "first prompt to try" already in front of them.
Five pitfalls to avoid
The workflows above are battle-tested, but a few patterns reliably create pain. Side-step them.
- Don't grant write access on day one. Start in read mode. Build trust. Promote when you've watched the AI behave on your real data for a week.
- Don't skip the audit log. Every team that gets serious about MCP eventually wants to know "who asked what, and when." ExpenseMonkey logs every MCP call by default — review it weekly until you stop being curious.
- Don't paste sensitive secrets into chat. The whole point of MCP is that secrets stay inside the server. If you find yourself pasting API keys or customer PII into the conversation to "help" the AI, your integration is missing a tool — open a support ticket.
- Don't over-prompt. "Categorise these expenses correctly" beats a three-paragraph instruction. The MCP tool definitions are descriptive enough that the AI usually knows what to do. Add detail when accuracy drops, not preemptively.
- Don't forget the human review step. AI is excellent at the boring middle and reasonably good at judgement. It is not infallible. Build a habit of approving categorisations in batches; don't auto-apply blindly. The accounting AI safety patterns documented across the industry all converge on this point.
Why this works when previous "AI for expenses" did not
If you've tried AI-flavoured expense tools before and felt like you'd been sold a search bar with extra steps, that's fair. Pre-MCP, "AI for expenses" usually meant a vendor-built chatbot trapped inside their app, with no access to anything else you use. The reasoning was capped at whatever the vendor decided to expose.
MCP inverts the model. You bring your own AI client. You bring your own MCP servers — ExpenseMonkey, Slack, Google Calendar, GitHub, anything. The AI reasons across all of them, in plain English, with you in the loop. This is why Anthropic's finance agents announcement reads less like a product launch and more like an architectural shift.
If you want a slower, more strategic perspective on what this means for the CFO stack as a whole, our companion piece on why MCP will reshape the CFO stack by 2027 goes deeper. If you want to know what MCP is in the first place, start with our plain-English guide.
Further reading: how MCP works with QuickBooks, FinOps Foundation's MCP working group, and the open MCP servers on GitHub.
