AI in FP&A: A Practical Guide for Finance Teams

2026-10-08 · Tommaso Maria Ricci

AI in FP&A is the most talked about and least practiced idea in corporate finance right now. The 2025 AFP FP&A Benchmarking Survey, built on answers from 362 finance and FP&A professionals, found that only 23% of FP&A teams use AI on a daily, weekly or monthly basis. Another 40% are testing it and plan to implement within a year. The remaining 36% are not even experimenting. Meanwhile, 92% of the same teams still open a spreadsheet every single day to plan the business.

That gap is the whole story. Finance leaders have been told for three years that AI will transform forecasting, budgeting and variance analysis. Most of them have seen a demo. Very few have changed how a forecast is actually produced on a Tuesday afternoon.

This guide is about closing that gap. It covers where AI in FP&A pays first, what it quietly breaks if you deploy it carelessly, how to score your own readiness, and a 90-day plan that a mid-market finance team can run without hiring a data science department. I have spent twenty years building companies and, in the last few, helping other founders and CFOs put AI to work on real numbers. The pattern is always the same: the technology is the easy part, the data and the process are where projects live or die.

Why AI in FP&A Lags Every Other Finance Use Case

Finance functions as a whole have adopted AI faster than FP&A specifically. Gartner's 2025 survey of 183 CFOs and senior finance leaders found that 59% of finance functions now use AI somewhere, up from 58% in 2024 and 37% in 2023. The momentum has clearly slowed. Marco Steecker of Gartner Finance put it plainly: the sharp jump happened between 2023 and 2024, and since then adoption has flattened.

Look at where that 59% actually lands and the FP&A problem becomes obvious. The most common use cases among adopters are knowledge management (49%), accounts payable automation (37%) and error and anomaly detection (34%). These are transactional, rules-heavy, high-volume activities. They are exactly what machine learning and large language models handle well out of the box.

Planning is different. A forecast is a judgment wrapped in assumptions wrapped in politics. The data lives in forty spreadsheets, three ERP instances and the sales director's head. The output has to survive a board meeting. That is why FP&A is where AI projects stall, and also why it is where the payoff is largest once they work.

Three structural reasons explain the lag:

Data fragmentation. The AFP survey reports that only 25% of teams have automated the creation of a single version of the truth for planning. Without it, every AI model trains on contradictory inputs.
Skill gap, not tool gap. The top barrier cited by FP&A practitioners, as CFO.com reported from the AFP data, is lack of expertise, scored 3.8 out of 5, ahead of unclear ROI at 3.7. The tools exist. The people who can connect them to the planning cycle are rare.
Trust. Finance signs its name to the numbers. A vendor survey released in October 2026 by Datarails, covering 270 CFOs at US organizations with more than 1,000 employees, found that finance teams spend about 26% of their working week verifying or correcting AI output. It is a vendor study and should be read as such, but the direction matches what I see: unverified AI output in finance is not a time saver, it is a liability.

The conclusion is not that AI in FP&A does not work. It is that it works only when you treat it as a planning process redesign with software attached, not as software with a planning process attached.

What AI in FP&A Actually Does Today

Strip away the vendor language and AI in FP&A does five things. Each one has a different maturity level, a different data requirement and a different payback period.

Driver-based forecasting with machine learning

Classic FP&A forecasting takes last year, applies a growth rate, and adjusts by gut. Machine learning forecasting takes the historical series plus the drivers (volume, price, headcount, seasonality, pipeline, macro indicators) and learns which drivers explain variance. The output is a forecast with a confidence interval, not a single number.

This matters most in businesses with seasonality or volatile demand. For a hotel group I worked with, the revenue forecast had been produced by the general manager from occupancy memory. We rebuilt it on booking pace, channel mix, lead time and local event calendars. Revenue went from 9 million to 10 million in the following year, and most of that gain came from pricing decisions the model surfaced weeks earlier than the old process would have.

Automated variance analysis and commentary

Month-end variance analysis is where FP&A analysts lose entire days. Pull actuals, compare to budget, identify the big movers, write the explanation, format the deck. Large language models connected to the ledger and the budget file can now produce the first draft of that commentary: "Marketing spend 14% over budget, driven by a 22% increase in paid search in region North, partially offset by lower agency fees."

The analyst's job shifts from writing the explanation to checking it and adding the context the model cannot see. In practice this cuts the commentary cycle from days to hours, and it is the single use case with the fastest payback because it requires no new data, only connection to data you already have.

Scenario modeling at scale

Running three scenarios in a spreadsheet is manageable. Running thirty, each with different assumptions on price elasticity, hiring pace, churn and FX, is not. AI-enabled planning tools can generate and evaluate scenario trees in minutes, and more importantly they can show which assumption the outcome is most sensitive to.

The value is not the thirty scenarios. It is knowing that your full-year EBITDA is three times more sensitive to gross margin than to volume, which changes where the CFO spends their attention.

Anomaly detection in actuals

Before the forecast can be trusted, the actuals have to be clean. Anomaly detection models flag transactions, journal entries and allocations that look wrong relative to history: a cost center that doubled, a vendor invoice booked twice, revenue recognized in the wrong period. Gartner's 2025 data puts this at 34% adoption among finance functions using AI, which makes it one of the most mature use cases. For FP&A it is foundational, because every downstream model inherits the errors you do not catch here.

Conversational analysis for business partners

The newest category: a finance business partner in operations types "why did logistics cost per unit rise in Q3" and receives a data-backed answer with the drivers ranked. This is where generative AI meets planning data, and it is the use case with the most hype and the least maturity. The AFP survey found that only 8% of FP&A professionals use generative AI daily. It will grow, but it depends on everything above it being in place first.

How to Use AI in FP&A: The Five Workflows That Pay First

If you ask vendors how to use AI in FP&A, they will start with a platform. If you ask finance teams that have done it, they start with a workflow. Here is the order that has produced returns in the companies I have worked with, ranked by speed of payback.

1. Month-end variance commentary. Payback in the first close cycle. Required: access to actuals and budget by cost center, and a language model with a controlled prompt template. Risk: low, because every output is reviewed before it leaves finance.

2. Cash flow forecasting. Payback in one quarter. Required: AR and AP aging, historical collection and payment patterns, bank balances. Machine learning on collection behavior outperforms the standard "DSO times revenue" shortcut in almost every business with more than a few hundred customers.

3. Demand and revenue forecasting. Payback in two quarters. Required: at least 24 months of clean sales history at the right granularity (SKU, customer, region) plus drivers. This is where the hotel example above sits, and where a sports distribution company I worked with built the demand signal that fed an AI-driven marketing program and lifted sales by 30%. The finance side of that project was the forecast; the marketing side acted on it. I covered the forecasting mechanics in more depth in my guide to AI for demand forecasting.

4. Headcount and opex planning. Payback in one budget cycle. Required: HRIS data, hiring plans, compensation bands. AI helps less with the arithmetic and more with catching inconsistencies: a department planning 12 hires with no corresponding software seats, or a travel budget flat while headcount grows 40%.

5. Scenario and sensitivity analysis. Payback at the next board meeting where someone asks "what if." Required: a driver-based model that already exists. If your plan is still a static spreadsheet with hardcoded numbers, this step cannot be automated. Fix the model first.

Notice what is not on the list: fully automated budgeting. The AFP survey shows only 18% of teams have automated baseline budgets and 18% baseline forecasts. The reason is not technical. Budgets are negotiated, and no model negotiates for you.

The Data Foundation: What You Need Before Any Model

Every failed AI in FP&A project I have seen failed on data, not on algorithms. Here is the minimum foundation, in order of importance.

A single chart of accounts and a single planning hierarchy

If the ERP, the CRM and the planning spreadsheet disagree on what a "region" or a "product line" is, no model can reconcile them. The fix is boring: a master data map, owned by finance, with a named person responsible for changes. Companies that skip this spend months later arguing about whether the model is wrong or the data is.

At least 24 months of actuals at the forecast granularity

Machine learning forecasting needs history. Twelve months is not enough to learn seasonality. Thirty-six is better. If you only have aggregate data, you can only forecast aggregates, which is usually not where the decisions are.

Driver data, not just financial data

The forecast improves when the model can see what drives revenue and cost: pipeline stages, web traffic, occupancy, production volume, headcount, price lists, weather if you are in retail or hospitality. Most of this lives outside finance. Getting it requires a conversation with sales, operations and marketing, and that conversation is usually the real project.

Documented assumptions

Every forecast embeds assumptions: churn rate, price increase timing, FX rate. If those live in an analyst's head, the model cannot use them and the auditors cannot trace them. Write them down in a structured assumptions register. This also becomes the input for scenario analysis later.

Clean actuals, verified by anomaly detection

Run anomaly detection on the last 24 months before training any forecasting model. You will find errors. Every company does. Fix them in the source system, not in the model.

A useful cross-check is the approach I described in the guide to AI in financial reporting: if the reported numbers cannot be reconciled to the ledger automatically, the planning numbers built on them cannot be either.

Build, Buy or Extend: Choosing the Right Approach

Finance teams face three paths, and the right one depends on size, complexity and what you already own.

Extend what you have. Most ERP and planning platforms now ship AI features: anomaly flags, forecast suggestions, natural language queries. For a company under 50 million in revenue, this is almost always the right first step. The cost is near zero, the integration is already done, and the features are good enough for variance commentary and basic forecasting. The AFP survey shows only 12% of FP&A teams use off-the-shelf tools with built-in AI daily. That number should be higher. Most teams are leaving paid-for capability unused.

Buy a dedicated FP&A platform with AI. For mid-market companies with 50 to 500 million in revenue, multiple entities and a planning cycle that spans departments, a dedicated platform pays for itself in consolidation time alone. The AI layer is a bonus. Evaluate on three criteria from the AFP data: defined roles for finance and business data entry (69% of practitioners rate this essential), a robust security framework (67%) and the ability for finance to administer the tool with minimal IT support (62%). Any vendor that cannot show all three in a live demo on your data is selling a roadmap.

Build custom models. Reserved for companies where forecasting is a competitive weapon: large retailers, marketplaces, subscription businesses with millions of customers, financial institutions. Here a data science team builds proprietary models on top of a data warehouse. Payback is longer and the key person risk is real. I walked through this decision in detail in build vs buy for AI software, and the conclusion holds for FP&A: build only what differentiates you, buy everything else.

A fourth option, increasingly common, is a hybrid: a bought planning platform for consolidation and workflow, plus a small custom model for the one forecast that matters most, fed back into the platform.

Where AI in FP&A Goes Wrong: Seven Failure Patterns

I keep a list of how these projects fail. It has not changed much in three years.

1. Automating a broken process. If the budget process takes four months because every department negotiates line by line, AI will produce a faster version of the same four-month negotiation. Redesign the process first, then automate.

2. Forecasting the wrong granularity. A model that forecasts total revenue beautifully is useless if decisions are made by product line. Match the model to the decision.

3. Trusting unverified output. The 26% of the week spent fact-checking AI in the Datarails survey is the cost of skipping this step. Every AI output that leaves finance needs a named reviewer and a documented check. This is not bureaucracy; it is what makes the output usable.

4. No feedback loop. A forecast that is never compared to actuals never improves. Build the forecast-versus-actual review into month-end, feed the error back to the model, and track forecast accuracy as a KPI.

5. Skipping the explanation. A board will not accept "the model says so." Every forecast needs its top three drivers stated in plain language. If the tool cannot explain, the analyst has to, which means the analyst has to understand the model.

6. Underestimating change management. Analysts who have built the forecast by hand for eight years will not hand it to a model without a fight, and some of their resistance is justified. Involve them in model design, make them the reviewers, and reward forecast accuracy, not hours spent.

7. Starting with generative AI chat. The conversational interface is seductive and useless without clean data underneath. Build the foundation, then the models, then the chat layer. Gartner's finding that 91% of finance functions see low or moderate impact initially is largely this: they started with the visible layer and skipped the invisible one.

The Human Role: What Analysts Do When the Model Forecasts

A legitimate question from every FP&A team: if the model builds the forecast and writes the commentary, what do I do? The honest answer is that the work changes shape and, in my experience, becomes more valuable.

From producing numbers to interrogating them. The analyst's time moves from building the forecast to challenging it. Why does the model expect a Q4 dip? Is it seasonality or the loss of a customer it cannot see? That interrogation is where finance adds value, and it was always crowded out by production work.

From reporting to partnering. With commentary drafted automatically, the finance business partner spends the saved hours in the operating review, not in the deck. The AFP data is clear that FP&A teams want this: the number one wish list item across the survey is more time for analysis and less for data gathering.

From spreadsheet owner to model owner. Someone has to own the assumptions register, the driver definitions, the forecast accuracy KPI and the review process. That is a more senior, more durable role than spreadsheet maintenance.

The teams that handle this transition well reskill before they deploy. The teams that handle it badly deploy, watch adoption stall, and conclude the tool was wrong. My playbook on enterprise AI adoption covers the organizational side in more detail, and almost all of it applies directly to finance.

Self-Assessment: Is Your FP&A Function Ready for AI?

Score one point for each statement that is true today. Be honest; the point of this exercise is to find the gaps before a vendor finds them for you.

Data (maximum 5)

We have one agreed planning hierarchy used by finance, sales and operations.
We have at least 24 months of actuals at the granularity we forecast.
Non-financial drivers (pipeline, volume, headcount, traffic) are available monthly in a structured form.
Our assumptions are written in a register, not in people's heads.
We run some form of anomaly or error check on actuals before close.

Process (maximum 5)

Our budget cycle takes less than three months end to end.
We compare forecast to actuals every month and track accuracy.
Variance commentary follows a standard template.
Scenario requests from leadership can be answered within a week.
One person owns the planning model and its changes.

People (maximum 5)

At least one FP&A team member can write a basic SQL query or a Python script.
The team has used an AI feature in an existing tool in the last 90 days.
Analysts would describe their job as analysis, not data gathering, at least half the time.
The CFO has stated AI in finance as a priority with a budget line.
There is a named reviewer for any AI output that leaves finance.

Interpretation

12 to 15: You are ready to deploy. Start with variance commentary and cash forecasting this quarter.
8 to 11: You have a foundation with specific gaps. Fix the data items first; they block everything else.
4 to 7: Do not buy AI tooling yet. Spend the next quarter on the planning hierarchy, the assumptions register and the forecast accuracy review.
0 to 3: The problem is not AI. It is that the planning process itself is not yet a process. Build that first.

This is the same diagnostic I run with finance teams at the start of an engagement, and in most mid-market companies the score lands between 6 and 9. That is not bad news. It means the first three months of work are cheap, internal and entirely within finance's control. If you want a second pair of eyes on where your function sits and which gap to close first, you can reach me through the consultation request page on this site and we can walk through it together.

A 90-Day Roadmap for AI in FP&A

This is the roadmap I recommend to mid-market finance teams. It assumes you scored at least 8 on the assessment. If you scored lower, prepend a quarter of foundation work.

Days 1 to 30: Foundation and first win

Week 1 and 2: Data audit

Map every data source feeding the current forecast: ERP, CRM, HRIS, spreadsheets, manual inputs.
Define the planning hierarchy and get sign-off from sales and operations.
Run anomaly detection on the last 24 months of actuals. Log and fix what you find.

Week 3 and 4: Variance commentary pilot

Connect a language model to last month's actuals and budget at cost center level.
Write a prompt template that produces commentary in your house format.
Have two analysts review the output side by side with their own write-up. Measure time saved and errors caught.

Expected output: A clean data map, a fixed set of historical errors, and a variance commentary draft that takes hours instead of days to finalize.

Days 31 to 60: Forecasting pilot on one line

Week 5 and 6: Pick the forecast that matters

Choose one P&L line where forecast error is costly and history is clean. Usually revenue for one business unit or cash collections.
Assemble drivers for that line. Interview the business owner about what moves it.
Set a baseline: what is the current forecast accuracy (mean absolute percentage error) over the last 12 months?

Week 7 and 8: Model and compare

Build or configure the forecasting model, using built-in platform features if available.
Run it in parallel with the manual forecast for the month. Compare both to actuals.
Document the top three drivers the model found and sanity check them with the business.

Expected output: A measured accuracy comparison, a documented driver set, and a decision on whether to extend.

Days 61 to 90: Scale what worked, govern what you built

Week 9 and 10: Extend

Roll variance commentary to all cost centers.
Extend forecasting to the second and third most valuable lines.
Add the forecast accuracy KPI to the monthly finance dashboard.

Week 11 and 12: Governance

Write the AI output review policy: who reviews what, with which check, before it leaves finance.
Assign ownership of the assumptions register and the model.
Train the broader team on how to interrogate model output and when to override it.
Calculate the pilot's return and present it to leadership with next quarter's plan.

Expected output: AI embedded in two recurring FP&A workflows, a governance framework, and a business case for the next phase.

The 90 days are deliberately modest. The companies that try to transform the entire planning cycle in one quarter end up in Gartner's 25% who are "unsure how to move from planning to piloting." Small, measured, repeated wins build the trust that lets you go further.

Measuring the Return on AI in FP&A

Finance should hold its own AI projects to the standard it holds everyone else's. These are the metrics that matter, in three tiers.

Efficiency metrics

Hours per close cycle spent on variance analysis and commentary, before and after.
Days from period end to forecast delivery.
Number of scenarios that can be produced per request.

Quality metrics

Forecast accuracy, measured as mean absolute percentage error by line and by horizon.
Number of actuals errors caught by anomaly detection before close.
Share of variance commentary accepted by business owners without rework.

Business metrics

Working capital improvement from better cash forecasting.
Decisions changed or accelerated because of scenario analysis, logged qualitatively.
Margin impact from pricing or cost decisions surfaced by the model, as in the hotel example above.

The first tier justifies the pilot. The second tier justifies the platform. The third tier justifies the CFO's attention. I use a broader version of this structure in my guide to AI ROI for business, and finance teams tend to find the three-tier split easier to defend at budget time than a single blended return figure.

One warning on measurement: forecast accuracy can be gamed by forecasting conservatively. Track bias (consistent over or under forecasting) alongside error, or the model will learn to sandbag just like humans do.

AI in FP&A by Company Size

The right approach scales with the company. Here is what works at each stage, based on the engagements I have seen succeed.

Startups and companies under 10 million in revenue. FP&A is often one person and a spreadsheet. AI here means using the features already in your accounting and planning tools, plus a language model for commentary and board deck drafting. Do not buy a platform. Do build the assumptions register from day one; it is the single most valuable habit for a growing finance function.

Mid-market, 10 to 100 million. This is where the 90-day roadmap fits exactly. Usually a small FP&A team, a planning tool or an advanced spreadsheet, and enough history to forecast. The biggest win is typically cash forecasting, because mid-market companies are often tighter on working capital than they admit.

Mid-market, 100 to 500 million. Multiple entities, consolidation pain, and a planning cycle that involves ten or more department heads. A dedicated planning platform with AI features is almost always justified. The second biggest win after consolidation is driver-based forecasting for the two or three lines that dominate the P&L.

Enterprise, above 500 million. Custom models become viable, data engineering becomes necessary, and governance becomes the main risk. The failure mode here is not lack of capability but lack of coordination: three business units building three incompatible forecasting models. Centralize the data foundation, decentralize the use.

For private equity backed companies the calculus shifts again, because the sponsor wants portfolio-wide comparability. I covered that in AI for private equity, and the FP&A implication is that standardization across the portfolio often matters more than sophistication within one company.

What Changes in the Next Two Years

Three shifts are already visible and will shape how to use AI in FP&A through 2027 and 2028.

Agents that run the close checklist. AI agents that execute multi-step tasks, such as pulling actuals, running variance checks, drafting commentary and routing exceptions, are moving from demo to deployment. The constraint is not the agent; it is whether your close process is documented clearly enough for an agent to follow. Teams that documented their process in the last two years will deploy agents in weeks. Teams that did not will spend the first six months writing the documentation.

Forecasting becomes continuous. The monthly forecast cycle exists because producing a forecast was expensive. When the model reruns nightly, the question changes from "what is the forecast" to "what changed since yesterday and why." This is a bigger cultural shift than it sounds, because it removes the ritual that structures most finance calendars.

Verification becomes the job. As output generation gets cheap, the value moves to checking it. The quarter of the week that finance teams currently spend verifying AI output will not disappear; it will become formalized, tooled and owned. Finance functions that build verification discipline now will be the ones trusted with autonomy later.

None of these require you to wait. The foundation work in the roadmap above is exactly what each of them depends on.

From Reading to Doing

The data at the top of this guide bears repeating: 23% of FP&A teams use AI regularly, 40% are testing and planning, and 92% still plan in spreadsheets every day. The gap between what finance leaders expect from AI and what their teams have deployed is wide, and it is mostly a data and process gap, not a technology gap.

That is good news for any team willing to do the unglamorous work. The planning hierarchy, the assumptions register, the forecast accuracy review and the variance commentary pilot are all within a finance team's control, cost little, and build on each other.

If you are a CFO or an FP&A leader trying to decide where to start, or you have already started and the pilot has stalled, the most useful next step is a structured look at your data, your process and your team against the assessment above. I do this with finance leaders regularly, and the consultation request page on this site is the place to open that conversation. Bring the spreadsheet; that is where the real diagnosis begins.

For the wider view of how AI reshapes the finance function beyond planning, my guide for CFOs on AI for finance leaders and the practical guide to AI in accounting cover the transactional and reporting side that FP&A sits on top of.

FAQ

How do I start using AI in FP&A if my data lives in dozens of spreadsheets?

Start with the one workflow that needs the least new data: variance commentary. Connect a language model to last month's actuals and budget at cost center level and let it draft the explanation, then have an analyst review it. In parallel, build a single planning hierarchy and an assumptions register. Those two artifacts turn spreadsheets into something a forecasting model can learn from. Do not buy a platform until the hierarchy exists, or you will pay to consolidate chaos.

What is the realistic cost of AI in FP&A for a mid-market company?

For companies under 50 million in revenue, the first phase costs almost nothing beyond time, because the AI features already shipped in your ERP or planning tool cover variance commentary and basic forecasting. A dedicated FP&A platform with AI for a company between 50 and 500 million typically runs from the low tens of thousands to a few hundred thousand per year depending on entities and users, plus an implementation project of two to four months. Custom models require a data engineer and a data scientist and only make sense above that range.

Will AI replace FP&A analysts?

No, but it will change what they do. The production work, pulling data, building the forecast mechanically and writing the first draft of commentary, moves to the model. The analyst's work shifts to interrogating the forecast, owning the assumptions, reviewing AI output before it leaves finance, and partnering with business leaders. The AFP survey shows FP&A teams already want more time for analysis and less for data gathering, and AI is the first technology that actually delivers that shift at scale.

How accurate is AI forecasting compared to a manual forecast?

It depends on the data, not the algorithm. With 24 or more months of clean history at the right granularity plus driver data, machine learning forecasts typically beat manual forecasts on mean absolute percentage error, and the gap widens in seasonal or volatile businesses. With short or dirty history, the model is no better than the analyst and may be worse because it lacks context. Always run the model in parallel with the manual forecast for at least one cycle and measure both against actuals before trusting either.

How to use AI in FP&A without breaking audit and control requirements?

Treat every AI output as a draft that needs a named reviewer and a documented check before it leaves finance. Keep the assumptions register and the model configuration under version control. Log which outputs were overridden and why. Run anomaly detection on actuals before any model trains on them. These four practices satisfy most internal audit requirements and, more importantly, they are what makes the output trustworthy enough to act on.

Which FP&A use case should a finance team automate first?

Month-end variance commentary, because it needs no new data, pays back within the first close cycle and carries low risk since every output is reviewed. Second is cash flow forecasting, which uses AR and AP data most companies already have and directly improves working capital. Demand and revenue forecasting comes third, once you have at least two years of clean history and driver data. Fully automated budgeting should come last, if ever, because budgets are negotiated and no model negotiates for you.

What skills does an FP&A team need to use AI well?

At least one person who can query data directly, with SQL or a basic script, so the team is not dependent on IT for every data pull. Everyone needs to understand what a driver-based model is and how to read a confidence interval. The most important skill is not technical: it is the discipline to challenge a model's output the same way a good analyst challenges a department head's forecast. Reskill before deploying, make analysts the reviewers, and reward forecast accuracy rather than hours worked.