How FP&A Teams Are Using AI for Forecasting

Financial planning and analysis is where AI promises the most and where it is easiest to get wrong. FP&A forecasts look months and years ahead, cover revenue, margin, overheads and headcount, and feed budgets, board packs, investor updates and bank covenants. Machine learning can find patterns in sales and cost data that a spreadsheet never will, and generative AI can draft the commentary that used to take a week of the month. But a plan is a set of choices about the future, and no model can make those choices for the business.

This guide is for Heads of FP&A, FP&A Managers and Analysts, Finance Business Partners and the finance leaders hiring them. It covers how AI and ML fit alongside driver-based planning, demand forecasting, anomaly detection, AI-drafted commentary, the main types of planning platform, model governance, the skills FP&A teams now need, and what FP&A roles currently pay.

It deliberately does not cover the weekly cash forecast. Short-term cash forecasting is transaction-driven and measured against the bank each week; that is covered in our companion guide to AI for cash flow forecasting. This page is about the revenue and P&L forecast and the planning cycle.

Where AI Fits in the FP&A Cycle

A typical FP&A cycle runs from data collection, through forecasting and scenario work, to reporting, commentary and decision support. AI can contribute at each stage, but the value is uneven:

The pattern is consistent. AI is strongest where there is lots of history and the question is “what happens if things carry on”, and weakest where the question is “what should we do”. That is why it works best inside a driver-based model rather than instead of one.

Driver-Based Models Plus Machine Learning

A driver-based model expresses the P&L as the result of a small number of operational drivers: customers multiplied by average order value, billable heads multiplied by utilisation and rate, units multiplied by price less unit cost, and so on. The method is set out in our guide to driver-based planning for management accountants. Its strength is that managers can see and argue about the assumptions.

Machine learning is most useful in forecasting the drivers themselves, not the P&L lines. Rather than asking a model to predict “revenue”, the team asks it to predict new customer volumes, churn, order frequency or utilisation, using history plus relevant external signals. The driver-based model then turns those forecasts into revenue, margin and overheads using logic the business understands. This hybrid has three advantages:

• Explainability. When the forecast moves, you can say which driver moved and why, rather than pointing at a black box.
• Override points. Commercial teams can override a driver (a new contract, a price rise, a lost customer) without breaking the model.
• Separation of baseline and plan. The ML forecast gives an unbiased “if nothing changes” baseline. Management initiatives and targets sit on top as clearly labelled adjustments, so the gap between baseline and plan is visible.

That last point is often the biggest win. Many budgets blend what is likely with what is hoped for. A separate statistical baseline makes the stretch explicit, which is exactly what boards want to see. It also supports a better rolling forecast, because the baseline can be refreshed automatically each month while the team focuses on the adjustments.

Demand and Revenue Forecasting

Demand forecasting is the most mature use of ML in planning, particularly in retail, consumer goods, distribution, subscription and hospitality businesses with high transaction volumes. Typical inputs include:

• Sales history by product, channel, region and customer segment, ideally weekly or daily.
• Seasonality, calendar and promotional effects: bank holidays, school holidays, events, marketing campaigns and price changes.
• Pipeline and leading indicators: CRM pipeline by stage, web traffic, quotes issued, renewal dates.
• External data where it genuinely explains demand: inflation, interest rates, consumer confidence or weather. Official series from the Office for National Statistics and the Bank of England’s Monetary Policy Report are common starting points for macro assumptions.

For business-to-business companies with a few large customers and long sales cycles, pure ML demand models add less. There, AI is more useful in scoring the pipeline (estimating the probability and timing of each deal from historical win rates) than in predicting revenue directly. Either way, the measure of success is the same: forecast accuracy and bias tracked over time against a simple benchmark such as last year plus growth. If the model cannot beat the simple benchmark consistently, it is not adding value.

Cost forecasting follows similar logic. Headcount, payroll and run-rate overheads are largely deterministic and are better handled by rules than by ML. Variable costs linked to volume, such as materials, freight and commissions, can benefit from the same driver forecasts used for revenue.

Anomaly Detection

Anomaly detection is a lower-profile but highly practical use of ML in FP&A. Models learn the normal pattern of each account, cost centre or product line and flag movements that fall outside it. Useful applications include:

• Actuals. Catching mispostings, missing accruals and unusual costs before the management accounts go out, which helps the month-end team as much as FP&A.
• Submitted forecasts. Flagging budget-holder submissions that break from their own history or from peers, such as a cost centre forecasting flat spend despite a hiring plan.
• KPI drift. Spotting early changes in churn, margin mix or conversion rate that would otherwise surface only at quarter end.

The key control is tuning. Too sensitive and the team drowns in false alarms and starts ignoring them; too loose and real problems slip through. Start with a small number of high-value accounts and expand once the team trusts the alerts. Every flag still needs a person to investigate and decide whether it matters.

AI-Drafted Commentary

Generative AI has changed the reporting end of FP&A faster than the forecasting end. Given a variance table and some context, a large language model can produce a competent first draft of month-end commentary, forecast change explanations or a board summary in minutes. Used well, this frees analysts to spend their time on the questions behind the numbers.

The risks are specific. Language models can misstate figures, attribute a variance to the wrong cause, or produce fluent explanations that sound plausible but are not true. They also tend to describe what changed rather than why it matters. A good working pattern is:

• Feed the model the verified variance table and the known business reasons (from budget-holder conversations), not raw ledger data.
• Ask it to draft, then check every number against the source and every causal claim against what the business actually said.
• Add the judgement yourself: what it means for the full-year outturn and what management should do.

We cover this in more detail in AI for variance commentary and AI for board pack drafting. Prompt patterns are in our guide to prompt engineering for finance.

Scenario Planning at Speed

One of the most valuable effects of AI-enabled planning is speed. When a driver-based model sits on a modern platform, the team can run dozens of scenarios in the time it used to take to run three. ML can also produce probability ranges around the baseline, so a board sees not just a central case but how wide the uncertainty is.

Speed is only useful if the scenarios are meaningful. The choice of which scenarios to run (a lost major customer, a price war, a cost shock, a delayed product launch) and how to respond to them is a management conversation. Our guide to scenario and sensitivity analysis covers how to frame them. The budgeting process around them is in the budgeting and reforecasting guide.

Planning Platforms: A Neutral Overview

Most AI forecasting in FP&A now happens inside a planning or analytics platform rather than in a standalone tool. The market broadly divides into:

• Enterprise connected-planning platforms such as Anaplan, Oracle EPM and SAP Analytics Cloud, suited to large, complex organisations with many planning users.
• Mid-market and newer planning platforms such as Pigment, Workday Adaptive Planning, Planful, Jedox and Board, many of which now include built-in forecasting and AI assistants.
• Spreadsheet-centred approaches, using Excel with AI features and add-ins, often combined with a BI layer such as Power BI for reporting. Still the norm in many owner-managed businesses.
• Data-science tooling, such as Python or R models built by an analytics team and fed into the planning model, used where forecasting needs are specialised.

The right choice depends on the size of the business, the complexity of the model, the number of contributors and the skills in the team, not on which vendor has the loudest AI marketing. Implementation effort, data integration and the availability of people who can run the platform usually matter more than feature lists. Our comparison of FP&A tooling: Anaplan, Pigment and Power BI goes further, and the AI tool landscape for finance covers general-purpose assistants. For spreadsheet-based teams, see using AI inside Excel for finance.

Model Governance

Once ML forecasts start shaping budgets and guidance, they need the same governance as any other material model. Banks are subject to formal model risk rules; the PRA’s model risk management principles for banks (SS1/23) is written for them, but its core ideas translate well to any FP&A team. A proportionate framework covers:

• Inventory and ownership. A list of every forecasting model in use, what it feeds, and a named owner accountable for it.
• Documentation. Inputs, method, key assumptions, known weaknesses and the date of the last retrain.
• Validation and back-testing. Accuracy and bias measured against actuals and against a simple benchmark, reviewed at least quarterly.
• Override logging. Every manual adjustment to a model output recorded with the reason and the person, so you can see whether overrides improve accuracy or just add optimism.
• Change control. Model and vendor updates logged, with accuracy watched closely after each change.
• Data protection and security. Particularly where models use customer or employee data; the ICO’s guidance on AI and data protection applies.

These sit alongside your AI usage policy for finance teams and the review steps in human-in-the-loop AI controls. Boards and audit committees increasingly ask about this; see what boards are asking about AI in finance.

Risks and Limits

• Structural breaks. Models trained on stable years struggle with pandemics, inflation spikes, new competitors or a changed business model. Judgement must take over when the world changes.
• False confidence. A model output with decimal places looks authoritative. Present ranges and be clear about uncertainty, especially further out.
• Garbage in. Inconsistent cost-centre mapping, restated history or poor CRM data will undermine any model.
• Loss of understanding. If analysts stop building models themselves, the team can lose the ability to challenge the output. Keep people close to the drivers.
• Accountability. The forecast is still the finance team’s and management’s. The ICAEW’s artificial intelligence resources discuss professional responsibility when using AI, and CIMA publishes material for management accountants and FP&A professionals on the same theme.

Skills FP&A Teams Need Now

AI is shifting FP&A work away from compiling numbers and towards interpreting them. In the briefs we see, employers increasingly want a blend of:

• Strong modelling fundamentals. Driver-based and three-statement modelling remain the base; someone who cannot build a model cannot sensibly challenge one.
• Data literacy. Comfort with data structures, SQL or Power Query, and enough statistics to understand forecast error, bias and overfitting. Python is valued in larger teams but not essential everywhere.
• Platform experience. Hands-on use of a planning platform, which is often the deciding factor in senior FP&A hires.
• Critical review of AI output. Checking numbers, testing causal claims and knowing when to override.
• Business partnering. Translating forecasts into decisions with commercial and operational leaders, which AI does not do.

The route from Analyst to Head of FP&A, and how these skills fit at each level, is covered in our FP&A career path guide. For interview questions that test AI judgement as well as modelling, see FP&A interview questions.

FP&A Pay

Candidates who combine modelling, platform skills and sound AI judgement are in demand. Typical London base salaries are:

Regional salaries are generally lower, and regulated-sector roles can command a premium. Planning platform expertise and data skills tend to push candidates towards the top of each band. More detail is in the FP&A Manager salary guide, the FP&A Analyst job description and salary page, and our page on the Head of FP&A role and when to create it. For budget-season or platform-implementation cover, see interim FP&A budget-season cover.

We recruit FP&A professionals at every level through our FP&A recruitment service, with permanent shortlists typically within one to two weeks and interim or fractional shortlists within days. Where the brief is a CFO or FD to lead finance transformation, we work with our sister firm FD Capital.

Frequently Asked Questions

Will AI replace FP&A analysts?

It is replacing parts of the job, mainly data gathering, baseline forecasts and first-draft commentary. It is not replacing the need for people who understand the business, build and challenge models and turn forecasts into decisions. In practice, the analyst role is moving towards interpretation and partnering.

Should we use ML instead of a driver-based model?

Usually not instead of, but alongside. Use ML to forecast the drivers (volumes, churn, utilisation) and keep the driver-based logic to turn them into the P&L. That keeps the forecast explainable and lets managers override specific assumptions.

How do we know whether an AI forecast is better?

Back-test it. Compare its accuracy and bias against actuals and against a simple benchmark, such as last year plus growth or your existing method, over several periods. Keep running the old method in parallel until the new one has proved itself.

Do we need a planning platform to use AI in FP&A?

No. Many smaller teams get real value from AI features in Excel and general-purpose assistants for commentary. Platforms become worthwhile when the model, the number of contributors or the data volume outgrows spreadsheets.

How is this different from AI in cash flow forecasting?

FP&A forecasting covers revenue, margin and the P&L over months and years and leans heavily on assumptions about the market. Short-term cash forecasting is transaction-driven and measured weekly against the bank. See AI for cash flow forecasting for that side.

What should we look for when hiring an FP&A lead for AI adoption?

Strong modelling, hands-on planning platform experience, data literacy, and evidence that they have measured forecast accuracy and acted on it. Ask how they would decide whether to keep or drop an ML forecast; good answers talk about back-testing, bias and override analysis.