MCP and Claude for Finance: How One CFO Stopped Waiting for Answers on Her Own Numbers

MCP — Model Context Protocol — is the piece of AI plumbing that finally lets finance teams stop exporting, stop waiting, and start asking.

In a recent CFO Connect masterclass, Pauline Babell (CFO at Spendesk) demonstrated to nearly 500 live attendees what changes when Claude for finance is wired directly into live spend data via MCP. Below is the full playbook: her exact demos, the prompts she used, and a step-by-step framework to get started today.

▶️ Watch the full session: AI in Finance Masterclass — Stop Waiting for Answers on Your Own Numbers

TL;DR

• MCP connects Claude directly to your live finance data — no manual exports required

• Spendesk's CFO reduced three core reporting workflows from hours to under eight minutes each

• Best first use cases: travel compliance, OPEX review, AP aging

• Spendesk MCP is read-only by design — the AI surfaces, you decide

• Start with one workflow, build it once, then schedule and automate it

Key Takeaways

• MCP gives Claude and other AI assistants direct, real-time access to your live Spendesk data — no more manual exports

• Three core workflows — travel compliance, budget vs. actuals, and AP aging — each took hours before MCP and now run in under eight minutes

• Spendesk MCP is read-only by design: the AI surfaces the insight, you make the decision

• Once built, workflows can be scheduled to run automatically (e.g. every fifth of the month, every Monday morning)

• The productivity gain compounds: build it once, schedule it, automate it

• The prerequisite is data consolidation — AI cannot fetch data it cannot see

• Start with one workflow, not a full transformation

1. What is MCP and why does it matter for finance teams?

MCP is the reason Claude for finance finally works the way you always hoped AI would.

MCP — Model Context Protocol — is an open standard that gives AI assistants like Claude a direct, real-time, read-only connection to your business tools. Instead of downloading a CSV, uploading it to Claude, and hoping the numbers are current, your AI assistant queries your live data directly and answers in plain language.

It is, as Pauline puts it, the difference between AI that can talk about your business and AI that can touch it. The protocol was developed by Anthropic and has since been adopted by OpenAI, Google, and Microsoft — making it an industry-wide foundation rather than a single vendor's bet.

What to do this week

Read the Spendesk MCP explainer and forward it to one person on your team who manages your finance tech stack. Ask them: which of our tools has an MCP server available?

2. Why are most Claude finance skills still failing to deliver?

The gap between AI promise and AI usefulness in finance has one root cause: disconnected data.

The average finance team uses seven different disconnected tools. Each one holds a fragment of the truth. None of them talk to each other. And the people caught in the middle — your best analysts — spend hours every week reconciling systems instead of generating insights.

For years, my week started the very same way. Open Excel, open my ERP, export the data, open the spend tool, export again, then perform VLOOKUPs, pivot, reformat everything. Maybe three hours later, I had a number I could actually trust enough to send to the board. That number was already a week old by the time anyone read it. — Pauline Babell, CFO, Spendesk

This is not a one-company problem. It is a European CFO problem. According to the Grant Thornton UK Finance Leaders Barometer 2025, 77% of UK finance leaders say data volume clouds rather than clarifies judgment. The ABV Group Baromètre des DAF 2025 found that 63% of French finance leaders see AI as promising — but only 39% find it currently useful in daily work. The Deloitte Germany CFO Survey, Autumn 2025 found 66% of German CFOs plan to automate both transactional and knowledge-based processes, yet execution lags significantly behind ambition.

The reason most Claude finance skills underdeliver is structural, not technical. AI cannot fetch data it cannot see.

The wait is the cost. The wait is what makes us reactive instead of strategic. The wait is what burns out our best people on data plumbing. — Pauline Babell, CFO, Spendesk

What to do this week

Map your finance tech stack. For each tool, note: where does the data live, who owns the export, and how long does it take to get a number you trust? That map is your MCP readiness audit.

3. What is Spendesk MCP and how does it connect to Claude?

Spendesk MCP turns your spend data into a conversational finance experience — built for speed, trust, and everyday decision-making.

Spendesk MCP is a server that connects AI assistants — Claude, Dust, and others — directly to the Spendesk Public API. When configured, your AI assistant can query live spend data, payables, suppliers, purchase orders, cost centres, and settlements in plain language, in real time, without downloading a single file.

There are three ways Spendesk makes consolidated spend data accessible:

• Native integrations — certified connectors to your ERP, HRIS, and productivity tools, no engineering ticket required

• Open APIs — custom-built connections for selected data sets, requires developer support

• Spendesk MCP — plain-language conversational access through Claude and Dust, set up in minutes

It's literally a copy-paste of an HTTP address. And then the MCP is connected. You can see Spendesk MCP in the skills. And then I can ask a simple question, plain English, conversational style. — Pauline Babell, CFO, Spendesk

Spendesk MCP is read-only by design. It can surface, query, and analyse your data. It cannot create payments, approve invoices, or modify anything. The AI surfaces the information. You make the decision. Data access is also scoped by your existing Spendesk user permissions.

What to do this week

If you are already a Spendesk customer, check whether your account has MCP access enabled. The setup takes under five minutes.

4. The MCP Finance Workflow Stack

Before you connect anything, understand the five stages that make MCP work.

Step 1 — Consolidate Get your spend, payables, and supplier data into one clean, structured place. This is what Spendesk does. Without this step, the rest does not work.

Step 2 — Connect Wire Spendesk to Claude via MCP. One HTTP address. Five minutes. No engineering ticket.

Step 3 — Query Ask questions in plain language and get real-time answers from your live Spendesk data. No exports. No pivots. No waiting.

Step 4 — Schedule Set the same report to run automatically — for example, on day five of every month. It lands in your inbox without you lifting a finger.

Step 5 — Automate Stack workflows. Each one you add compounds the time saved. The fifth workflow costs a fraction of the effort the first one did.

Most finance teams trying to adopt AI are jumping straight to Step 3. The ones getting results have completed Steps 1 and 2 first. Sherilyn Kamga, founder of SmartPlans, frames the goal well:

That's the difference between automating a task and automating an entire process. No one reopens the spreadsheet, no one reruns the script, nobody pings anyone on Slack to ask if something is done. The loop closes itself automatically. — Sherilyn Kamga, Founder, SmartPlans

Consolidation is the prerequisite, not the optional extra.

5. Workflow 1 — Travel and Expense Compliance: From 4 hours to 3 minutes 30 seconds

Travel compliance is one of the highest-volume, most time-consuming monthly finance tasks — and one of the clearest early wins for Claude for finance with MCP.

Before MCP, Pauline's travel compliance workflow involved exporting data from Spendesk, cross-referencing the company travel policy, building a pivot, and formatting the output for the leadership team. Start to finish: approximately four hours — every single month.

The prompt she used: Generate travel policy compliance report, April 2026.

Time to output: 3 minutes 30 seconds.

The output is a dynamic HTML dashboard containing a policy compliance overview by payable type, a full list of flagged violations with payable ID, date, and amount, cost centre and employee-level breakdowns (filterable), suggested remediation actions already drafted, and month-on-month variance automatically calculated.

I can focus my time analysing it rather than crunching the data. It's already flagging the variances the way I want to disclose them. I can drill down to each violation in a snapshot. — Pauline Babell, CFO, Spendesk

The follow-up question is where the real power shows. Instead of rebuilding the report, Pauline simply asks: "Which cost centre has the highest rate of violations?" The conversation continues without the report being rebuilt. Once the workflow is built, it can be scheduled to drop into her inbox every Monday morning — automatically, without her lifting a finger.

Sherilyn Kamga puts it well: analysts usually have that one routine task that's 80% data prep and 20% actual analysis. MCP flips that ratio entirely.

What to do this week

If your team produces a monthly compliance or policy adherence report, document the current steps: where does the data come from, how long does it take, who builds it? That is your MCP candidate.

6. Workflow 2 — Budget vs. Actuals: From a week-long cycle to four minutes

The OPEX review is the meeting every finance team dreads preparing for. MCP removes the preparation entirely.

The traditional OPEX review cycle involves FP&A building a Google Slides deck, the CFO reviewing it, revisions going back and forth, and numbers aging before the conversation even happens. One week of preparation for a sixty-minute meeting.

The prompt Pauline used: Generate budget report for Q4 2025.

The Spendesk Budget Analyser MCP agent compiled a full dynamic dashboard in four minutes, including budget vs. actuals with variances in value and percentage, unplanned spend flagged, cost centre breakdown drillable without leaving the page, and risks and action plan by cost centre already drafted.

I'm able to drill down without ever opening Spendesk or NetSuite — from cost centre, to expense category, to individual contractor. All on one page. That level of granularity with that level of fluidity. That's where I find it powerful and beautiful. — Pauline Babell, CFO, Spendesk

The scheduling capability is the compounding factor. Build the workflow once, set the trigger for day five of each month, and it runs without human intervention.

No more Google Slides for OPEX review where it's pending review. That's powerful. — Pauline Babell, CFO, Spendesk

For a broader view of how AI is reshaping spend planning at the strategic level, read Spendesk's piece on SP&A in the AI era.

What to do this week

Take your last OPEX review deck. Count the hours it took to build. That is the number MCP eliminates from your monthly calendar.

7. Workflow 3 — AP Aging: Full overdue exposure in 7 minutes 20 seconds

Accounts payable aging is one of the most consistently painful month-end tasks — multiple systems, manual compilation, and a final number often already out of date by the time it is shared.

The traditional AP aging workflow requires logging into multiple systems, cross-referencing payment data, building the aging bucket structure, and formatting it for review. For most finance teams, several hours and at least two people.

The prompt: AP aging report — overdue exposure this week.

Time to output: 7 minutes 20 seconds.

The output included total overdue exposure broken down by aging bucket, the number of open invoices by risk level, cash curve visibility based on scheduled payables, and a prioritised action list with the highest-risk items first. Build it once and schedule it — the same compounding effect as every other workflow.

Alex Altman, a finance leader building Claude-powered automations in the CFO Connect Live Build series, described a near-identical breakthrough when he connected Claude to QuickBooks for revenue recognition:

"The old way would take me anywhere from four to six hours a month. It was the last thing I wanted to do every month — I was always the one holding up close. Once I was comfortable, this is now the only way I do things. It's literally button clicks." — Alex Altman, Head of Finance, Coram AI

For a deeper look at why traditional AP automation often relocates manual work rather than removing it — and how MCP changes that — read Spendesk's guide to accounts payable automation.

What to do this week

Time your next AP aging report from data pull to final formatted output. That number is your baseline.

8. What does "read-only by design" actually mean for security-conscious finance teams?

The biggest concern finance leaders raise about AI and live data is the one Spendesk MCP was built to address from the start.

Spendesk MCP cannot create payments, approve invoices, modify records, or take any action in Spendesk — it retrieves and analyses only. The MCP also only transmits data your user account can already access; no elevated permissions are granted via the AI.

When using Claude Enterprise (the recommended tier for finance teams), Anthropic contractually commits not to train on your company's data. The trust is built into the architecture, not added as an afterthought.

Kevin Steele, who built a full FP&A forecasting app using Claude Code with no prior coding experience, makes the same point about human oversight:

You need to check and verify everything on an iterative basis. If you're doing this for clients, it's got to be right — it cannot be wrong. People are making business decisions with your information. — Kevin Steele, Founder, Inflective Intelligence

For CFOs thinking about the broader human dimension of AI adoption, Spendesk's piece on whether AI will affect finance careers addresses this in practical terms.

What to do before you connect MCP

Confirm your AI assistant subscription tier. Confirm the data you plan to query sits within your existing permission scope. Have a five-minute conversation with your security team. That is the full checklist.

9. What has to be true before MCP works? The consolidation prerequisite

MCP is not a shortcut around data consolidation. It is the reward for doing it.

If your spend data is fragmented across five tools with no single source of truth, connecting an AI assistant will not produce reliable answers. It will produce fast, confident, wrong answers.

Consolidation comes first. You cannot connect what is not consolidated and clean enough to get AI plugged on. You need to map your tech stack and don't procrastinate it. — Pauline Babell, CFO, Spendesk

The sequencing matters: consolidate first, then connect, then query, then schedule, then scale. According to the Kyriba OPR Index 2026, only 33% system connectivity was reported across Spanish finance teams — a figure that illustrates how structural this problem remains across European markets.

Six months ago I was really at level zero. We were running these workflows every week, every month — super painfully. And now this is a game changer. We are genuinely gaining time on getting the data ready to focus on analysis. — Pauline Babell, CFO, Spendesk

For CFOs earlier in their AI journey, Spendesk's piece on one CFO's path from AI sceptic to believer maps this progression in practical terms.

What to do this quarter

Score your current data consolidation on a scale of one to five: how confident are you that your spend data is clean, current, and in one place? If the answer is below four, that is where to focus before configuring MCP.

10. What is the compounding effect — and why does it matter for European CFOs?

The real value of MCP is not the time saved on one workflow. It is what happens when you build, schedule, and automate five of them.

When workflows run automatically on day five of each month without anyone opening a laptop, the savings are not just time. They are cognitive load, coordination overhead, and the management tax of chasing numbers before every leadership meeting.

With MCP, you can build it, then schedule it, then automate it. That is the compounding effect of it. — Pauline Babell, CFO, Spendesk

Cristian Sanford, co-founder of Quantify and fractional CFO to businesses across the US and Europe, frames the strategic direction clearly:

Really where we're heading is a departmental transformation where the talent structure looks different. The controller is still the controller. The CFO is still the CFO — but a lot of that execution work starts to move around and be automated. And it's really interesting to think about the acceleration over the last six months and where we're going. — Cristian Sanford, Co-founder, Quantify

According to Gartner's 2026 CFO Technology Survey (300+ CFOs), 75% of CFOs are raising tech budgets this year, 88% rank finance staff productivity as a top-three priority, and headcount growth is collapsing from 6% to just 2%. The output expectation is rising. The headcount to deliver it is not. MCP is one of the few levers that changes that equation without adding people.

What to do this quarter

Identify your top three recurring finance workflows by time cost. For each, ask: could this be prompted once and then scheduled? Start with the one where the answer is most obviously yes.

11. A practical 60-day plan for getting started with MCP and Claude for finance

You do not need a full transformation programme. You need one workflow, one prompt, and one month.

Start small. Pick one workflow that costs your team the most hours — compliance, budgeting, anything — and start from that. — Pauline Babell, CFO, Spendesk

Kevin Steele started with zero coding experience six weeks before building a fully functioning FP&A forecasting app. Alex Altman puts the approach plainly:

"Start simple, build a block, build your foundation, and then take it from there." — Alex Altman, Head of Finance, Coram AI

Days 1–10: Map and consolidate Audit your finance tech stack. Confirm your spend data in Spendesk is clean and current. Identify your one highest-cost monthly workflow.

Days 11–20: Connect and configure Enable Spendesk MCP (copy-paste the HTTP address into your AI assistant settings). Run your first plain-language query against live data. Note what comes back.

Days 21–40: Build and schedule Design your first workflow agent. Run it once with human review at the output stage. Schedule it to run automatically next month.

Days 41–60: Review and expand Review the scheduled output when it lands. Note what it got right and what it missed. Refine the prompt. Identify your second workflow.

For a wider view of how AI is being adopted across European finance teams right now, download CFO Connect's State of AI in Finance 2026 report.

FAQ: MCP and Claude for Finance

What is MCP in finance? MCP stands for Model Context Protocol. It is an open standard that allows AI assistants like Claude to connect directly to business tools — including Spendesk — and query live data in plain language. For finance teams, it eliminates the manual export step between a question and a reliable answer.

What is the difference between uploading a CSV to Claude and using Claude for finance with MCP? A CSV is a static snapshot — potentially hours or days out of date. With MCP, Claude queries your live data in real time. Outputs are current, follow-up questions work without re-uploading, and workflows can be scheduled to run automatically.

Is it safe to connect Spendesk data to Claude via MCP? Spendesk MCP is read-only by design — it cannot create payments or approve invoices. Data access is scoped by your existing Spendesk permissions. Claude Enterprise does not train on your company's data. Consult your security team before connecting live finance data to any AI tool.

What claude finance skills do I need to use MCP? None beyond basic prompting. Setup is a copy-paste of an HTTP address. The learning curve is writing the initial workflow prompt, which improves with each iteration.

What is a good claude prompt for financial analysis to start with? Pauline's recommendation: describe your most time-consuming recurring report in plain language. For example: "Generate travel policy compliance report for [month]." Start with one output format, then add follow-up questions once the base workflow is running reliably. For 25 more prompts across FP&A, compliance, and controls, see CFO Connect's Claude prompts for finance teams.

Does Spendesk MCP work with tools other than Claude? At launch, Spendesk MCP supports Claude and Dust. Support for ChatGPT and other major AI assistants is planned.

What does "conversational finance" actually mean? It means asking follow-up questions without rebuilding the report. In Pauline's demos, she moved from a high-level cost centre overview to individual contractor spend in one conversation — without opening Spendesk, without exporting data, without leaving the page.

Closing thought: the wait is the cost

Finance has always been the function that turns data into decisions. The problem has never been access to data — it has been the time it takes to make that data usable.

MCP does not change what finance teams are trying to do. It removes the step that was slowing everything down: the export, the pivot, the reformat, the three hours on a Monday morning before the number is trustworthy enough to send to the board.

The wait is the cost. Every hour your teams spend pulling data is an hour in 2026 that is lost — not allocated to analysis, not to strategy, not to the decisions that actually matter. — Pauline Babell, CFO, Spendesk

The teams building these workflows now — one prompt, one agent, one scheduled report — will not be waiting for answers on their own numbers six months from now. The teams still exporting CSVs will.

Start with one workflow. Build it once. Schedule it. That is the compounding effect.