Know What Numbers Can and Cannot Do

AI in Finance: Opportunities and Risks

When the Spreadsheet Stopped Being Enough

The shift wasn't magic. It required the finance team to rethink which tasks genuinely needed human expertise and which tasks were just expensive, slow, manual processes dressed up as expertise. Reconciling transactions? AI handles it faster and with fewer errors. Writing the first draft of a board-ready variance report? AI can produce a solid 80% draft in minutes. Deciding whether a 12% cost overrun in Q3 signals a structural problem or a one-time event? That still needs a finance professional with context, history, and business judgment. Klarna's experience illustrates the central tension of AI in finance: the tools are genuinely powerful, and the risks of misusing them are genuinely real.

This lesson is about navigating that tension without either dismissing AI as a gimmick or blindly trusting it as an oracle. Finance and operations professionals who get this balance right will work faster, produce better analyzis, and make stronger recommendations. Those who get it wrong, who accept AI outputs without scrutiny, or who refuse to use the tools at all, will find themselves increasingly behind. The goal here is practical competence: knowing exactly which AI tools do what, where they genuinely save you time, and where they can lead you quietly off a cliff.

The Principle Hidden Inside Klarna's Story

The principle Klarna discovered isn't unique to them. It shows up in almost every finance team that successfully adopts AI tools: the biggest gains come from separating volume work from judgment work. Volume work is anything that involves processing, formatting, summarizing, or drafting, tasks where the inputs are clear and the output follows a pattern. Judgment work is anything that requires weighing ambiguous information, applying organizational context, or making decisions with consequences. AI excels at volume work. It assists with judgment work but cannot replace it. When teams blur this line, when they let AI make judgment calls or when they refuse to let AI touch volume work, they either create risk or leave enormous efficiency gains on the table.

Think of it this way: if you hired a brilliant but very literal assistant who could read and write faster than any human, had no ego about doing repetitive tasks, but had never actually worked at your company and didn't know your industry's quirks, you'd give them very specific, bounded tasks. Draft this report based on these numbers. Summarize these vendor contracts. Flag any line items that exceed $10,000. You would not ask them to decide whether to renew a supplier relationship or explain to the CFO why Q2 margins compressed. That's the right mental model for AI in finance today. It's a fast, tireless, pattern-recognizing assistant, not a financial strategist.

This distinction matters even more in finance than in other departments because the stakes of errors are concrete. A poorly written marketing email can be corrected. An incorrectly summarized contract clause, an AI-generated forecast accepted without verification, or a compliance document that sounds authoritative but contains a hallucinated regulatory requirement, these can cause real financial and legal damage. The tools are powerful enough to be worth using and consequential enough to require discipline. Both things are true at once.

Click to enlarge

What AI Finance Tools Actually Do: A Story From Healthcare

Bon Secours Mercy Health, one of the largest hospital systems in the United States, faced a problem that will feel familiar to anyone who manages operational budgets: their finance team spent enormous time producing reports that people read for about four minutes before asking follow-up questions the report didn't answer. The analyzts would go back, pull more data, reformat, redistribute. The cycle repeated. When their team began using Microsoft Copilot, the AI tool built into Microsoft 365, integrated with their existing Excel and PowerPoint workflows, something useful happened. analyzts started generating initial report narratives directly from their data, then spending their time refining and contextualizing rather than building from scratch.

The critical detail is what Copilot was actually doing in that workflow. It wasn't analyzing strategy. It was reading structured data that the analyzts had already validated and producing clear, readable prose that described what the numbers showed. "Revenue per adjusted patient day increased 3.2% year-over-year, driven primarily by outpatient volume growth in the cardiology service line." That sentence used to take an analyzt ten minutes to construct while toggling between tabs. Copilot produced it in seconds. The analyzt's job shifted to verifying accuracy, adding context Copilot couldn't know (like the fact that a new cardiology partnership launched in March), and making recommendations, the genuinely high-value part of the role.

A Different Industry, The Same Pattern: Retail Operations

Patagonia's operations team deals with a complexity that most retail finance professionals will recognize: hundreds of SKUs, multiple suppliers across different countries, fluctuating raw material costs, and sustainability commitments that add a layer of reporting requirements that purely financial metrics don't capture. Their operations and finance teams began using AI tools, specifically Claude and ChatGPT, to accelerate the synthesis of supplier cost data into plain-language summaries that procurement managers and finance partners could actually discuss together. The tools didn't replace their ERP systems or their analyzts. They bridged the gap between raw data and human conversation.

What made this work in practice was discipline about the handoff points. The AI tools handled the drafting, summarizing, and reformatting. Human analyzts handled the interpretation, the supplier relationship context, and the final recommendations. A procurement manager might paste a 40-row cost comparison table into Claude and ask for a clear summary of which three suppliers showed the most significant cost increases and what patterns might explain them. Claude would produce a readable summary in under a minute. The manager would then apply their knowledge of those supplier relationships, a factory fire in one region, a new labor agreement in another, to turn that summary into an actual recommendation. Speed from AI. Judgment from the human. That combination is where the real value lives.

The CFO Who Almost Trusted the Summary: A Cautionary Story

A CFO at a mid-sized manufacturing company, not a name you'd recognize, but a real scenario shared in a finance leadership forum, used ChatGPT to summarize a 60-page supplier contract ahead of a renegotiation meeting. The summary was excellent: well-organized, clearly written, hit all the major commercial terms. The CFO went into the meeting confident. Midway through, the supplier's legal team referenced a pricing escalation clause that the AI summary had described as capped at 3% annually. The actual contract language allowed for escalation up to 3% or CPI, whichever was greater, a meaningful difference in an inflationary environment. The CFO caught it in the room, but barely. The AI hadn't lied. It had summarized imprecisely, and the imprecision had financial consequences.

This story isn't an argument against using AI to summarize contracts, it's an argument for using it correctly. The right workflow is: use AI to get to 80% comprehension quickly, then focus your own careful reading on the highest-stakes clauses. The AI summary tells you where to look and what questions to ask. It doesn't replace reading the original for anything consequential. Finance professionals who adopt this habit. AI for orientation, human attention for the critical details, get the speed benefit without the risk. Those who treat AI summaries as substitutes for source documents are building on sand.

Click to enlarge

Putting It Into Practice: Your First AI Finance Workflow

The most practical place to start using AI in finance is with written deliverables, reports, summaries, variance explanations, email updates to stakeholders. These are tasks where you already know what good output looks like, which means you can immediately spot when AI gets it wrong. They're also tasks that consume a disproportionate amount of finance professionals' time relative to the strategic value they create. A monthly operations report that takes a senior analyzt four hours to write delivers the same information whether it took four hours or forty-five minutes. The time saved is real; the quality difference is often minimal when AI assistance is used well.

The key to making this work is what's called prompt quality, essentially, how clearly and specifically you tell the AI what you need. Think of it like briefing a contractor. A vague brief gets mediocre work. A specific brief, with context about the audience, the purpose, the tone, the key facts, and any constraints, gets work you can actually use. In finance contexts, this means always including the actual numbers you want discussed, the audience for the document (board? operations managers? external auditors?), the tone required (formal? concise? explanatory?), and any specific points that must be included or avoided. The more specific your brief, the better the output.

Start with one recurring deliverable you produce every month that you find tedious rather than intellectually engaging. Maybe it's the monthly expense summary email to department heads. Maybe it's the first-draft narrative for your budget-versus-actual report. Maybe it's formatting and summarizing meeting notes from your weekly ops review. Pick one, try AI on it this week, and compare the output to what you'd have written yourself. Most finance professionals are surprised to find the AI draft is 70-80% there, and that improving it takes a fraction of the time starting from scratch would have taken. That's the starting point. Not a wholesale transformation. Just one workflow, improved.

Key Principles From Part 1

1. AI in finance creates the most value when it handles volume work, drafting, summarizing, formatting, while humans retain ownership of judgment work: interpretation, context, and decisions.
2. The tools that matter for non-technical finance professionals right now are ChatGPT Plus, Claude Pro, Microsoft Copilot (built into Excel, Word, and Teams), and Google Gemini, no coding or technical setup required.
3. Always supply your own verified numbers to AI tools. AI cannot access your financial systems and will not tell you when it's fabricating a plausible-sounding figure.
4. AI-generated summaries of contracts, compliance documents, or regulatory requirements should be treated as orientation, not as source-of-truth. Always verify high-stakes details against the original document.
5. The 80/20 rule applies: use AI to reach 80% of your output quickly, then apply human judgment to the final 20%, the part that requires context, accuracy verification, and professional accountability.
6. Prompt quality determines output quality. Treat AI like a talented contractor: specific briefs produce usable work; vague briefs produce vague results.
7. The right entry point is one recurring, tedious deliverable in your current workflow, not a wholesale transformation of how your team operates.

When AI Meets the Monthly Close: A Story From the Trenches

In early 2023, a mid-sized logistics company in Ohio, about 400 employees, three regional offices, was drowning in its month-end close process. The finance team of six spent the last four days of every month manually reconciling invoices, chasing down department heads for expense approvals, and compiling variance reports in Excel. The controller, a 15-year veteran named Sandra, described it as 'organized chaos.' They brought in Microsoft Copilot for Finance (then in preview) alongside their existing Microsoft 365 setup. Within 60 days, their close cycle dropped from four days to two and a half. Sandra didn't lose a single team member. But she did reassign two of them to forward-looking analyzis work they'd never had bandwidth for before.

What changed wasn't the accounting rules or the underlying data. What changed was where human attention went. Before AI, skilled finance professionals spent enormous energy on data movement, copying numbers between systems, formatting reports, writing routine variance commentary like 'Sales were 8% below forecast due to delayed Q3 shipments.' After AI, those same professionals spent their time on questions the data raised, not on transcribing the data itself. That's the core shift. AI handles the description. Humans handle the interpretation.

Sandra's team also discovered something less expected: the AI flagged two expense anomalies during the first automated reconciliation run, a duplicate vendor payment worth $14,000 and an expense category that had been miscoded for three consecutive months, quietly inflating one department's overhead by about $4,200 per month. Neither error was deliberate. Both had been invisible in the manual process simply because no one had time to look closely enough. This is where AI's pattern-recognition capability earns its place in finance, not replacing judgment, but creating space for it.

The Three Jobs AI Actually Does in Finance

To use AI well in finance, it helps to think of it as filling three distinct roles, not one vague 'AI assistant' blob. The first role is data narrator: AI reads structured numbers and writes plain-English summaries. The second role is pattern spotter: AI scans large datasets for anomalies, trends, and outliers that humans would miss simply due to volume. The third role is draft generator: AI produces first versions of reports, memos, board summaries, and variance explanations that a human then reviews, edits, and approves. Each role has a different risk profile. Understanding which role you're asking AI to play in any given task is the key to using it responsibly.

The data narrator role is the lowest risk and the most immediately accessible. You paste a table of numbers into ChatGPT or Claude, you ask for a plain-English summary suitable for a non-finance executive, and you get a readable paragraph back. The AI isn't making financial decisions, it's translating. The risk here is mostly about accuracy: did it read the numbers correctly, and did it interpret context you didn't provide? A good habit is to always include the time period, the currency, and the comparison baseline in your prompt. 'Revenue was $2.3M' means something very different depending on whether last year's figure was $1.8M or $3.1M.

The pattern spotter role is more powerful but requires more setup. Tools like Microsoft Copilot integrated into Excel, or AI features inside accounting platforms like QuickBooks Advanced or NetSuite, can scan transaction histories and flag statistical outliers. This is genuinely useful for expense auditing, vendor payment reviews, and budget variance tracking. The risk in this role is over-reliance: an AI might flag something as anomalous that has a perfectly legitimate explanation, or miss a subtle pattern because it wasn't trained to recognize that specific business context. Human review of flagged items isn't optional, it's the whole point.

Click to enlarge

A Retail CFO's Lesson About Forecasting With AI

A specialty outdoor retailer with 22 stores across the Pacific Northwest used AI-assisted forecasting for the first time ahead of their 2023 holiday season. Their finance director, working inside a Planful forecasting platform with AI features enabled, asked the system to generate a revenue projection for Q4 based on three years of historical data, current inventory levels, and regional weather patterns the tool had access to. The AI produced a range: $8.2M to $9.6M, with a central estimate of $8.9M. The finance director felt good about it. It matched her intuition. She presented it to the executive team with confidence.

Actual Q4 revenue came in at $7.4M. The miss wasn't the AI's fault in any simple sense, two unexpected factors hit simultaneously: a competitor opened three new stores in their highest-performing markets in October, and an unusual warm spell in November delayed the seasonal demand spike by nearly three weeks. Neither factor was in the historical data the model used. The lesson the finance director drew wasn't 'AI forecasting doesn't work.' It was more precise: AI forecasting is only as good as the assumptions baked into the inputs, and it has no way to know what it doesn't know. Her practice now is to treat AI forecasts as a starting point for a structured assumptions conversation, not a conclusion.

Click to enlarge

AI Tools for Finance: What Each One Actually Does

The HR Budget Conversation: A Different Kind of Finance AI User

Finance AI isn't only for finance teams. Consider Marcus, an HR director at a professional services firm with 180 employees. Every year, Marcus has to defend his department's budget to the CFO, headcount requests, training spend, benefits cost projections, recruiting fees. He's good at the people side of his job. He finds financial modeling uncomfortable. For his 2024 budget presentation, he used Claude Pro in a way that changed how the CFO perceived his department entirely. He fed Claude his previous two years of HR spend data (copied from Excel into the chat), described his three hiring goals for the year, and asked it to help him frame the budget as a return-on-investment story rather than a cost center narrative.

Claude helped Marcus restructure his presentation around cost-per-hire benchmarks, turnover cost calculations (using a standard 1.5x salary estimate for replacing an employee), and projected savings from reducing time-to-fill by 15 days through a new ATS investment. The CFO, who had approved flat HR budgets for three consecutive years, approved a 12% increase. Marcus is quick to say the AI didn't win the argument, his data and his relationships did. But the AI helped him speak the CFO's language. That's a transferable skill for anyone in a non-finance role who has to defend budget requests: use AI to translate your operational goals into financial framing.

Practical Application: Building AI Into Your Finance Workflow

The most effective finance professionals using AI right now aren't doing anything exotic. They've identified three or four specific, recurring tasks that consume disproportionate time, and they've built simple AI habits around those tasks. A common starting point is monthly reporting commentary. Most finance managers write essentially the same variance explanations every month, with different numbers plugged in. This is exactly what AI does well: take structured inputs (actual vs. budget, key drivers, notable events) and produce readable, professional prose. The time savings are real, typically 45 minutes to two hours per report cycle, depending on complexity.

A second high-value application is preparing for financial conversations, board presentations, budget reviews, investor updates. AI tools like ChatGPT and Claude are excellent at anticipating questions. Give the AI your presentation deck content or key financials and ask: 'What are the five toughest questions a skeptical board member would ask about these numbers, and what's a strong response to each?' This kind of adversarial preparation used to require a senior colleague willing to play devil's advocate for an hour. Now you can do a version of it in 10 minutes, alone, at 10pm before a morning board meeting.

The third application is financial communication across teams, translating finance-speak into language that operations, sales, or HR teams can actually act on. Finance teams often produce accurate reports that nobody outside the department reads carefully because the format is impenetrable. AI can rewrite the same underlying analyzis in multiple registers: a one-paragraph Slack summary for a department head, a three-slide executive summary, a detailed variance memo for the audit committee. Same data, different audiences, different formats. The finance professional remains the author and the expert. AI is the translator and the formatter.

Key Principles From This Section

1. AI in finance works best as a cognitive bandwidth tool, it handles description so humans can handle interpretation.
2. The three distinct roles AI plays are: data narrator, pattern spotter, and draft generator. Each has a different risk profile.
3. AI forecasting is bounded by its inputs, it cannot anticipate what isn't in its training data or the context you provide.
4. Non-finance professionals (HR, operations, sales) can use AI to translate their goals into financial language that resonates with CFOs and budget committees.
5. Monthly reporting commentary, board preparation, and cross-functional financial communication are the three highest-ROI starting points for most professionals.
6. Every AI-generated financial output requires human verification of the numbers before it goes anywhere, the AI's job is to draft, not to certify.
7. Choosing the right tool matters: native accounting integrations (QuickBooks AI, Copilot in Excel) for number-heavy tasks; general AI tools (ChatGPT, Claude) for narrative and communication tasks.

What to Carry Forward

• AI doesn't close your books, it frees up the people who do to think more clearly about what the numbers mean.
• The most dangerous AI risk in finance isn't that it will replace finance professionals. It's that professionals will trust AI output without verifying it, and a confident-sounding error will make it into a board deck.
• Start with one recurring task that costs you time and produces a predictable output, variance commentary is the easiest entry point for most finance-adjacent professionals.
• The pattern-spotting capability of AI (finding duplicate payments, miscoded expenses, statistical outliers) is genuinely valuable, but flagged items always need human review before action.
• Speaking the CFO's language is a learnable skill, and AI can help anyone, not just finance professionals, frame their work in financial terms that get budget approved.

When AI Caught What the Auditors Missed

In 2022, a mid-sized logistics company in the UK ran its quarterly expense review the same way it had for a decade, a finance team of four manually sampling roughly 10% of submitted receipts. Then their CFO piloted an AI-assisted expense analyzis tool. Within 48 hours, the system flagged 34 duplicate reimbursements, three vendors who didn't exist in any business registry, and a pattern of Friday-afternoon submissions that consistently bypassed the approval threshold. Total exposure: just over £180,000. None of it had surfaced in the previous three audits. The CFO didn't celebrate immediately. Her first question was: 'How do we know the AI isn't wrong?' That question, not the fraud itself, turned out to be the most important thing she asked all year.

She assigned her most experienced analyzt to manually verify every single flag the AI produced. He confirmed 31 of the 34. Three were false positives, legitimate expenses from a contractor who used an unusual billing format. The system was 91% accurate on its first run, with zero prior training on this company's data. But here's what mattered most: the analyzt's verification process took two days, not two weeks. The AI didn't replace his judgment. It focused it. Instead of searching for problems across thousands of transactions, he was evaluating a shortlist. That's the shift worth understanding.

The principle this company discovered is one that holds across every finance and operations context: AI is a precision spotlight, not a verdict. It narrows the field dramatically, but a human still has to walk into the light and look. When organizations treat AI output as a final answer, they inherit its errors. When they treat it as a first-pass filter, they get faster, sharper human decisions, which is exactly what good financial oversight requires.

Click to enlarge

The Risk Hiding Inside the Efficiency Win

A marketing agency in Chicago started using AI to generate monthly budget variance reports for client accounts. The tool was brilliant at it, pulling actuals from their project management system, comparing against planned spend, and writing a clean narrative summary in minutes. Account managers loved it. Reports that took three hours now took fifteen minutes. Then, in Q3, a client noticed their report showed a 4% underspend in paid media, but their own platform data showed an overspend of the same amount. The AI had silently used the wrong date range when pulling actuals. Nobody had checked, because the report looked so clean and professional.

The agency lost the client. Not because of the error itself, errors happen, but because the client felt the agency had stopped paying attention. The efficiency win had quietly eroded the human review step. This is the most common failure pattern in AI-assisted finance work: the output looks authoritative, so verification gets skipped. A well-formatted AI report feels finished. It isn't. Build the check into your process as a non-negotiable step, not an optional one.

The fix the agency implemented was simple and effective. Every AI-generated report now includes a mandatory 'data source confirmation' line at the top, completed by the account manager before the report goes to the client. It takes 90 seconds. It requires the human to actively confirm the date range, the data source, and the account ID. That tiny friction point restored accountability without eliminating the efficiency gain. Small structural checks like this are the difference between AI making you faster and AI making you sloppy.

Click to enlarge

How a School District Finance Director Used AI Without an IT Team

A school district finance director in rural Ohio had no data science team, no BI software, and a budget that hadn't changed much in four years. What she did have was a spreadsheet with six years of monthly expenditure data and a free ChatGPT account. She copied her data into the chat, asked it to identify which spending categories had grown faster than inflation over the period, and asked it to write a plain-language explanation she could present to the school board. In 20 minutes, she had a clear narrative identifying that special education transportation costs had risen 34% in three years, driven by increased enrollment and fuel costs, while instructional supply budgets had stayed flat.

She used that output as a first draft, verified the percentages against her own spreadsheet (they matched), added local context the AI couldn't know, and presented it to the board. The conversation that followed led to a successful grant application for a shared transportation consortium with neighboring districts. The AI didn't write the grant. It didn't attend the board meeting. It helped one person without a big budget or technical team do in 20 minutes what would have taken two evenings. That's the accessible version of AI in finance, and it's available right now.

Putting It Into Practice Without Overcomplicating It

The professionals who get the most value from AI in finance aren't the ones with the most sophisticated tools. They're the ones who've identified three or four specific, recurring tasks where volume or formatting slows them down, and applied AI there deliberately. A sales operations manager might use Claude to summarize pipeline reports every Monday. An HR director might use ChatGPT to draft budget justification memos for headcount requests. A small business owner might use Gemini to compare two quarters of expenses and write a plain summary for their accountant. These aren't dramatic transformations. They're consistent time savings that compound.

The discipline required is small but important. Every time you use AI for financial work, confirm one thing before you act on the output: where did the numbers come from, and did the AI have access to the right data? If you typed the data in yourself, you already know. If the AI pulled from a connected system, verify the date range and scope. This 60-second check is the difference between AI as a reliable workflow tool and AI as a liability. It's not distrust, it's professional practice, the same way you'd reconcile a bank statement even if you trusted your bookkeeper.

Start with one use case this week. Pick the financial task that eats the most time with the least creative payoff, formatting a report, summarizing actuals, drafting a budget narrative. Run it through ChatGPT or Claude. Verify the output. Use it. Next week, add one more. Within a month, you'll have a personal AI workflow that's faster, more consistent, and still entirely under your control. That's the version of AI in finance that actually works for non-technical professionals, not the version that runs autonomously in the background, but the version that makes your Monday morning noticeably better.

1. AI in finance works best as a first-pass filter, not a final authority, humans verify and decide.
2. The biggest risk isn't AI making errors; it's professionals skipping the verification step because the output looks polished.
3. Expense anomaly detection, variance analyzis, trend summarization, and budget narrative drafting are all accessible to non-technical professionals using free tools today.
4. A simple structural check, confirming data source, date range, and scope, catches the majority of AI errors before they reach anyone else.
5. Small, consistent AI use cases compound into significant time savings without requiring any technical skill or expensive software.
6. AI cannot know your organizational context, upcoming events, or the business reasons behind numbers, that knowledge is yours, and it must be added before any AI output is shared.
7. The professionals getting the most value from AI in finance are not the most technically sophisticated, they are the most deliberate about where and how they apply it.

• AI is a spotlight, not a verdict, it narrows the field, you make the call.
• Efficiency gains disappear fast if verification steps get skipped, build the check in structurally.
• Free tools like ChatGPT and Claude can handle real financial summarization and analyzis tasks right now.
• Your contextual knowledge, the business reasons, the upcoming events, the relationships, is what makes AI output actually useful.
• Start with one recurring task, verify the output, use it, and add one more next week.