AI in finance: the brake is governance, not technology

AI is already in finance, but the return isn't showing up. The brake isn't technical: roles, decisions and governance aren't ready. How to prepare them.

October 22, 202610 min read

Jose Luis Romero · Process Mining Expert

The CFO wants AI. The structure and governance aren’t ready.

AI has already entered the finance function, but the return isn’t showing up. The problem isn’t the technology. It’s that nobody redesigned the roles, decision rights or control mechanisms to work with it. And without that, scaling AI agents means scaling risk.

The CFO wraps up the AI strategy presentation to the audit committee. Pilots in accounts payable, an assistant for forecasting, agents that flag anomalies before the close. All good until a board member asks: if the agent books a wrong provision and it reaches the financial statements, who is accountable?

Silence. Not the vendor. Not the IT team. Maybe the accountant who approved it without reviewing it.

That pause sums up the state of AI in finance. The technology arrived before the structure to govern it. Deloitte published a broad analysis of how AI is redefining the finance function, and its own figures, cross-checked with Gartner and KPMG, point to the same place: the bottleneck isn’t technical, it’s organizational. This article gives you the arguments to close that gap before an auditor closes it for you.

Figure	What it means
21%	Active AI users in finance who report clear, measurable value
~75% vs. 21%	Organizations planning agentic AI within two years versus those with mature governance for it
84%	Organizations that haven’t redesigned jobs around AI
7%	Share of the AI budget going to redesigning work and people

The real gap is between deploying and getting value

The usual story says finance wants AI but hasn’t managed to implement it. The data says something else.

87% of CFOs consider AI extremely or very important to their operations in 2026, according to Deloitte’s CFO Signals survey. The firm’s Finance Trends 2026 study, covering more than 1,300 finance leaders, shows 63% already have it deployed and in active use. But only 21% report measurable return, and just 14% have fully integrated agents.

Gartner sees the same plateau: adoption in finance jumped from 37% in 2023 to 58% in 2024 and practically stalled in 2025.

The technology is already installed. What wasn’t installed is an organization able to make use of it.

The problem is no longer switching AI on. It’s that nobody redesigned the house where it was switched on.

AI doesn’t fit in the current org chart

Deloitte puts it bluntly: AI doesn’t slot into the existing operating model. It forces it to change across capabilities, service delivery, organizational design, ways of working, technology and, above all, decision rights.

The underlying shift is a redistribution of effort. Today, between 45% and 55% of finance work goes to transactional processing. With AI, that share could fall to 5–15%, while analysis and strategic direction would rise to 35–45%.

That shift doesn’t happen on its own. If roles, handoffs between teams and responsibilities aren’t redesigned before automating, the freed-up hours dissolve into reviews, emails and meetings.

There’s a more delicate point. In an agent-based model, decision authority becomes variable: it moves from the system to the person depending on model confidence, risk appetite or how ambiguous the case is. No traditional finance structure is designed for that.

Current structure	AI-ready structure
Approvals by amount and hierarchy	Thresholds based on risk and model confidence
Sample-based controls at close	Continuous monitoring with defined escalation
Roles built around transactional tasks	Roles built around exception oversight and analysis
Implicit accountability	An explicit owner for every automated output

Nobody knows who answers for the model’s decision

When an AI output feeds into provisions, disclosures or accounting flows, it stops being an experiment. It becomes a risk management has to manage.

Before auditors and regulators, finance has to be able to answer three questions about every use of AI:

What exactly did the tool produce?
Who reviewed that output?
Who is accountable for it?

Most organizations can’t answer them today. According to Deloitte’s State of AI in the Enterprise report, almost three quarters plan to deploy agentic AI in the next two years, but only 21% have mature governance frameworks for those systems.

The gap widens every quarter, because pilots in reporting, forecasting and controls move ahead without policies, accountability structures or oversight mechanisms.

Data governance is part of the same problem. AI doesn’t fix inconsistent information: it reproduces it at scale. But the fix isn’t only technical. It’s defining who owns each data set, who certifies its quality and who decides whether it’s fit to feed an automated decision.

Governance: 20% policy, 80% behavior

Deloitte proposes assessing AI controls against five criteria: transparency, fairness, privacy and security, reliability and accountability. It’s a good framework. But the same analysis warns that effective governance is, proportionally, 20% policy and 80% behavior.

A policy without habits is a PDF on the intranet. A team with good habits but no policy won’t survive an audit.

In practice, governing financial AI takes four elements:

A complete inventory. Every model, agent or assistant that touches a finance process, registered with its purpose and scope.
An explicit owner and reviewer. One person accountable for the output and another who validates it, by name and title.
Escalation criteria. Which cases the system resolves on its own and which must go to a person.
Active internal audit. Not as an after-the-fact reviewer, but as a function that tests controls in practice and reports findings to the audit committee on an ongoing basis.

The core recommendation is to build governance into the investment thesis, not bolt it on as a control layer at the end. Organizations that do can show auditors, regulators and investors that their transformation rests on a reliable foundation.

An agent without an owner isn’t automation. It’s unsigned risk.

One project, two owners: CFO and CIO

The governance structure runs into an ambiguity few organizations resolve. Gartner’s 2024 CIO survey showed that 71% of CFOs believe they’re in charge of enterprise technology, and 77% of CIOs think the same.

AI deepens the overlap. According to KPMG’s 2025 CFO-CIO collaboration survey, 51% of CFOs say AI expanded their responsibility for data and analytics, versus 31% of CIOs. And while 41% of CIOs see AI as having integrated planning and budgeting, only 20% of CFOs see it that way.

Both believe they lead. Both see different progress. That ownerless space is where projects stall.

The answer isn’t deciding who’s in charge, but splitting responsibilities in writing:

Finance decides	IT guarantees
Which processes to prioritize and which KPIs measure them	Architecture, integration and security
Risk thresholds and each agent’s level of autonomy	Traceability, versioning and model monitoring
Who reviews and approves each output	Access control and operational continuity
Criteria for escalating to people	The technical record of the AI inventory

The jobs didn’t change; the tasks did

84% of organizations haven’t redesigned jobs around AI. The result is predictable: people with job descriptions built for producing numbers, working alongside systems that now produce them.

Deloitte describes the profile replacing the traditional specialist: a cross-functional professional who works with AI, brings business context to model outputs and applies judgment when exceptions arise. Core finance skills don’t change. What they’re applied to does.

The budget, however, isn’t keeping up. According to Deloitte’s Tech Trends 2026 report, only 7% of AI investment goes to redesigning work and people; 93% goes to data and technology. The consequence: reported AI use fell 15% even with the tools available.

The cause is trust. An analyst who doesn’t trust the model’s number recalculates it by hand, and the process gains a step. When trust is high, finance professionals are 2.7 times more likely to use generative AI daily and save 2.3 times more hours per week, according to Deloitte’s TrustID Workforce Index.

Trust can’t be decreed. It’s built with structure: clear roles, known review criteria and the certainty that someone answers when the model gets it wrong.

A roadmap in five moves

Inventory before expansion. Register every AI use that already exists in finance, including the ones the team adopted on its own.
Decision rights. For each process, define what the system decides, what a person decides and at what threshold accountability shifts.
A written CFO-CIO agreement. Separate what finance decides from what IT guarantees, and take it to the executive committee.
Role redesign. Shift job descriptions toward exception oversight, analysis and validation, and set aside real budget for training.
KPIs across four dimensions. Measure financial, operational, strategic and employee value. If all you can show is hours saved, the operating model hasn’t absorbed the change yet.

Reasonable objections

“Let’s test AI first; we’ll set up governance when we scale.” That’s the path that led to today’s gap. Pilots without governance become de facto processes, with outputs already feeding decisions and nobody formally accountable. Governing later costs more, because it means dismantling habits that are already in place.

“The figures come from large North American companies. Do they apply to Latin America?” The CFO Signals survey covered 200 CFOs at companies in the United States, Canada and Mexico with revenue above $1 billion. The proportions may vary. But in organizations with smaller teams and overlapping functions, ambiguity over roles and responsibilities tends to be greater, not smaller.

“Another AI committee is more bureaucracy.” Only if it’s designed as an approval gate. Designed well, governance speeds things up: it defines in advance what can be automated without consultation and what needs review, and removes the case-by-case debate that’s holding projects back today.

AI doesn’t replace control: it demands it

The finance function’s mandate doesn’t change: close the books, manage risk, allocate capital and guide decisions. What changes is who executes each step, and that forces a redefinition of who decides, who reviews and who answers.

Organizations that treat AI as a software purchase will keep piling up pilots and explaining to the board why the return isn’t coming. Those that treat it as a redesign of structure and governance will be the ones able to scale it without fearing the next audit.

Before you automate decisions, decide who answers for them.

Your next step this month: bring the CFO, the CIO and the head of internal audit together for a session with a single goal: build the inventory of AI uses that already exist in finance and assign each one an owner, a reviewer and an escalation criterion. That document will be the foundation of your governance model and the answer to the question the audit committee will ask sooner or later.