The CFO's Guide to AI and Agentic AI

The CFO decides whether AI is a technology expense or a new form of capacity, and that decision shapes everything that follows: how projects are funded, what counts as a return, and what controls the company demands before an agent touches a ledger. This guide sets out the economics, the budgeting model, and the finance controls that fit agentic AI.

Why is agentic AI a finance question?

An AI agent that reads a supplier invoice, matches it to a purchase order and receipt, posts it, and routes exceptions is doing the work of an accounts payable clerk. The economics of that work change in three ways.

1. Cost becomes variable. Each completed task consumes model inference and a little compute. Cost scales with volume, not headcount.
2. Marginal cost is low and falling. Once the agent is built and proven, the next thousand invoices cost a small fraction of the first thousand.
3. Capacity becomes elastic. Month-end peaks no longer require temporary staff or overtime; the agent handles the surge.

This is why the useful unit of measure is cost per completed task, compared with the human or outsourced baseline. FISTA's Digital FTE economics whitepaper works through the math in detail.

How should a CFO structure the AI budget?

Two lines, reviewed on different cadences.

Budget line	What it contains	Behaves like	Review cadence
Build	Discovery, process design, engineering, integration, evaluation sets, security review, initial human-review staffing	A project with a defined scope and completion	Milestone-based, tied to production go-live
Run	Model inference, hosting, observability, maintenance, ongoing evaluation, residual human review, vendor fees	A variable operating cost that scales with tasks	Monthly, against tasks completed and cost per task

Two practical rules follow. First, do not approve build spend without a run-cost estimate per task and a named production date. Second, expect run cost to become the larger line within a year or two as agents take on more volume, and manage it accordingly: model routing (cheaper models for easy tasks), caching, and periodic review of which agents justify their run cost.

The AI total cost of ownership guide lists the cost drivers most budgets miss, including evaluation, monitoring, and the human review that persists after launch.

What counts as a return?

Only production numbers count. A pilot that "would save 3,000 hours" has produced a forecast, not a return. The CFO's discipline is to require a baseline, a target, and an actual for each committed outcome:

• Cost per task: measured before and after, on the same process, at comparable volume.
• Cycle time: measured from trigger to completion, because faster quotes, closes, and settlements have revenue and working-capital value.
• Quality: error rates, rework, exception rates, and downstream corrections.
• Capacity: hours released, counted only when they are reinvested in measurable work or actually removed from a budget.

The how to calculate AI ROI guide gives a worked structure for these comparisons.

What controls should finance insist on?

Agents that act inside finance systems need the controls you already apply to people with the same access, translated into software:

• Identity: each agent has its own scoped credentials, never a shared service account, so activity is attributable.
• Approval thresholds: the agent prepares; a person approves above defined amounts or risk levels. The threshold is a finance policy, not an engineering default.
• Segregation of duties: the agent that creates a vendor cannot also approve payments to it.
• Audit trail: every action logged with inputs, reasoning summary, and outcome, retained for the audit period.
• Reconciliation: agent activity reconciled to source systems on a schedule, exactly as you would reconcile a bank feed.
• Access review: agent permissions reviewed periodically and revoked when the process changes.

External auditors increasingly ask about automated decision-making in financial processes. Designing these controls in from the start is far cheaper than retrofitting them under audit pressure. FISTA's guidance on tool permissions for AI agents explains how the controls are implemented technically.

What should the CFO ask before approving an AI project?

• What is the baseline cost per task today, and who measured it?
• What run cost per task is projected at launch, and at three times the volume?
• What is the production date, and what happens to the budget if it slips?
• Which actions will the agent take without a person, and what are the approval thresholds?
• How will we know the agent is wrong, and how quickly?
• What is the exit plan if the model or vendor changes price or is discontinued?

These questions also filter out AI theater. Projects that cannot answer them are not ready for funding, however impressive the demo. The cost of AI that does not ship explains why pilot portfolios are the most expensive way to learn.

How does the CFO's role change over time?

In the first year, the CFO is mostly a gatekeeper: funding a small number of committed outcomes and enforcing the evidence standard. In the second year, the role becomes portfolio management: which agents earn more autonomy, which are retired, and how run cost is trending. By the third year, finance is usually operating its own agents for close, reconciliation, collections, and reporting, and the CFO is a producer of AI evidence, not just a consumer of it.

Financial reporting, tax, and audit obligations vary by jurisdiction and framework; this guide is general guidance, not accounting or legal advice.

How can FISTA Solutions help a CFO?

FISTA Solutions builds AI agents for finance operations with approval gates, scoped identities, audit trails, and evaluation evidence designed in, and works with CFOs through its AI enablement practice to set up cost-per-task measurement and the reporting that makes returns visible. Since 2017, FISTA has delivered 150+ projects for 50+ companies across 12+ countries; clients report efficiency gains of up to 47% on automated processes.

To pressure-test the economics of an agent before you fund it, talk to FISTA on WhatsApp, or read the AI agent unit economics whitepaper first.