Master Excel Copilot's Latest AI Features with Christian Wattig
Speakers

Christian Wattig, Founder, Inside FP&A

Jaco Schoeman, Associate Director of Strategic Finance, Drivetrain
It’s the middle of October, and sales has booked only 20% of the month’s target. That’s a signal to act, and flagging it is the weekly pacing model’s job. To do that, the model has to calculate the percentage of the month elapsed, whether the month has 28, 30, or 31 days. Building that formula by hand takes an analyst time that could otherwise be spent working with the head of sales to figure out why the month is behind.
Christian Wattig, founder of Inside FP&A, expects AI to win back some of that time. In his survey of finance leaders, lack of time was the top reason they didn’t work more with cross-functional leaders.
He predicts that reconciliations, copy-pasting between spreadsheets, and producing standard reports like the P&L will largely disappear from finance work within a few years, and he said some companies are already using AI for that work.
In a Drivetrain webinar hosted by Jaco Schoeman, associate director of strategic finance, Christian showed how by handing parts of his Excel model-building to ChatGPT and Copilot.
This article will give you the four-part prompt structure Christian uses and the check he recommends for each output, from testing a formula against a known total to back-testing a forecast on months whose actuals you already have. We’ve also included the webinar Q&A and the full transcript at the end.
61:07Know the risks of machine learning and generative AI before you start
Before jumping into his demos, Christian offered his thoughts on the risks of using AI in finance and what he does to mitigate them.
Christian split AI into two flavors, generative AI and machine learning. Most of the session focused on generative AI, the technology behind ChatGPT, Gemini, and Claude, and Christian named three risks to manage before using it.
- Data security: Christian said tools like ChatGPT and Gemini can be configured so uploaded data stays on-premises and isn’t used for training. But you should confirm with your manager and IT that your company’s setup does that before uploading anything confidential.
- Hallucination: Models confidently make things up, so Christian recommends two defenses. Use a reasoning model, which thinks more slowly and runs built-in checks that reduce (but do not completely eliminate) hallucination. Ask the model for its references and click through each one, because a model can invent a citation, complete with a plausible title and a convincing link, for a source that doesn’t exist.
- Math: Language models often get basic counting and arithmetic wrong. Christian pointed to a classic example: asking how many R’s are in “strawberry,” which many models miss repeatedly.
Machine learning, which has been around for more than a decade, forecasts by finding trends in your historical results and projecting them forward. Christian compared it with judgment-based forecasting.
It's powerful for forecasting because its strengths and weaknesses are almost the mirror image of judgment-based forecasting: Human forecasts carry bias but capture nuance and context. Machine learning has no human bias, but it's entirely bound by the data you feed it. Quality issues or unusual one-offs in recent data get amplified rather than filtered out.
Because the two err in different ways, Christian said combining expert judgment with machine learning can be extremely powerful. Getting machine learning right also takes experimenting with its inputs. At a company Christian worked with, the machine learning models stopped performing well in mid-to-late 2020 and regained accuracy only after the team fed in economic data, unemployment figures, and shutdown-related data.
Prompt ChatGPT to write Excel formulas without giving it your data
In his first demo, Christian showed how to get working Excel formulas from ChatGPT without uploading a workbook. The method works without an enterprise-grade ChatGPT setup, and more importantly, your company data never leaves your hands.
He built the formulas for a weekly pacing model, which compares actual revenue with the monthly target to show whether the business is ahead or behind. The model needs two calculations that update when you change a date cell: month-to-date revenue and the percentage of the month elapsed.
Christian described the workbook’s layout in his prompt, including which rows held revenue and dates, and structured the rest using his favorite framework for ChatGPT for finance prompts. It comes from OpenAI’s Greg Brockman and runs in a fixed order: goal, return format, warnings, then context.
Christian said context needs the most effort, because the model is like a well-read intern that has absorbed the internet but knows nothing about your business. The finance teams he trains most often skimp on context and get back generic answers that are hard to act on, a problem Drivetrain’s fourteen ChatGPT prompts for finance addresses by building that context into prompts for common FP&A tasks.
His pacing prompt covered all four parts.
- Goal: “I’m a finance manager at a software company building a weekly pacing report in Excel to track actuals against monthly forecast. I need help writing the Excel formula.”
- Return format: “Provide formulas I can paste directly into Excel, with short explanations of each component.”
- Warnings: “Keep formulas as simple as possible for auditability, and verify they truly do what’s intended.”
- Context: “This is a high-visibility project for my CFO, someone I don’t work with often, and I want it to help me stand out.”
Christian also explained why a bad formula costs little.
[Formulas are] actually a good example of a low-hallucination-risk use case. If the formula it returns doesn’t work, nothing breaks; there’s no factual claim at stake. I just paste the error back and ask it to try again. And you’ll find that iterative back-and-forth gets you a working result even when the first attempt fails.
In his demo, ChatGPT’s first formula, a SUMIFS that totals revenue up to the selected date, checked out against total revenue and recalculated correctly when Christian changed the date cell. The percent-of-month formula ChatGPT returned was long and complex, so he checked its logic by hand: On day 15 of a 31-day month, the answer should be 15 divided by 31, or about 48%. The math held up, so he then asked ChatGPT for a simpler version and used that one.
Christian also recommended reading the explanation ChatGPT returns with each formula, because it teaches the Excel behind the answer.
Prompt Copilot’s agent mode to draft a three-year P&L from a handful of assumptions
Christian’s second demo showed how to get a working first draft of a financial model from one prompt, using Copilot’s agent mode, which builds an Excel model from the instructions you give it.
Agent mode goes beyond the Copilot features Carl Seidman, FP&A advisor and fractional CFO at Seidman Financial, demonstrated in his Drivetrain session, How to Master Excel Copilot for FP&A in 60 Minutes. At the time of Christian’s October 2025 session, agent mode was an experimental feature released only a few weeks earlier. It required a Microsoft 365 or paid Copilot subscription and ran through Add-ins > Excel Labs, with desktop Excel support expected later.
Christian’s prompt asked for a three-year P&L for a new product launch, a pro forma income statement built from assumptions written like meeting notes. The assumptions included a $500 target price, a cost of goods sold (COGS) at 14% of price, marketing spend, sales targets, headcount and associated costs, and general and administrative (G&A) expenses at 5% of revenue.
After about a minute of reasoning, agent mode returned a model that calculated revenue and gross margin, listed its assumptions, and linked the summary to the underlying calculations. Revenue, operating income, and the chart it created all appeared to be formula-driven, with nothing hard-coded.
At first glance it did a solid job, though before using this for anything real, I’d absolutely do the same diligence you’d apply to a model handed to you by an analyst: check every cell, verify the logic, because mistakes are still possible.
Prompt ChatGPT to write Python for an ARIMA forecast, then run it in Excel
In his third demo, Christian built a machine learning forecast without asking ChatGPT to do any math. He had ChatGPT write the forecasting code and ran it with Excel’s built-in Python. He called it a fairly advanced use case, one that builds on the ChatGPT techniques in Nicolas Boucher’s ChatGPT for FP&A masterclass.
The dataset for this demo held eight years of monthly retail deposits for a bank, with actuals through April 2025. Christian asked ChatGPT for a forecast of May through December 2025 using ARIMA (autoregressive integrated moving average), a model that accounts for seasonality and adjusts based on its own past errors. His prompt asked for the point forecast plus confidence bounds, the range where the actual outcome is likely to fall, and told ChatGPT to double-check that the code runs.
Christian selected a cell, chose Insert Python on Excel’s Formulas tab, and pasted the code into it. He said he isn’t a developer and can’t verify the code himself, so he recommends scanning it or asking another AI to explain what it does before running it.
Christian didn’t upload the dataset to ChatGPT. It stayed in Excel, where Excel’s Python engine ran the code on it with no AI involved.
During the demo, the code failed three times. The first attempt stalled without returning a result, and the next two returned errors. Christian said it usually takes two or three rounds of troubleshooting, feeding errors back into ChatGPT to get a successful result. So, in the interests of time, he stopped his demo there.
Later in the session, Christian reported that the revised code ChatGPT wrote in response to the third error ran successfully. The output showed the forecast with its 80% and 95% confidence bounds, the range where the actual outcome is likely to fall. He said the next step would be back-testing it against actuals already on hand.
Back-test the forecast against the actuals you already have
Back-testing shows how well a model would have forecast a period for which you already know the results. Christian’s method is to set aside your most recent actuals, for example, the last eight months, have the model forecast those months from the earlier data, and check whether the real results fall inside the forecast’s confidence bounds. When he tested the working code on his demo dataset before the session, he said the results were surprisingly accurate, though forecasts on more irregular data will vary.
Christian described running the back-test in a fresh ChatGPT chat with no history, using the shortened dataset. That step puts company data in ChatGPT, which the live demo avoided, so clear it with IT first.
Work around AI’s limitations with connected data
Christian pointed out that even when you successfully build a workflow with AI, there are other limitations you need to keep in mind. General-purpose AI tools still require manual data uploads, don’t support real collaboration, and don’t connect to live data, so reforecasting remains tedious.
Those limits get harder to work around as a company adds data sources, entities, and stakeholders to coordinate, which Christian said is the point at which AI-native FP&A platforms start to make sense. Christian asked Jaco how a purpose-built FP&A platform would handle these problems. And Jaco answered him with a demo of Drivetrain.
Refresh, reforecast, and run scenarios in one connected model
Jaco’s platform demo addressed two of the limitations Christian described: manual uploads and live data. He also named coordinating with stakeholders like the CMO or head of sales as a common planning challenge, and walked through what happens at month-end and when one of them asks for an alternative plan.
- At setup: Jaco connected two Xero instances, Google Sheets, and Salesforce, sources a team would otherwise export and upload by hand. He said Drivetrain offers more than 800 native connectors.
- During the month: Each connection automatically refreshes the data daily, weekly, or monthly on a schedule the team sets. Analysts can also refresh their data on demand with a manual sync, so the forecast runs on current numbers.
- When the books close: An analyst clicks Update on the closed month, and Drivetrain replaces its forecast with actuals. The months after it keep forecasting as before.
- When an assumption needs rethinking: An analyst can change how a line is forecast. The sales and marketing forecast, for example, starts from the company’s connected accounting data, and Jaco said it can be tied to a driver such as 20% of revenue.
- When the head of sales asks “what if”: An analyst adds an “aggressive case” scenario that starts from the budget assumptions and layers on changes such as a 20% revenue lift or updated R&D averages, all inside the same model.
Ground Claude's answers in your Drivetrain data with the MCP server
Christian mentioned another limitation in using AI for finance, which is that it doesn’t know your business context. As a result, any insights it provides are generic.
Jaco showed how his Drivetrain account was connected with Claude through Drivetrain’s MCP (Model Context Protocol) server. Connecting Claude via MCP gives it direct access to your data in Drivetrain so that responses are grounded in your real data, and every number is traceable back to the metric it came from. Jaco added that with the MCP server connection to Claude, access control carries through: Users who can't view the data inside Drivetrain can't see it through Claude either.
Jaco then asked Claude for revenue per full-time employee (FTE). He explained that Claude can either fetch the metric directly from Drivetrain or, if it isn’t already defined, calculate it from revenue and headcount.
Claude returned almost the revenue per FTE result along with the recent trend. To save time, Jaco had run two follow-up prompts before the session and, during the demo, showed what came back. Claude had searched the web for benchmarks from similarly sized companies, compared his revenue per FTE against them in a table, and drafted a playbook for improving the metric based on those benchmarks and his Drivetrain data. Jaco said the exchange took him from a single data point to a full strategy in just a couple of minutes.
To ask questions like these of your own numbers, book a personalized demo.
Summary
Most of this session focused on the different ways Christian uses ChatGPT and Excel’s Copilot as AI copilots for Excel financial workflows:
- Christian prompted ChatGPT to write the pacing formulas by describing the workbook’s layout, so he never had to upload the file. Then he pasted the formulas into Excel and tested them against figures he could verify.
- He gave Copilot’s agent mode a prompt listing assumptions such as a $500 price and cost of goods sold (COGS) at 14% of price, and the agent returned a formula-driven three-year P&L. Christian added that he’d check every cell and its logic before using it for anything real.
- Because language models are unreliable for arithmetic calculations, in his forecasting demo, Christian had ChatGPT write Python code for the forecast and pasted it into Excel to run on Excel’s built-in Python. It wasn’t successful on the first try, so he fed each error back to ChatGPT until the code ran properly.
- Christian recommended back-testing an AI forecast by removing your most recent actuals, asking the AI in a fresh chat to forecast those months, and then checking whether the real results fall inside the forecast’s confidence bounds.
- Christian said general-purpose AI tools still require manual data uploads, don’t support real collaboration, and don’t connect to live data, which keeps reforecasting tedious.
- Jaco demoed how Drivetrain, an AI-native FP&A platform, replaces uploads with direct connections to source systems, keeps data current with scheduled refreshes, lets analysts build scenarios for business partners inside one model, and, through its MCP server, gives Claude business context with access to the company’s real data. Access controls in Drivetrain carry through.
Questions from the webinar
What should FP&A professionals focus on once AI handles the repetitive work?
Christian said the repetitive tasks are less enjoyable anyway, and finance is paid to help leaders make better decisions through business partnering and deeper analysis. Earlier in the session, he named challenging assumptions as the part of finance work AI will be slowest to replace, because AI answers questions well but isn't yet a good challenger.
How can I use Excel Copilot without enterprise access?
Christian said Excel Copilot is built on some version of ChatGPT, though Microsoft doesn't disclose which. Without enterprise-grade access, use it for brainstorming and formula help, and keep confidential data out of it. His pacing demo shows the method in ChatGPT: he described the sheet's layout without uploading the workbook.
Which finance tasks can I hand to ChatGPT with the least risk of a hallucinated answer?
Tasks where a wrong answer shows up immediately and costs nothing carry the least risk, Christian said. A formula is his example: if it fails, you paste the error back and ask again, and no factual claim is at stake. Answers built on facts or citations carry more risk, so click through every reference before using one.
How are finance teams using generative AI at work today?
Attendees listed reviewing memos, advanced Excel and PowerPoint functions, contract analysis summaries, meeting minutes, and first drafts of communications. Christian drafts his own emails first, then has AI fix grammar and tone, flag anything missing, and tighten the language. He said most companies now give employees access, yet the tools go underused because people struggle to find concrete use cases, a pattern Drivetrain's State of AI in FP&A report examines across finance teams.
Which AI-powered formulas and functions in Excel did Christian use for finance work?
Christian had ChatGPT write standard Excel functions, such as a SUMIFS with a date condition for revenue to date, and tested each one against figures he could verify. For heavier math such as an ARIMA forecast, he had ChatGPT write code for Excel’s built-in Python, which ran the calculation with no AI involved. He also had it write a formula for the percentage of the month elapsed, which he checked by hand and then asked ChatGPT to simplify.
How does Drivetrain keep financial data secure when Claude is connected?
Jaco said access control carries through the MCP server, so anyone without permission to view data inside Drivetrain can't see it through Claude either. Drivetrain is SOC 1 Type 2 and SOC 2 Type 2 certified.
Speaker spotlight

Christian Wattig
Founder, Inside FP&A
Christian Wattig is an FP&A expert with 14+ years of experience in MNCs (P&G, Unilever) and high-growth tech startups (Squarespace). After earning his MBA from NYU Stern, he helped take Squarespace public. Now an entrepreneur, Christian directs the FP&A certificate program at Wharton and runs his firm, FP&A Prep, offering corporate training, online programs, and sharing insights through LinkedIn, newsletters, and his 'FP&A Tuesday' series.

Jaco Schoeman
Associate Director of Strategic Finance, Drivetrain
Jaco Schoeman is Associate Director of Strategic Finance with over a decade of hands-on finance experience in financial planning, budgeting, and FP&A system implementation. He spent over four years at Vena Solutions as Senior Services Solutions Manager, where he partnered with sales and services teams to scope projects, position FP&A solutions, and guide clients through software implementations. Before Vena, Jaco held finance and planning roles at Predator Ridge Resort, Caron Business Solutions, and BASF, where he managed commercial planning and forecasting cycles as Company Controller. At Drivetrain, Jaco writes about financial planning and budgeting best practices, FP&A software implementation strategies, and how finance teams can optimize their planning processes.
- 00:00
Jaco Schoeman
Everybody, this is really, truly great to see — some people already here with us, obviously to talk Excel and the latest AI features. I'm joining you from British Columbia in Canada this morning. Why don't we take a couple of minutes while people filter in — type in the chat where you're dialing in from, we'd love to know. Perfect, seeing people come in — Egypt, London, NYC, Florida, India, Las Vegas, UAE, North Texas, Latvia, Lagos, Dallas, Buffalo — welcome everybody. So, why should you be listening to us today? Just by way of introduction, my name is Jaco, I'm the associate director of strategic finance here at Drivetrain — think of us as Excel's smarter cousin: we turn your static sheets into live, automated business models, helping your FP&A teams report, plan, and forecast without the copy-paste chaos. I'm really excited to introduce you to Christian Wattig this morning. Christian is an FP&A expert with over fourteen years of experience across multinational corporations like P&G and Unilever, as well as high-growth tech startups like Squarespace — after earning his MBA from NYU Stern, he helped take Squarespace public. Now an entrepreneur, Christian directs the FP&A certificate program at Wharton and runs his own firm, insidefpa.com, offering corporate training, online programs, and sharing insights through LinkedIn, newsletters, and his FP&A Tuesday series. Christian, welcome, it's great having you.
- 02:15
Christian Wattig
Thanks so much for having me today, Jaco, excited to be here.
- 02:20
Jaco Schoeman
That's so good. So what are we discussing today? We're going to learn about generating financial models using Excel's agent mode, building forecasts in minutes by combining ChatGPT, Python, and Excel, and also the limitations of LLMs and how to address them. Christian, I'll hand the presentation over to you — we're grateful you're joining us and looking forward to learning from you.
- 02:58
Christian Wattig
Fantastic — you should see the screen now. I'm really excited to dive into some more advanced AI use cases in finance. Before that, I'd like to welcome everyone, and especially the CPAs in the audience today — if you're a practicing CPA based in the US and want a free CPE credit for this session, you'll need to participate in three of the four polls we'll launch over the next fifteen minutes; simple quiz questions, you don't need to answer correctly, just participate so we have evidence with NASBA that you attended live. With that said — please don't hesitate to put questions in the Q&A box, we'll answer them at the end. Let me pick things up from the beginning: FP&A is changing. When I train corporate finance teams, I still often see them dealing with reconciliations between systems, formatting data, removing duplicates, copy-pasting between spreadsheets, handling exceptions, and producing standard reports — your P&L, balance sheet, cash flow statement, and other recurring metrics — plus writing repetitive variance commentary that sounds similar every month. I'm convinced all of this will largely disappear within a few years; some companies are already further along using AI for exactly this, and you'll see how today. So the question becomes: what will finance focus on once these manual tasks are automated? I want this to be interactive — type your thoughts into the Zoom chat. I'm seeing a lot of responses around business partnering, strategic decision-making, real-time forecast updates, predictive analytics, and agent orchestration — someone will need to manage not just people but the AI agents themselves in the future. Business partnering is what I see mentioned most. I actually ran a survey with finance leaders a while back, and the top reason people gave for not working more with cross-functional leaders was simply lack of time — I think AI will make a real difference there. You've covered most of it already — I'll launch our first poll here, related to this, mainly so our CPA attendees can show participation; I'll close it in a minute or two. So: in-depth analysis — peeling the onion, asking why not just of the data but of your business partners, understanding variances against budget, forecast, or prior year, and pushing until you reach the "so what" behind a number. It's about advising cross-functional partners, surfacing metrics earlier so they can decide and act faster — which takes real time, because you have to build deep business understanding and genuinely step into the shoes of a product manager, engineering lead, marketing lead, or sales rep. But sometimes adding value isn't just helping others decide better — it's challenging assumptions, helping people think outside the box by asking the right question at the right time. That's a deeply human task, and I think it's one of the last things AI will fully automate, because AI is fundamentally an assistant — it answers your questions well, but it's not yet a good challenger. And then there's connecting strategy to tactics and measurable action plans — looking beyond the obvious data, since AI is only as good as what it's fed. A big part of adding value in the age of AI is actively looking at datasets you wouldn't normally check — insights hidden in contracts, HR data, even news — things not automatically connected to what your AI already has access to. Okay, closing that poll. But let's step back — what do we actually mean by artificial intelligence? In a nutshell, it means trying to mimic human cognition, that's really it. In finance specifically, there are two meaningfully different flavors: machine learning and generative AI. We'll spend most of today on generative AI, but I don't want to skip machine learning entirely, since it's a much older, more mature technology that's been around over a decade — about eight years ago I led the effort to implement Unilever North America's first machine learning model, forecasting working capital. Machine learning takes your actual historical results, identifies trends, strips out one-offs, and projects forward. It's powerful for forecasting because its strengths and weaknesses are almost the mirror image of judgment-based forecasting: human forecasts carry bias but capture nuance and context; machine learning has no human bias, but it's entirely bound by the data you feed it — quality issues or unusual one-offs in recent data get amplified rather than filtered out. Because those biases are so different in nature, combining expert judgment with machine learning can be extremely powerful. It used to require an expensive data science team; today you can use third parties or software with ML already built in, so it's much more accessible — though it remains something of a black box, since even data scientists often can't fully explain why a model produced a given output, since the underlying math is genuinely complex. It really comes down to experimenting with your inputs. For example, at a website-builder company I worked with during the pandemic, our machine learning models suddenly stopped performing well in mid-to-late 2020 — it wasn't until we fed in different datasets, like economic data, unemployment figures, and shutdown-related data, that the model regained accuracy. That experimentation is what makes machine learning so interesting — you can make it far more sophisticated than anything Excel-based. If you haven't tried it at your company, I'd encourage you to — and tools like ChatGPT already make a simplified version of this accessible; I'll show you how to build a simple ML-based model with ChatGPT later. That's a good segue into generative AI — GPT, Gemini, Claude, and others, with enormous capital pouring into the space as everyone races toward the most effective tool. But even after running corporate AI workshops and full-day trainings, I find people still struggle to get the most out of it — most companies now give employees access, but it's underutilized, often because people struggle to come up with concrete use cases. I hope by the end of today you'll leave with plenty. Before covering the great use cases, let's briefly talk about the risks — what would you say are the dangers of using generative AI? I'm seeing hallucination, confidentiality issues, bad data, loss of critical thinking ability mentioned — there was actually a recent study showing that the more people rely on AI, the less brain activity is measured during the task, so differentiating yourself going forward may mean continuing to do that deep thinking alongside AI use. Also over-reliance on output, and security issues. So, data security is arguably danger number one — though fortunately, after three or four years of this technology being mainstream, the major AI companies have built ways to use tools like ChatGPT and Gemini safely at your company; just triple-check with your manager and IT before uploading anything confidential, since you can configure these tools so uploaded data stays on-premises, isn't sent to the cloud, and isn't used for training — but you need to verify that's genuinely how it's configured. Second: hallucinations — the model confidently making things up. It's gone as far as an attorney in the US being disbarred after submitting AI-generated legal citations to a court that referenced trials that didn't exist, along with links nobody bothered to click and verify. A few ways to work around hallucinations: use reasoning models, which think more slowly and have built-in checks that reduce (not eliminate) hallucination; always ask for references and actually click through to verify them, since a link can look legitimate without being real; and, third, use AI specifically for tasks where hallucination doesn't really matter, which I'll show examples of. Third limitation, which nobody mentioned: it's genuinely not great at math — the classic example is asking how many R's are in "strawberry," which many models get wrong repeatedly. The workaround is having the AI write Python code instead of doing arithmetic directly, then running that in Excel, since Python handles math far more reliably. Before we get there — how do people use AI in their everyday personal life? I love asking this since people always have interesting use cases: wedding speeches (getting past that blank-page anxiety), meal planning, hike recommendations, calorie tracking, workout planning, card and letter writing, travel recommendations, understanding Excel formulas and file logic — great, lots of personal uses already. Now, at work: reviewing memos, advanced Excel/PowerPoint functions, project planning commentary, contract analysis summaries, meeting minutes, first drafts of communications. We'll relaunch our second poll here, again mainly for our CPA attendees. Let's cover a couple of use cases you may not have mentioned. Emails: I like to draft the email myself quickly first, then paste it into AI and ask it to clean up grammar and tone, suggest anything I've missed, and tighten the language — I much prefer that to having AI draft from scratch, since a from-scratch draft usually needs more editing than my own first pass improved by AI. Excel: it's a phenomenal Excel teacher, which is honestly part of why I built my business teaching what happens outside of Excel to succeed in FP&A — it's great at explaining formulas and helping build models, which we'll do shortly. It's also great at explaining tools generally — say you use SAP as your ERP, it can help you navigate it, though it can lag a bit behind the newest software updates. With some back-and-forth prompting it's still very effective, and I genuinely think in the near future every software tool will have AI built in, essentially eliminating the learning curve. Then Python and SQL: Python is excellent for advanced analysis, even forecasting and modeling, which I'll show later; SQL lets you slice and query large datasets quickly, and rather than painstakingly learning the syntax yourself, you can just have ChatGPT write those queries for you. And lastly, brainstorming — I love using it for that, and I'll show an example. But across all these use cases, the core challenge is that AI doesn't inherently know your business context — it can give generic insights that are hard to act on without that context. So, Jaco, can Drivetrain help address that limitation?
- 25:20
Jaco Schoeman
Absolutely, we can. We recently launched a Drivetrain MCP server that does exactly this — I've connected Claude to my Drivetrain account, and you can see it's already logged in here. LLMs are very powerful, but unlocking that power usually means manually uploading your data, which is time-consuming and can be risky — without your actual data, it's hard to get anything beyond generic advice. This MCP server bridges that gap, so when you ask Claude a question, the response is grounded in your real data. For example, I can ask Claude, "what's my revenue per FTE?" — it interprets the request and goes to fetch that metric; if it's not already defined in Drivetrain, it'll pull revenue and headcount separately and calculate it. Importantly, if you don't have the security permissions to view that data inside Drivetrain itself, you won't be able to see it via Claude either — access control carries through. Scrolling down, our most recent figure, April 2025, comes in at almost $170,000, with a recent trend shown as well. In the interest of time, I've already run a couple of follow-up prompts I'll show you: I asked how this compares to industry benchmarks for similarly-sized companies, requesting the answer as a table — Claude searched the web, pulled in and filtered relevant industry benchmarks, and returned a genuinely useful comparison, honestly comparable to what you'd get from a senior analyst. I also asked how I could improve revenue per FTE, and it searched the web to curate a playbook grounded in best practices, benchmarks, and my actual data, with some concrete quick wins. So in just a couple of minutes, I went from a single data point to a full strategy — that's the power of AI inside Drivetrain.
- 28:18
Christian Wattig
Thanks so much for that quick demo, really cool to see. Let me pull my screen back up, since now we're getting into the more advanced material — how to do similar things with ChatGPT directly. Before diving into use cases, I want to quickly cover how to structure your prompts, since this is crucial — garbage in, garbage out. I'll launch our third poll here while I explain. My favorite prompting framework was shared by Greg Brockman, co-founder and current president of OpenAI, the company behind ChatGPT — if anyone knows how to prompt these models well, it's him. He recommends structuring prompts around goal, return format, warnings, and then a context section. I won't read the whole example, but broadly: clearly state your goal, specify how you want the output formatted, add relevant warnings (like double-checking that a cited fact or timeframe is actually accurate), and then provide context. Context is where I see most finance teams fall short in training — think of AI as a very well-read intern: it's absorbed the entire internet, but knows nothing specific about your business or what you actually need, so you have to give it a real context dump to work with. Let's get into use cases, starting with agent mode. For those unfamiliar, agent mode in Copilot was only released in test form a few weeks ago — very recent. You need a Microsoft 365 or paid Copilot subscription, and access it via Add-ins > Excel Labs, an experimental feature that should eventually reach desktop Excel. Agent mode is quite powerful for building models from scratch — I'll paste a prompt now: create a three-year P&L model for a new product launch based on some quick meeting-note-style assumptions — $500 target price, cost of goods sold at fourteen percent of price, some marketing spend, sales targets, headcount and associated costs, G&A overhead at five percent of revenue. I'll submit it; it'll take a few minutes to reason through, and I'll warn you upfront this is cutting-edge and early, so it may or may not work well live. While that's running, let me show you something else: a weekly pacing model, which answers, are we tracking to our monthly target, and are we ahead or behind? Companies typically do this weekly — finance shares a tracker comparing realized revenue and expenses against how much of the month has elapsed; if you're halfway through October but only twenty percent toward your sales goal, that's a signal to act. This model needs to dynamically calculate, based on a changeable date cell, both how many days of actuals exist and what percentage of the month has elapsed — correctly handling months of 28, 30, or 31 days. Here's how I'd prompt this: "I'm a finance manager at a software company building a weekly pacing report in Excel to track actuals against monthly forecast — I need help writing the Excel formula." That's the goal. Return format: provide formulas I can paste directly into Excel, with short explanations of each component. Warnings: keep formulas as simple as possible for auditability, and verify they truly do what's intended. Context: this is a high-visibility project for my CFO, someone I don't work with often, and I want it to help me stand out. That's the format — being specific here is exactly where finance teams often fall short. Note I'm not even uploading the actual file, to show this works even without a secure enterprise ChatGPT tier — I'm just describing where the relevant data sits (actual revenue in row 22, dates in row 20) and how the sheet is structured. This is actually a good example of a low-hallucination-risk use case: if the formula it returns doesn't work, nothing breaks, there's no factual claim at stake — I just paste the error back and ask it to try again, and you'll find that iterative back-and-forth gets you a working result even when the first attempt fails. It gave me a SUMIFS-style formula with a less-than-or-equal condition — let's paste it into the actual pacing model... $47,498, and checking against total revenue confirms that's correct. If I change the date, you can see it recalculates dynamically — so this is a great way to have AI essentially teach you Excel as you go. Let's tackle the trickier formula — the percentage of the month elapsed — by going back to the chatbot; it's kind enough to explain its reasoning too, which I'd encourage you to actually read rather than just grabbing the formula, so you learn something. This one came back fairly long and complex — let's test it... it's giving a slightly odd result, so let's ask for something simpler; checking fifteen divided by thirty-one confirms the underlying math is right, so I'll just ask it to simplify the formula, which works better. Let's check back on that P&L build — it reasoned for about sixty-four seconds, quite fast. It calculated revenue and gross margin, listed its assumptions, and linked the summary correctly back to cell B8 — everything appears dynamic rather than hard-coded, which is genuinely nice, including revenue, operating income, and even a chart. At first glance it did a solid job — though before using this for anything real, I'd absolutely do the same diligence you'd apply to a model handed to you by an analyst: check every cell, verify the logic, because mistakes are still possible. But it's genuinely impressive that from a single prompt you get a fully dynamic starting model that saves real time — and this will only keep improving, so it's worth keeping an eye on Copilot. Next, machine-learning-based forecasting combining ChatGPT and Python — a fairly advanced use case I'll walk through. Here's a dataset: eight years of monthly data, in this case retail deposits for a bank, with actuals ending in April 2025. I want the model to forecast the rest of 2025, May through December, using an ARIMA model specifically — ARIMA stands for autoregressive integrated moving average, a fairly sophisticated algorithm that accounts for seasonality and can adjust based on its own past errors. I'm not having ChatGPT do the arithmetic directly, since it's not reliably good at math — instead I'll have it write Python code that I insert into Excel, using a relatively new native feature Microsoft added. I'll ask it to write Python code that also returns confidence intervals, not just a point forecast, and to double-check that the code runs. While that's thinking, let me show how the Python integration itself works — under Formulas, on a current Excel version, there's an Insert Python button; that's where the generated code will go. It's produced something — fits a small ARIMA model. I'm not a developer myself, so I can't independently verify the code, but I can copy and insert it; a quick security note: since you're pasting code into Excel, it's worth scanning it, or even asking another AI to explain exactly what it does, before running it, and always check with your IT team and manager before doing this with real company data (to be clear, I'm not uploading anything to ChatGPT here, just executing code it generated). This uses Excel's native Python execution, not Copilot or any AI at runtime. While that runs — let me address a couple of questions from the chat. Tisha asked: if AI handles repetitive tasks, what should FP&A professionals focus on instead — just judgment? Those repetitive tasks are honestly less enjoyable anyway; you're fundamentally paid to help leaders make better decisions, which you do through business partnering and deeper analysis — that's where your focus shifts. Another question: how can I best use Excel Copilot without enterprise access? Excel Copilot is essentially built on some version of ChatGPT, though Microsoft doesn't fully disclose which — you can still use it for brainstorming and formula help the way I've shown, just avoid uploading anything confidential without enterprise-grade access. Back to our running example — it's still stuck in the Python cell, so let me feed it that as an error and see what it suggests, perhaps a simpler approach that's less computationally heavy. While it thinks — John asked: what happens when AI makes a mistake in a forecast, how do you catch and fix it? Great question, and exactly what I wanted to cover next. Let me copy this new code from ChatGPT and try executing it — this has worked for me in testing before, though we may not have time to fully resolve it live. It's returned an actual error this time, an attribute error — so I'd feed that back to ChatGPT, and in my experience it typically takes two or three iterations to land on a working forecast. To John's underlying question — how do you know if a forecast is actually good? Through back-testing: deliberately remove your most recent known actuals (say, everything from May 2024 onward), start a fresh chat with no history, re-upload the truncated dataset, and ask it to forecast through December 2024 — then compare that forecast directly against the real actuals you already have for that period, checking whether they fall inside the eighty-five or ninety-five percent confidence range. When I tested this with this exact dataset previously, the working code produced surprisingly accurate results — though your mileage will vary with more irregular data. We're hitting a different error now, so I'd continue that back-and-forth, though in the interest of time I won't keep troubleshooting live — in my own testing it usually takes two or three rounds of back-and-forth with the AI before it produces a working forecast. That's a natural segue into limitations more broadly: beyond not always working on the first try, AI tools currently require manual data upload, don't support real collaboration, and don't connect to your real-time data — so reforecasting stays somewhat tedious even once you've got a working machine learning setup. If you're scaling, that's where AI-native FP&A tools — platforms like Drivetrain with AI built directly in — start to make more sense. Jaco, could you show that briefly before we wrap up, so people can see how this looks inside a purpose-built FP&A tool?
- 53:14
Jaco Schoeman
Sure thing — thanks for passing that over, Christian. Drivetrain was built primarily to streamline day-to-day FP&A work — let me give you a quick glimpse. The biggest time sink for FP&A teams is usually just getting data into the system in the first place; whether that's multiple ERPs or multiple entities across different currencies, that's straightforward in Drivetrain. Here I've got a couple of Xero instances, Google Sheets, and Salesforce data connected — we support over eight hundred native connectors, and once connected, you can set refresh frequency (daily, weekly, monthly) or trigger a manual sync anytime. On planning specifically: the typical challenges are keeping data real-time for forecasting, coordinating with stakeholders like your CMO or head of sales to get timely inputs, and the sheer time investment involved. Inside Drivetrain, our Drive App Hub lets you spin up base models for things like headcount planning, restatement models, and pipeline movement walks — launching one of these gives you balance sheet, cash flow, and P&L models along with out-of-the-box reporting, powered by an AI agent working in the background; a few minutes later you have a fully populated three-statement model. Drilling into the income statement here, forecasting is already populated based on the connected NetSuite data — if I wanted to shift my sales and marketing forecast assumption from six months to three, that's a quick click-and-apply change, or I could switch it to a data-driven or fully driver-based assumption, like a flat twenty percent of revenue. Updating actuals as books close is similarly simple — select the month, hit update, and it actualizes that column while everything forward-looking from July onward stays intact, since it's still driven by the underlying forecast formula. If your CMO or head of sales needs an alternative scenario, you don't need a separate model — scenarios are built in natively; I can add an "aggressive case," and by default it inherits my existing budget assumptions, then I can layer in adjustments like a twenty percent revenue lift or updated R&D averages, giving you multiple live versions to forecast from as a business partner. We've got one more poll running — while that's up, I'll mention we're SOC 1 Type 2 and SOC 2 Type 2 certified, work with a range of customers globally, and Drivetrain generally eliminates a lot of the manual upkeep that comes with maintaining an Excel-based forecast. We'll share this deck and the recording afterward, along with some additional resources — definitely check out Christian's bootcamp, plus our "fourteen ChatGPT prompts for finance" resource and our State of AI in FP&A report. Christian, I think we're running slightly over, but happy to extend a bit if there are unanswered questions — feel free to drop them in chat. Otherwise, I just want to say a genuine thank-you, Christian, for your time today — this was really insightful, and I've personally learned a lot; Excel is close to my heart too. If anyone would like a personalized demo, feel free to scan the QR code on screen. Thanks again, Christian, for walking us through Copilot and its AI capabilities today.
- 1:00:15
Christian Wattig
Thank you, Jaco. Just really quickly before we close out, for anyone still here — I fed that third error message back into ChatGPT, and the resulting code actually ran successfully. So let me briefly show you the outcome: once it works, you get a data-frame cell, and clicking into the array preview shows the forecast along with the lower-eighty and lower-ninety-five confidence bounds — the range where the actual outcome is likely to fall. The natural next step would be back-testing it against last year's actuals to validate accuracy, but just to wrap up — echoing what Jaco said, thank you all so much for joining today, really appreciate the time and engagement. Thanks so much.





