Generative AI use cases in finance: Nicolas Boucher's 5 examples

You've probably moved past experimenting with AI. Now the question is which tools actually save you time on real finance work. Generative AI uses large language models to produce original outputs like benchmark tables, research reports, and scenario analyses from prompts and existing data. It doesn't retrieve pre-written answers. It builds new ones. The practical result: work that used to take a week can now take 14 minutes, and work that required a designer can now take 2.

The short version

Benchmark hunts eat a week. Tariff analysis goes stale before the deck is built, and categorizing a column of credit card transactions is the unspoken Friday tax on every controller's calendar. You've probably used ChatGPT for quick lookups but still run the heavy work by hand. That's usually because no one has shown you which tools to trust on the work that actually matters.

Nicolas Boucher, founder of the AI Finance Club, joined Ramp's Dave Wieseneck to walk through five tools you can put to work this quarter. He live-built a SaaS benchmark report, generated a slide from it, ran a tariff impact analysis across alternative sourcing countries, demoed Microsoft Copilot accessing files and emails inside the Microsoft stack, and used the new =AI formula in Google Sheets to auto-categorize transactions.

Three ideas connect everything Nicolas covered: pick the right model for the task, structure your prompts so a CFO can defend the output, and verify every result like it's a junior analyst's first draft.

Most useful for: Controllers, FP&A leads, and CFOs who already use ChatGPT for quick lookups and want to move it into research, slides, and spreadsheet workflows.

About the speaker: Nicolas Boucher is the founder of the AI Finance Club. He teaches AI-powered finance workflows to a community of CFOs and a network of 1M+ followers on LinkedIn, drawing on his finance experiences like audit work at PwC and a senior role at a French international group focused on defense and aerospace.

What is generative AI in finance?

Generative AI in finance is the use of large language models like ChatGPT, Microsoft Copilot, and Google Gemini to produce original outputs (analyses, reports, slides, variance narratives, formulas) from prompts and existing data. Where traditional automation moves data between systems, generative AI creates new content based on the brief you give it.

The practical result for a finance team is that work that used to take a week now takes minutes. Benchmark research, tariff impact analyses, variance commentary, and transaction categorization are all tasks where the bottleneck has historically been time, not skill. Generative AI compresses the mechanical work without removing the judgment step.

The use cases split into two broad audiences:

• Corporate finance team applications include FP&A research, financial modeling, variance commentary, expense categorization, and slide design
• Financial services applications include automated compliance reporting, fraud detection, customer service chatbots, and investment research at scale

Nicolas Boucher's five tools in this webinar focus on the corporate finance side, with live demos of what works in a real workflow and where the verification step has to stay manual.

The most common generative AI use cases in finance

Generative AI shows up across six concrete areas of finance work today. Four of them are where Nicolas spends the bulk of the webinar. The other two sit more squarely in the financial services industry's domain.

• Document review and summarization: Reading contracts, regulatory filings, credit memos, and long internal reports at speed, then producing summaries a finance lead can act on. Nicolas's deep research demo below is one practical version of this category applied to benchmark research, and Microsoft Copilot's connection to your actual files (use case 5 below) extends it to internal documents.
• Market research and competitive intelligence: Pulling industry benchmarks, competitor analyses, and policy impact reports from public data. Nicolas's SaaS benchmark report and tariff analysis (use case 1 and 3 below) are both examples of this category run end to end, including the verification follow-up that makes the output defensible.
• Variance and forecast commentary: Drafting variance narratives, scenario explanations, and board commentary from the underlying numbers. The prompt anatomy Nicolas uses for image generation (use case 2 below) applies directly to narrative drafting as well, since both rely on a clear brief that gets the model to a presentable first draft.
• Spreadsheet and reconciliation automation: Auto-categorizing transactions, classifying vendors, and populating routine fields that historically demanded manual review. Nicolas's =AI formula demo below is the native Google Sheets version, with equivalent functionality available in Excel.
• Risk monitoring and fraud detection: Real-time anomaly detection on transaction data and flagging unusual patterns at scale. This sits outside the workflows Nicolas walks through in this workshop. It's where the financial services industry has invested most heavily and where most enterprise vendors have built dedicated products.
• Customer-facing financial services: AI advisors, automated loan underwriting, and personalized portfolio recommendations. More common in retail banking, insurance, and wealth management than inside a corporate finance team.

The five tools Nicolas demos are the hands-on version of the first four categories. The last two (risk monitoring and customer-facing services) belong to a different audience and a different tooling stack, but they're worth knowing exist when someone asks about generative AI in finance more broadly.

What does AI for finance actually look like in practice?

AI works best in finance when you treat it like a junior analyst. Give it a clear brief, review its work, and own the final output. Maybe you've used ChatGPT to summarize a memo or write a formula, and the leap from one-off lookups to real workflows is where Nicolas focused.

His framing is supervised delegation, not autopilot. Every AI run is a junior analyst who shows their work, which means you give them a clear brief upfront, answer their clarifying questions before they research the wrong vertical, read the output critically, and challenge anything that feels off.

Your expertise is what turns AI output into something you can put in front of a CEO, because you know what a credible source looks like and what a typical EBITDA range is for your segment.

Every tool below follows this same principle. Each one compresses mechanical work dramatically, but none of them removes the judgment step. As Nicolas put it, if you've never worked in SaaS, you take a real risk presenting deep research output to your CFO because you can't judge whether the report is credible. If you've done the work the slow way for years, you can validate it in minutes.

The question isn't what these tools do. Vendor websites have covered that for 2 years. It's why you're still getting mediocre results from them. Nicolas's argument cuts straight to it: the tools are ready, the workflows aren't, and the five use cases below are where that gap closes fastest.

• Building benchmark reports: You can run dozens of searches and get a fully cited benchmark report in minutes. That's work that used to take days of digging through industry surveys and analyst PDFs.
• Turning data into presentation-ready slides: You can convert a raw data table into a polished slide on demand, with no designer needed.
• Analyzing tariff and policy changes: You can analyze tariff changes, trade policy shifts, and compliance updates across multiple sourcing scenarios in a single prompt session, then verify the results with a follow-up.
• Automating spreadsheet work: Native AI formulas in Google Sheets (and equivalent functionality in Excel) let you auto-categorize transactions, classify vendors, and populate fields you currently handle manually.
• Answering questions from your own files: Tools like Microsoft Copilot connect generative AI to your actual files, emails, and calendar, so you can answer operational questions that require internal data, not just the public web.

Nicolas's 5 generative AI use cases for finance teams

1. Deep research compresses a week of work to 14 minutes

In Nicolas's live demo, a single prompt for SaaS financial benchmarks ($10M–$50M ARR) ran 83 searches across 30 sources in 14 minutes. It returned a fully sourced table with median, top-quartile, and full-range figures.

You know the work it replaces: scrolling Google for industry reports, asking your network for survey PDFs, and manually compiling a comparison table. That's a multi-day cost you've probably just accepted.

To run it, open ChatGPT, activate deep research, and select o3 for the most thorough results. Write the prompt with role, ask, and target metrics, then answer the clarifying questions it surfaces (vertical, geography, top-quartile vs. full range) and let it run.

The right-hand panel documents every search the model ran, in order, so you can audit the work the way you'd audit a junior analyst's report. Nicolas says you can't skip the verification step. In the demo, David recognized KeyBank and OpenView reports inside the source list and could vouch for them in seconds because he'd used those reports in past work.

"Even if it looks really smart, you need to be here, owner of the output. And if you own the output, it means that you have verified it. That's like something where we are, as human, quite lazy, and we are going to take risk if we don't here verify all of the work because we have to think that it's just a tool. It's not going to replace us."

2. Image generation is a finance designer on demand

You can turn a benchmark table into a presentation-ready slide in about 2 minutes using ChatGPT's image generation. You probably produce dense, table-heavy slides because that's how the underlying data looks, and your audience checks out the moment the slide goes up. Your design team isn't building a one-off chart for an email, so the slide stays ugly.

Nicolas's prompt structure is format, content, visual aids, style in that order:

• Format: names the canvas ("a 16:9 slide")
• Content: names the data ("paste the screenshot, name the columns")
• Visual aids: name the layout ("table with icons for each benchmark category")
• Style: names the look ("modern, minimalistic, brand colors blue and green")

This sequence consistently produces slides that fit straight into PowerPoint or Google Slides on the first generation.

When you spot a typo or a wrong number in the rendered slide, use the selection tool in the top-right of the image. Highlight the specific area and prompt a targeted change so the rest of the slide stays untouched. In the demo, Nicolas fixed a wrong range on the benchmark slide without rebuilding the whole image, and this is the feature most people miss, which is what turns image generation from a novelty into a real workflow.

"In finance, our big problem is that we are really good at presenting tables, but they always look ugly. They are, like, really hard to digest. And it's because maybe we don't have the genes for visuals. Well, now it's good because we have actually a designer available for us."

If you want to go further with turning finance data into executive-ready visuals, the Ramp webinar How to use AI to turn Raw Data into Executive-Ready Insights covers the full data-storytelling workflow for CFOs and FP&A teams.

3. Always run a verification follow-up on deep research

Your first deep research pass is never your last. In the tariff demo, Nicolas's first pass produced a report covering current tariffs, recent policy changes, and alternative sourcing countries including Mexico, Vietnam, India, and Thailand. The model pulled from Reuters, the Office of the United States Trade Representative, trade.gov, and auto industry trade publications. Reading it, he sensed something was missing.

One follow-up prompt changed everything. When Nicolas asked the model to verify against the latest news, it ran another round of searches, surfaced policy updates the first pass missed, and recalculated the impacts. The original pass took 14 minutes and 167 searches, and the follow-up changed the conclusion.

Nicolas says the verification follow-up is what makes the difference between a one-off experiment and a workflow you actually trust. The first answer is never the last answer, so build the second prompt into your workflow from the start. On tariff work specifically, where information changes every few days, skipping the follow-up means you're presenting stale numbers to a CFO who's been reading the news.

"That's why I want everybody to be aware. Don't take the first answer as the last one and the best one. Review it, challenge it, proof check it, and then you can use the work if you are sure that you can use it."

4. GPT-4o for daily speed, o3 for high-stakes work

GPT-4o answers in seconds and handles most day-to-day finance prompts, while the o3 reasoning model takes 2–50 minutes but produces output Nicolas estimates is 10–20x better on complex tasks. They're not interchangeable, and using the wrong one is the most common mistake Nicolas sees from people new to ChatGPT.

Nicolas's rule: use o3 for cost reduction strategy, multi-variable models, deep benchmarks, and tariff analysis. Basically, reach for o3 any time you'll need to defend the output in front of a CFO or board, and use GPT-4o for everything else.

If a GPT-4o answer feels thin, it's probably a model mismatch, not a tool limitation. Rerun the same prompt through o3 and wait, because most "ChatGPT isn't that useful" stories Nicolas hears trace back to running an o3-grade question through GPT-4o and judging the tool by the wrong output.

"GPT-4o is for everything that is really quick, where you don't want to wait, you want a quick answer, which will be ninety five percent of the time always good. Then for more complex topics where you want the AI to think about it, let's imagine that you have to work on cost reduction. It's not something you want to get answer like this in ten seconds. You want to find the best strategies. Well, that is where o3 is the best tool for it."

If you want to apply these outputs to FP&A planning work, the Ramp webinar How to Build Scenarios Like a Wharton Program FP&A Leader covers how finance teams use AI for scenario planning in a way that builds directly on the model-selection and benchmark research workflows Nicolas covers here.

5. Copilot's edge is access to your actual work

ChatGPT doesn't know your emails, your calendar, or which Excel file you opened this morning, but Copilot, on a business Microsoft 365 license, does. For independent research and image generation, ChatGPT still wins, while for the daily question "which product had the highest revenue today," Copilot gets you the answer because it can see the file.

In the demo, Nicolas opened Copilot and toggled to Work mode. The alternative, Web mode, behaves like ChatGPT for external research. He used a slash command to reference a specific file, /SKU supermarket, and asked the question in plain language. The answer came back in seconds with a clickable link directly into the source file, and the same applied to email when he pulled up his last received messages, summarized one, and extracted action items inside the Copilot interface.

If you're already on the Microsoft stack, the real value isn't the AI itself. It's the AI plus the data it already has access to. Nicolas recommends pushing for the Copilot license before you add another standalone AI tool.

"The problem with ChatGPT is that it doesn't know anything about your work. It doesn't know anything about you. Or even with memory, it knows a bit, but it doesn't have access to your emails. It doesn't know your calendar. It doesn't know what you worked on today, which Excel file. That's really the difference between having just an AI which knows everything outside of your work or having AI which has access to all of your information at work."

How do you apply Nicolas's advice this quarter?

Pick one of these steps, not all three. The point is to get one workflow off the manual track this quarter and judge the results before scaling further.

1. Run one =AI formula on a real dataset. In Google Sheets, type =AI("your prompt instructions", target_cell). Pick a column you'd normally categorize manually (vendor types, expense buckets, state abbreviations). Generate one cell, verify the output, drag down. Excel users can access equivalent functionality through the Excel AI Lab. Hand the result to your controller and ask if it saves them 1 hour a week
2. Run one deep research benchmark using the o3 model. Pick a question your CEO has actually asked. Industry margins, sales efficiency benchmarks, working capital ratios. Write a prompt with role, ask, and target metrics. Answer the clarifying questions. Read the output critically and verify 2 sources before you share it
3. Generate one slide with ChatGPT image generation. Take a table you'd normally hand to a designer. Use Nicolas's four-part structure (format, content, visual aids, style). Drop the result into your next deck

If you want to apply these tools to dashboards, Nicolas walks through 3 live builds in the Ramp webinar Build 3 AI-Powered Dashboards in 60 Minutes. And if you want to pair these AI workflows with a broader financial resilience strategy, The Cash Flow Playbook: How to Engineer Financial Resilience Before Your Business Needs It covers the planning work that makes your analysis actionable.

Final thoughts

"Everybody here is actually already in the top percentage of people who understand how to use AI in finance. Finance is a great place to become the AI champions of business."

You'll get the most out of these tools not by writing better prompts but by treating every AI output the way you'd treat a junior analyst's first draft. Read it, challenge it, verify the sources, and own the version that goes upstairs. The tools compress the mechanical work, but the judgment is still yours.

See how Ramp fits in

The five tools Nicolas demoed cut your research and analysis time. The other half of your time goes to expense categorization, invoice approvals, and reconciliation. You can automate all of it with Ramp before the data ever hits your spreadsheet. You'll move faster on analysis when Ramp has already coded, approved, and synced your transaction data to your general ledger.

Common questions about generative AI in finance

What is generative AI in finance?

Generative AI in finance is the use of large language models like ChatGPT, Microsoft Copilot, and Google Gemini to produce original outputs (analyses, reports, slides, variance narratives, formulas) from prompts and existing data. Unlike traditional automation that moves data between systems, generative AI generates new content based on the brief you give it.

For corporate finance teams, the highest-leverage applications today are benchmark research, slide generation, scenario modeling, and spreadsheet automation. Nicolas Boucher's five tools above are the practical hands-on versions of each.

What are the most practical generative AI use cases in finance right now?

The five Nicolas demoed (deep research benchmarking, AI-generated slides, policy impact analysis, model-aware prompt routing between GPT-4o and o3, and spreadsheet automation via =AI) are the best places to start for most finance teams. They all share one trait: they compress mechanical, repeatable work without removing your judgment.

The use cases that tend to disappoint are the ones where you can't verify the output against a known standard, which is why Nicolas's supervised delegation approach matters for all of them.

What are the benefits of using generative AI in finance?

The three benefits most finance teams see in the first quarter of structured use are time savings on mechanical work (research, slide design, transaction categorization), broader analysis (running scenarios you wouldn't have had time to run manually), and stronger communication (variance commentary, board narratives, and visual slides that used to need a designer).

The trade-off is the verification overhead Nicolas emphasizes throughout the workshop above. Every output needs a human review, which keeps the judgment with the finance team while the mechanical work goes to the tool.

Which ChatGPT model should we standardize for the team?

Both. Nicolas pushed back on the idea that one model fits every workflow. GPT-4o handles roughly 95% of daily prompts in seconds, while o3 takes longer but produces noticeably better results on complex or high-stakes work. The key is getting your team comfortable switching models based on the task, not habit.

How do we know we can trust deep research output?

Verify 2 sources before you share anything. Nicolas's process is to scan the right-hand panel listing every search the model ran and look for at least one trusted publisher per claim, which in SaaS might be KeyBank or OpenView and on trade policy might be Reuters or USTR. Then read the linked source for any number that drives a recommendation, and run a follow-up prompt asking the model to check against the latest news, because the first answer is never the last answer.

How do smaller finance teams without dedicated AI resources get started?

The same way larger ones do, one workflow at a time. Nicolas's supervised delegation model doesn't require a dedicated AI team or a formal implementation project, but rather one person who knows the domain well enough to verify the output.

For a two-person finance function, that's usually the controller or the CFO themselves. Start with the =AI formula on a transaction categorization column, or run one deep research benchmark on a question your CEO has already asked. The verification step is the same regardless of team size, and only the coordination gets lighter.