What AI Changes in Financial Planning and Analysis Work

Which parts of FP&A work are changing first?

AI most readily changes the repeated, digital preparation around financial planning and analysis (FP&A): gathering inputs, organizing them, producing a first draft of commentary, and exploring possible patterns or scenarios. That does not mean the whole analyst role is on a path to disappearance. The work also includes checking reconciliation, assumptions, and implications for a business decision. A useful starting point is therefore to map tasks, rather than label an occupation as exposed or safe. The U.S. occupational profile for Financial and Investment Analysts on O*NET includes activities such as interpreting financial information, preparing reports, financial modeling, monitoring developments, and using spreadsheets, databases, enterprise resource planning systems, and business-intelligence tools. It also covers investment and client work, not corporate FP&A alone. The three-part map is an editorial synthesis, not an official classification or measured ranking. First is preparation: bringing together recurring inputs and making them usable. In a monthly close, that might mean arranging approved actuals by period and business unit or drafting a summary from a reconciled table. In a forecast update, it could mean organizing submitted assumptions. Repetition, digital inputs, stable formats, and outputs that can be checked against a source make these plausible places to investigate assistance. A draft can be compared with approved figures. This is a reason to test, not proof of reliable performance or employer adoption. Second is stewardship: keeping the numbers and model trustworthy. An analyst may need to reconcile totals, trace a metric to its source, check that periods and business-unit definitions match, and understand how a spreadsheet or planning model turns assumptions into results. Automating a copy or a calculation does not settle whether the input was complete, the mapping was appropriate, or a changed definition makes a comparison misleading. System differences, missing context, and review requirements can make routine work harder to hand over. The analyst must verify lineage and explain changes. Third is interpretation: deciding what the movement means and what deserves attention. A variance can be described without explaining its cause. A scenario can be calculated without establishing that its assumptions are plausible. The analyst may need to distinguish a one-time timing effect from a change in demand, identify uncertainty, and communicate implications to the people responsible for operating decisions. A system may surface patterns or draft wording, but does not establish causality or choose an acceptable trade-off. A polished explanation can still rest on a weak assumption. The boundaries are not fixed by job title. One analyst may spend much of a week on repeated reporting; another may work chiefly on model maintenance or discussions with budget owners. Within a forecast cycle, these tasks overlap: an input must be organized, checked, and understood before supporting a recommendation. List recurring tasks and note whether each has a stable source, checkable result, and consequences requiring judgment. This can identify a practical experiment without settling the occupation-level question. The distinction matters when comparing tools: generative systems can draft or summarize, while analytics and rules-based automation may flag patterns or apply defined steps. In all cases, the task still needs a suitable input and a check appropriate to its use.

Sources: Financial and Investment Analysts; Generative AI and jobs: A 2025 update

What does exposure evidence establish—and what does it leave open?

Exposure evidence estimates which work tasks appear technically susceptible to generative AI under a defined task model. It does not answer whether an employer has bought or approved a tool, whether the tool can complete a finance process reliably, how management will redesign roles, whether demand for analysts will change, or whether a particular worker will lose a job. Treat exposure as a reason to inspect your work, not as a personal forecast. The International Labour Organization’s *Generative AI and jobs: A 2025 update* summarizes a global occupational assessment built from task-level data, expert input, and AI predictions. The method evaluates nearly 30,000 tasks at six-digit occupational detail and assigns four exposure gradients based on scores and variation among tasks. The brief estimates one in four workers worldwide are in occupations with some GenAI exposure and says most jobs are more likely to be transformed than made redundant because human input remains necessary. These are modeled occupational findings, not observations of what FP&A teams have deployed or a count of jobs that will disappear. The ILO’s global estimate combines occupations and labor markets; it cannot be read as a rate for financial planning and analysis analysts in a particular country or company. That distinction matters because “financial analyst” is not a precise synonym for corporate FP&A. The U.S. Department of Labor’s O*NET profile, *Financial and Investment Analysts*, includes planning analyst among reported job titles, but also covers investment, valuation, and client work. Its listed tasks include using financial models, interpreting data, monitoring developments, and presenting reports. This profile is useful for seeing the breadth of work grouped under an occupational label. It does not isolate internal budgeting and forecasting, say how many hours an FP&A analyst spends on each activity, or estimate AI exposure. Neither this profile nor the ILO assessment supplies an FP&A-specific task share or an individual replacement estimate. For an analyst, several separate steps sit between exposure and employment consequences: a task must be technically addressable; the organization must permit and adopt a system; it must fit the data, permissions, and existing workflow; reviewers must be able to check its output; and leaders must decide what to do with any time or cost change. Weak data quality, inconsistent definitions, limited system connections, control requirements, or the cost of changing a process can slow or prevent adoption. An approved tool may also assist one step without automating a role. These are reasons to investigate the local workflow, not predictions about a particular employer. So begin with evidence close to your work: identify recurring tasks, note which inputs and outputs are digital and checkable, and find out whether your employer has an approved tool or a production workflow. Ask what has actually changed in review steps, cycle timing, or responsibilities. An exposure label cannot settle those questions. It becomes useful only when paired with observed use and the controls and decisions around it.

Sources: Generative AI and jobs: A 2025 update; Financial and Investment Analysts

How far has adoption moved from capability to regular workflow?

Finance-function AI adoption is real, but surveys do not show how routinely an FP&A analyst’s own tasks are handled by AI. They report what finance leaders say their organizations use, not a representative record of analyst work or job changes. Track whether a workflow has moved from access, through a pilot, into recurring production use with measured effects. Gartner Survey Shows Finance AI Adoption Remains Steady in 2025 reports that 59% of 183 CFOs and senior finance leaders surveyed in May and June 2025 said their finance function used AI. Gartner’s 2024 survey reported 58%. This is leader-reported function-level use, not the share of FP&A analysts using AI or planning tasks automated. Among organizations that had implemented AI, respondents most often named knowledge management (49%), followed by accounts payable automation (37%) and error or anomaly detection (34%). They do not establish how often generative tools prepare forecasts or commentary. The same release shows why “uses AI” can conceal different stages. Gartner says data literacy or technical skills and inadequate data quality or availability remained major obstacles. It reports 25% were uncertain how to move from planning to piloting, while 16% had no implementation planned for the coming year. Even after a pilot begins, 91% of respondents initially observed low or moderate impact. This is reported initial impact, not failure, and does not isolate FP&A. This suggests friction between intent, testing, and useful production; access alone does not show a changed analyst role. A separate survey in How finance teams are putting AI to work today, published by McKinsey, asked 102 CFOs across industries and global regions about generative AI. Forty-four percent said their organizations used it for more than five use cases in 2025, compared with 7% in the previous year’s survey. The studies differ in population, question, technology label, and comparison; the percentages are not a time series. McKinsey also reports that nearly two-thirds of respondents in a separate recent survey said their organizations had not begun scaling AI enterprise-wide. It describes pilots breaking under real-world conditions, failing to adapt as new data emerge, or remaining poorly integrated into core processes. The reports together show activity alongside incomplete integration, not universal routine use. McKinsey describes applications including commentary drafts, performance summaries, scenario modeling, and decision support. These selected examples clarify possible workflows but do not estimate prevalence. A recurring report may already automate fixed steps; adding generated narrative or a natural-language scenario interface is a different change. Calling both “AI adoption” can obscure what shifted and what still needs analyst review. To locate your team’s workflow, ask: Is an approved tool available for this task, or merely discussed? Has anyone tested it on permitted data with a reviewer? Does it run in the recurring close or forecast process with defined ownership and controls? Has the team measured the whole workflow, including corrections, reconciliation, and decision usefulness? Access or a pilot is not the same as sustained production impact. If unclear, ask a manager what evidence would justify expansion. That local check informs a career decision but cannot predict staffing changes. These surveys establish reported organizational use and persistent barriers; they do not reveal one analyst’s exposure, productivity change, or displacement risk.

Sources: Gartner Survey Shows Finance AI Adoption Remains Steady in 2025; How finance teams are putting AI to work today

What changes in the workflow once a tool is actually adopted?

Take a monthly forecast and variance cycle. Actuals arrive from finance systems; operating teams add information about volume, price, staffing, freight, or production; mappings connect inputs to the planning model; an analyst checks calculations, explains movements, and prepares a review for managers deciding what to change. A tool inserted here may retrieve approved commentary, summarize it by business unit, compare it with financial data, or flag a movement outside an agreed range. That can speed a first pass and help the analyst investigate exceptions sooner. It does not establish that an explanation is correct or that a manager should act on it. “Adopted” can mean access to a tool, occasional use for a draft, or integration into a recurring workflow with defined data access, checks, and decision rights. In “How finance teams are putting AI to work today,” McKinsey describes finance use cases including drafting commentary, summarizing performance, supporting scenarios, and surfacing alerts. Its selected cases and proprietary survey do not establish a typical analyst workload or general productivity result. They show possible uses, not a standard outcome. The analyst still needs to check whether the tool used the correct period and business unit, whether a measure means the same thing across systems, and whether the result follows approved model logic. A reported margin decline might reflect a change in mix, a late adjustment, a mapping error, or a mismatch between operational and finance calendars. A fluent summary can omit the detail that distinguishes these explanations. Review should trace numbers to source records, confirm dimensions and definitions, inspect model lineage, and test the explanation against available evidence. O*NET's broad Financial and Investment Analysts profile includes financial modeling, interpreting financial information, monitoring developments, and preparing reports. It is not an FP&A task breakdown, but it shows why the work extends beyond producing a report. “How AI agents can help FP&A better steer the business,” also from McKinsey, describes a more integrated design: agents ingest, clean, and format data, staff validate it, a first-pass forecast goes to human review, and separate agents identify possible actions to reduce plan gaps. The article reports a selected telecommunications-company case, including a threefold reduction in forecasting cycle time, and describes changes in coordination and roles. It is a selected redesign example, not representative evidence. It illustrates assistance versus end-to-end automation: preparation may be automated within defined controls, while choosing assumptions, judging materiality, and authorizing a response still depend on people and organizational rules. Faster preparation may create capacity for earlier analysis, but management decides where that capacity goes. Shorter cycles could let finance examine operating drivers while pricing, spending, or production options remain open. McKinsey describes examples of linking operational signals with financial outcomes and testing scenarios earlier in the period. These cases show a possible mechanism, not proof every team will gain influence or save net time. A department could instead increase report frequency, request more scenarios, or expand commentary. The analyst may then spend less time gathering inputs but more time validating a larger stream of outputs. More generated analysis can increase review work, particularly when definitions or source ownership are unsettled. So the useful measure is not whether one step gets quicker. It is whether the complete cycle improves: trusted inputs, a traceable explanation, a useful decision, and the time spent checking and correcting. A tool that drafts commentary quickly but leaves unexplained differences to reconcile may move effort rather than remove it. One that surfaces a meaningful exception with clear source links may focus investigation. A role's consequences depend partly on whether saved preparation supports interpretation or becomes more frequent reporting. The sources describe opportunity alongside integration, validation, and governance needs; they do not establish a universal reallocation of FP&A work.

Sources: How AI agents can help FP&A better steer the business; How finance teams are putting AI to work today; Financial and Investment Analysts

Why can broader analysis create a new bottleneck?

More sources, scenarios, and polished commentary can widen an analyst’s view without making a forecast more reliable. The scarce work may shift from assembling a first draft to choosing which signals matter, reconciling contradictions, checking assumptions, and explaining uncertainty. A useful warning comes from the working paper “Generative AI for Analysts,” which examines sell-side financial research. It does not measure corporate FP&A, so it is evidence of a possible information-processing bottleneck, not a prediction about an internal planning team. The paper studies the 2023 integration of generative AI into FactSet as a plausibly external change in analysts’ access to the technology. In its September 2026 revision, the authors report that FactSet-associated reports contained 26% more distinct information sources, 24% broader topical coverage, and 21% more analytical methods, with improved timeliness. Yet relative forecast accuracy declined when analysts faced greater information-processing demands. A machine-learning benchmark using the same observable inputs did not show an analogous decline; the paper interprets this contrast as consistent with a human attention constraint. Placebo comparisons with other vendors made a general platform-wide technology trend less likely as the explanation. These results matter within their setting, but the setting matters. Sell-side analysts synthesize market and company information for research reports and forecasts. Corporate FP&A analysts work with internal actuals, budgets, operating assumptions, and planning calendars; they support managers who own operational decisions. Objectives, data access, incentives, and review arrangements differ. The paper cannot establish that AI worsens corporate forecasts, nor that every analyst’s output changes in the same way. It identifies a risk worth testing: when gathering information gets easier, human capacity to interpret the expanded material may become a constraint. Other explanations for weaker forecasts in a new workflow are plausible. Inputs may be incomplete, definitions may vary across systems, or summaries may obscure the period or population behind a number. Adding a tool without clarifying data ownership or review can make problems harder to see. The study’s comparisons address some alternatives, including a broad vendor trend, but cannot settle transfer to internal planning. McKinsey’s “How finance teams are putting AI to work today” offers context rather than a direct test: its November 2025 article describes a proprietary survey of 102 CFOs and selected finance cases, while noting that many organizations had not scaled AI enterprise-wide and pilots can falter when poorly integrated into core processes. These reports do not prove that extra information causes errors; they show why workflow and integration belong in the explanation. For FP&A, the practical response is to value traceability and synthesis, not output volume. Each material statement should trace to a source, reporting period, metric definition, and assumption. Each scenario should expose its drivers: whether a margin change follows price, volume, mix, input cost, or another stated factor, and how the conclusion changes when that factor moves. A reviewer should distinguish a recorded result from an explanation and proposed action. These checks do not make a forecast certain; they make disagreement inspectable and let decision-makers challenge assumptions before they become a plan. The analyst should measure work that faster drafting can hide: total preparation and review time, corrections, unsupported statements, reconciliation effort, and whether analysis helped answer the business question. A shorter draft is useful only if checking and interpretation do not consume the saved time or exceed the result’s value. More scenarios help when they clarify a consequential choice; they clutter when they multiply alternatives without changing a decision. The conclusion is to pair added coverage with disciplined selection and verification. The bottleneck may move toward human synthesis, but whether it does in a particular team remains a question for its data, process, and measured results.

Sources: Generative AI for Analysts; How finance teams are putting AI to work today

How should an analyst test a task before trusting or resisting it?

Choose one recurring task with clear inputs, a checkable output, and limited consequences if a draft is wrong. For example, use a closed month and an approved variance table to prepare a first-pass summary for review. Do not begin with a live forecast, board submission, or action affecting spending or staffing. The point is to learn whether assistance improves this specific step under your team's rules, not to test confidential or consequential work. Before starting, confirm the tool is approved for the data, name the reviewer, and agree how the output may be used. If policy does not permit the proposed data or use, choose a safe alternative or stop. A tool available on a work computer is not automatically authorized for every finance record.

Set a baseline: record usual preparation and review time, required checks, and what makes the summary useful. Run the assisted version on the same bounded inputs. Keep source tables and periods visible; reconcile each number, definition, and stated cause. A plausible sentence can still assign a movement to the wrong period or imply a cause the table does not establish. Track omissions, unsupported explanations, corrections, reviewer edits, and rework, alongside elapsed time. Count review and correction as part of the workflow: a quicker first draft is not a saving if validation takes longer than writing. Ask the intended reader whether the artifact helps answer the real question, such as which variance needs follow-up. Polished wording or more observations alone do not prove better analysis.

Adoption figures describe organizations, not your task. Gartner Survey Shows Finance AI Adoption Remains Steady in 2025 reports a May–June 2025 survey of 183 CFOs and senior finance leaders: 59% said their finance function used AI. Among organizations that had implemented AI, common use cases included knowledge management, accounts payable automation, and error or anomaly detection. Respondents cited data quality and availability, plus data literacy and technical skills, as obstacles; 91% initially observed low or moderate impact. These leader-reported findings do not establish how often FP&A analysts use a tool or whether a specific workflow improves. They support checking data readiness, reviewer capability, and demonstrated impact rather than assuming general adoption means a task is ready to automate.

How AI Agents Can Help FP&A Better Steer the Business describes a selected telecommunications-company example: agents ingested and formatted data with staff validation, built a first-pass forecast for human review, and were compared with human-led outputs over multiple cycles before wider reliance. This consulting case is neither representative nor a result to expect elsewhere. Its useful lesson is procedural: keep validation visible, compare outputs over repeated cycles when tasks vary, and define decision rights before expansion. One successful summary cannot establish reliability across periods, definitions, or higher-stakes forecasts.

Decide in advance what would make you continue, redesign, or stop. Continue only if the output is traceable, corrections manageable, total effort improves or the decision becomes more useful, and the reviewer accepts the workflow. Redesign when errors cluster around unclear definitions, missing context, or weak source trails; investigate these workflow problems without treating them as proof that every use fails. Stop when data are unapproved, output cannot be checked, correction work outweighs benefit, or the task demands certainty the trial cannot establish. Keep a person accountable for material assumptions and approvals. The result is local evidence about one task and its controls. It does not measure displacement, prove reliability elsewhere, or guarantee saved time will become higher-value analysis.

Sources: Gartner Survey Shows Finance AI Adoption Remains Steady in 2025; How AI agents can help FP&A better steer the business

What should an analyst learn next—and which path fits the goal?

If you want to stay in FP&A, strengthen the abilities that let you frame a finance question, verify an answer, and explain what it means; then learn the approved tools attached to an workflow. If you want an adjacent move, add the data, business-intelligence, planning-system, or process skills that local roles request and demonstrate them in a relevant project. AI exposure alone is not a reason to enroll in a degree. Choose the smallest learning investment that closes a demonstrated gap for a named task or role.

Start with the work outcome, not a technology label. “Prepare traceable monthly variance commentary using an approved assistant” is bounded; “become an AI expert” is not. The first goal draws on durable abilities: state the question clearly, understand the metric and period, check figures against source data, spot unsupported explanations, and communicate uncertainty. These foundations travel across tools. Interfaces change, and knowing one vendor’s menus is narrower than being able to validate output.

O*NET’s “Financial and Investment Analysts” profile is broad and includes planning analysts among reported job titles. It lists applying financial models, analyzing financial data, presenting information, and updating professional knowledge. Its software categories include spreadsheets, query tools, business-intelligence software, ERP, and financial-analysis systems. This is not an FP&A curriculum, but it supports an in-role path: deepen modeling, metric definitions, and explanation, then learn the approved system needed for a task. A short course may address a defined gap; supervised practice shows whether learning transfers.

For a finance-data, BI, or planning-systems move, identify the bridge between your experience and the target work. Depending on actual postings, that could mean querying, dashboard design, data mapping, metric lineage, system configuration, or process documentation. Do not collect all of these as badges by default. Review current vacancies where you can work, note repeated requirements, and compare them with your evidence of skill. Then build one small, non-confidential project: document how a sample metric travels from a source table into a report, or create a public-data dashboard with definitions and checks. A project makes applied judgment visible; it does not guarantee an interview or replace required experience.

A course, certificate, project, degree, and self-study solve different problems. Self-study is flexible and inexpensive, but feedback may be limited. A focused course offers sequence and exercises; check its syllabus, support, and tool availability. A certificate may provide structure or a signal some employers recognize, but value depends on the role and local practice. A project tests application. A degree offers broader study and formal credentialing, but takes more time and money; choose it when prerequisites or a longer-term goal justify it. None alone proves job readiness or causes an employment outcome.

A larger change calls for more evidence before more training. Confirm prerequisites, compare current postings, and estimate training time and cost against your capacity. Set a salary floor and geography first; include health, care responsibilities, schedule, and household tolerance for income risk. If these constraints rule out a full-time program, a part-time project, internal assignment, or adjacent role may be a workable test. An employer’s sanctioned pilot can reduce cost and let you observe the work before a pivot.

For U.S. readers, the Bureau of Labor Statistics’ 2025–2035 projections are context, not a forecast of AI’s effect on FP&A. O*NET’s Financial and Investment Analysts profile reports projected growth of at least 7% and 24,700 openings, reflecting growth and replacement. The occupation is broader than corporate planning, and figures do not say which tasks will change or whether a particular analyst will be hired. Readers elsewhere should check local sources. Weigh demand alongside vacancies, skill requirements, and personal constraints.

A practical order is to test a low-risk learning need in the current role; build a relevant project if an adjacent move requires new capabilities; and consider a degree or larger change when prerequisites and personal economics support it. Revisit the choice if postings, employer workflow, or your experience point elsewhere. The goal is not to accumulate credentials against a vague threat. Make the next capability observable, useful in context, and affordable in time and money.

Sources: Financial and Investment Analysts; Occupational Projections and Characteristics

What is the next move this month?

This month, map three FP&A tasks: what you prepare, reconcile, and explain to decision-makers. Choose one repeatable, low-consequence task using approved data and a permitted tool. Before trying it, agree with your reviewer on the data boundary, baseline, and what would count as useful. Compare the assisted version with normal work on total time including review, corrections, traceability to source figures, and usefulness to the decision-maker. Check every number and explanation before sharing. Continue only if the result is verifiable, policy-compliant, and useful after review; if corrections erase the time saved or sources cannot be traced, stop or redesign the trial. The International Labour Organization's “Generative AI and jobs: A 2025 update” estimates occupational exposure from task-level data. That measure does not forecast your employer's adoption or your individual job outcome. Next, compare deepening your current FP&A role, moving toward adjacent finance data, BI, or planning systems work, and a larger change. Set your salary floor, location, learning time, health, and family constraints first. Check local job requirements before paying for training. If a trial is not permitted or the role no longer fits, ask your manager what work is changing and compare options with postings. The free task checker organizes change-pressure signals, not a validated displacement probability. The paid roadmap compares scenarios but guarantees neither employment nor income.

Sources: Generative AI and jobs: A 2025 update

Questions readers ask