AI Adoption in Finance: How CFOs Move from Pilots to Controlled Value

A finance function pilots an AI assistant for management reporting.

The tool drafts variance commentary, summarises business-unit submissions and identifies unusual movements. Early users estimate that the monthly reporting process is faster. The CFO sees a route to greater insight with less manual effort.

The first difficult reporting cycle exposes the gap.

The assistant explains a variance using an outdated classification. One business unit has changed its operating model, but the relevant source data and narrative are inconsistent. Reviewers correct the output manually without recording why. The finance team cannot distinguish time saved in drafting from time added through verification, correction and reconciliation. Because the pilot sits outside the reporting workflow, the final commentary is copied between systems and reviewed through the old process anyway.

The function has demonstrated a useful capability. It has not yet established controlled AI adoption in finance.

KOR’s central argument is:

AI adoption in finance succeeds when useful automation, data quality, auditability, human judgement and measurable value operate as one finance process.

The distinction matters because finance is moving beyond the period when low adoption could be explained by simple caution.

Gartner reported that 59% of finance leaders surveyed in 2025 were using AI in their finance function, broadly unchanged from 58% in 2024. Knowledge management, accounts payable automation and error or anomaly detection were the most commonly adopted use cases among respondents using AI. Yet Gartner’s 2026 roadmap research reported that 84% of finance organisations had implemented or planned AI while only 7% reported high or very high impact.

KPMG’s 2026 Global AI in Finance report reaches a similar conclusion: adoption is broad, but stronger performance is concentrated among organisations that combine governance, measurement, audit evidence and workforce capability.

The adoption problem is therefore no longer only access.

It is the ability to move from planning, pilots and embedded features into dependable production use with evidence strong enough for finance decisions.

Framework boundary: This article applies KOR’s proposed Enterprise AI Adoption Model to corporate finance. It is not accounting, audit, regulatory or investment advice. Requirements depend on the system, process, organisation and jurisdiction.

AI adoption in finance is the sustained use of approved AI within finance workflows, supported by reliable data, clear control ownership, auditability, competent human review, redesigned processes and evidence of operational or financial value.

Finance is sometimes described as culturally slow because it places a high value on accuracy, evidence, segregation of duties and repeatability.

That description is incomplete.

These requirements can slow careless deployment. They can also create the discipline required for adoption at scale.

Finance AI must operate inside processes where:

• figures reconcile;
• assumptions are understood;
• data lineage matters;
• material changes require explanation;
• reviewers are accountable;
• evidence may be examined later;
• outputs can affect capital, reporting, controls or business decisions.

The correct objective is not to make finance behave like a technology laboratory.

It is to create controlled experimentation that can mature into auditable, useful work.

Knowledge management and finance support

Use cases include searching policies, accounting guidance, prior analysis, process instructions and internal knowledge.

This can be a lower-risk starting point where the assistant retrieves authoritative sources and users verify the answer. Adoption weakens where source libraries are incomplete, outdated or poorly permissioned.

Accounts payable and transaction processing

AI can support invoice capture, coding, matching, exception detection, duplicate identification and query handling.

The value depends on process and master-data quality. Automating poor purchase-order discipline or supplier data does not remove the underlying issue. It may make errors faster and less visible.

Close, reconciliation and control

AI can help identify unusual balances, propose reconciliations, organise support and summarise exceptions.

The adoption boundary should be explicit: what may be prepared automatically, what requires review, what evidence must be retained and who signs off.

Error and anomaly detection

Models can identify patterns that rules or manual sampling miss.

The output is not a conclusion. Finance users need to understand the alert threshold, false-positive and false-negative trade-offs, data coverage and escalation route.

Forecasting and planning

AI can support driver analysis, scenario creation, forecast updates and narrative explanation.

Forecasting is especially vulnerable to false precision. A more complex model is not automatically a better decision tool. Finance should compare performance with credible baselines and examine how the model behaves under structural change.

Management reporting and business partnering

Generative AI can draft commentary, summarise performance and help finance partners explore questions.

The opportunity is not merely faster writing. It is a shift from report production towards decision support. That shift requires role redesign and a clear standard for evidence, interpretation and challenge.

Code and model assistance

Finance professionals increasingly use AI to create spreadsheet logic, SQL, scripts and low-code automation.

This expands local innovation but also introduces control questions: testing, documentation, access, version control, model ownership and dependency on citizen-developed solutions.

Finance users should know:

• which tasks AI may support;
• which outputs may enter books, reports or decisions;
• who owns the final judgement;
• what review is mandatory;
• when independent checking is required;
• how changes are documented;
• what records must be retained;
• when use is prohibited or suspended.

A generic statement that finance remains accountable is not enough. Accountability must be matched by information, authority and time to review.

Finance AI capability includes:

• understanding the data and process;
• interpreting uncertainty;
• testing calculations and code;
• distinguishing correlation from explanation;
• verifying source material;
• recognising structural breaks and unusual conditions;
• documenting assumptions;
• challenging polished but unsupported narrative;
• knowing when to escalate.

Gartner identified data literacy and technical skills as major finance adoption barriers in both 2024 and 2025. The answer is not to turn every accountant into a data scientist. It is to define the capability required for each role and provide access to specialist support.

Finance employees will use consumer tools where approved systems are unavailable or too restrictive.

The approved route must address:

• financial and commercially sensitive data;
• access by entity, business unit and role;
• source-system integration;
• retention of prompts and outputs;
• supplier use of customer data;
• code and model storage;
• evidence and audit trail;
• local tools and spreadsheet integration.

A policy that prohibits practical work without offering a credible alternative encourages hidden use.

Finance culture can produce both under-trust and over-trust.

Under-trust leads to duplicate manual work and prevents value. Over-trust turns a generated explanation, prediction or anomaly score into apparent fact.

Calibrated trust asks:

• What evidence supports the output?
• Which source data was used?
• How current is it?
• What assumptions are embedded?
• How does performance compare with the existing method?
• What errors are material?
• Who must review?
• How are corrections captured?

The review standard should follow consequence. Drafting an internal email does not require the same assurance as supporting a forecast, control conclusion or external disclosure.

CFO sponsorship is important because finance employees take signals from how leaders respond to error, uncertainty and workload.

Leadership should:

• define an ambition level for finance AI;
• identify where finance will be an end-user, co-developer or pioneer;
• fund data and process foundations;
• allocate time for testing and redesign;
• require evidence of value;
• reward disclosure of issues;
• avoid presenting headcount reduction as the only benefit;
• model appropriate use in leadership work.

Gartner’s 2025 research on finance leadership roles distinguishes finance as an end-user advocate, co-developer or pioneer depending on the solution and available central capability. The cultural point is that the function must know what responsibility it is accepting.

Finance professionals need permission to challenge both the system and the business narrative around it.

They should be able to report:

• a recommendation that does not reconcile;
• a model that performs poorly under current conditions;
• pressure to accept an output because it supports the desired answer;
• a mandated process that creates duplicate work;
• a local AI solution operating without appropriate review;
• a claimed saving that is not visible in the end-to-end process.

A strong control culture should support challenge. A punitive implementation culture can suppress it.

Finance adoption is distorted when teams are asked to add AI review while targets and close timetables remain unchanged.

Poor measures include:

• number of prompts;
• licences activated;
• pilot count;
• hours estimated by users;
• volume of reports drafted;
• nominal automation rate.

Better measures include:

• cycle time across the full process;
• exception and correction rates;
• reconciliation quality;
• control effectiveness;
• forecast performance against an appropriate baseline;
• decision timeliness and usefulness;
• sustained approved use;
• capacity released and redeployed;
• total cost of operation;
• finance-validated benefit.

AI should change the finance workflow rather than become another step around it.

For management reporting, redesign may involve:

• automatic retrieval of approved data;
• structured first-draft commentary;
• explicit evidence links;
• exception-focused review;
• recorded corrections;
• reduced copying and reformatting;
• more time for business challenge.

For accounts payable, redesign may involve different exception routes, supplier communication and control sampling. For planning, it may change how scenarios are generated, reviewed and communicated.

The released capacity should have an intended use. Otherwise time savings disappear into workload without creating visible value.

Finance adoption is verified when:

• approved use persists beyond the pilot team;
• outputs are reviewed appropriately under pressure;
• data and controls remain reliable;
• corrections are visible and acted on;
• roles and timetables have changed;
• duplicate manual processes have reduced;
• outcomes improve against baseline;
• value remains credible after total costs;
• the process survives model, supplier and personnel changes.

Provisional

• leadership optimism;
• training attendance;
• pilot demonstrations;
• estimated hours saved;
• user satisfaction;
• supplier benefit claims.

Substantiated

• approved use cases;
• process and control design;
• data lineage;
• documented review rules;
• test results;
• supplier and model information;
• role guidance;
• implementation and value plan.

Verified

• production usage;
• retained audit evidence;
• observed review and escalation;
• performance against baseline;
• reduction in errors or cycle time;
• sustained process change;
• finance-validated benefits;
• continuing control after material change.

A finance AI business case should not be treated as realised value.

1. Inventory existing and embedded AI. Include ERP, EPM, AP, analytics, reporting, productivity tools and local code.
2. Define the finance problem. Start with a process or decision, not a technology category.
3. Set the consequence and control boundary. Determine what the output may influence and what evidence is required.
4. Assess data and process readiness. Identify master-data, integration, ownership and standardisation gaps.
5. Choose the finance role. Decide whether the function is end-user, co-developer or technical pioneer.
6. Establish a baseline. Measure current cost, time, quality, errors, exceptions and decision impact.
7. Run a bounded pilot. Include difficult cases and normal operating pressure.
8. Redesign the workflow. Remove obsolete steps and define exception handling.
9. Verify impact. Separate task savings from end-to-end value.
10. Scale conditionally. Expand only where controls, adoption and outcomes support the decision.
11. Monitor material change. Reassess data, models, suppliers, thresholds and user behaviour.

A group finance team pilots generative AI to draft monthly variance commentary.

The initial pilot shows that first drafts can be produced more quickly.

The adoption review finds:

• source data is retrieved manually from several systems;
• business-unit classifications are inconsistent;
• the assistant does not link statements to supporting evidence;
• reviewers correct outputs but corrections are not logged;
• the old reporting process remains intact;
• user estimates count drafting time but omit checking and copying;
• local teams apply different review standards.

The organisation does not scale immediately.

It redesigns the process:

• data is drawn from approved sources;
• commentary follows a structured template;
• statements link to evidence;
• material explanations require named review;
• corrections are recorded;
• local differences are tested;
• cycle time and quality are measured across the full process;
• the old first-draft step is removed where evidence supports it.

The use case moves from a writing assistant to a controlled reporting workflow.

Automating before standardising

AI cannot resolve ownership, definitions and master-data issues by itself.

Treating a generated narrative as analysis

Clear language can conceal weak reasoning or incomplete evidence.

Counting task time as financial value

A faster task does not create value unless capacity, quality, cost or decisions change.

Adding review without changing workload

This creates duplicate work and encourages superficial checking.

Ignoring embedded AI

Finance platforms may introduce AI through vendor releases before the function has classified or approved the use.

Leaving local code outside governance

AI-generated spreadsheet logic and scripts can become critical finance assets without documentation or ownership.

Treating control culture as resistance

Challenge is valuable. The objective is to make evidence proportionate enough that strong controls do not become permanent duplication.

The relevant question is not whether finance should adopt AI faster.

It is whether a use case has:

• a clear finance problem;
• data and process foundations;
• defined control ownership;
• capable users;
• meaningful human review;
• an auditable workflow;
• evidence of sustained adoption;
• value that survives full cost and risk.

Finance becomes an AI leader not by approving the greatest number of tools, but by turning selected capabilities into dependable, measurable work.

How widely is AI used in finance functions?

Gartner reported that 59% of surveyed finance leaders were using AI in their finance function in 2025. Adoption does not necessarily mean production scale or high impact.

What are common finance AI use cases?

Knowledge management, accounts payable automation, anomaly detection, forecasting, management reporting, reconciliation support and code assistance are common or emerging uses.

Why does finance AI adoption stall?

Common barriers include data quality, technical and data literacy, unclear use cases, fragmented automation, weak workflow redesign, uncertain ownership and difficulty converting pilots into measured value.

What does human oversight mean in finance AI?

It means a competent person has the evidence, authority and time to review the output, challenge assumptions, correct material errors and remain accountable for the finance conclusion.

How should finance measure AI value?

Use end-to-end baselines covering cost, cycle time, quality, exceptions, control outcomes and decision impact. Do not rely only on estimated task hours.