A Finance Team’s First 90 Days With AI

The Problem Is No Longer Finding an AI Use Case

By late 2025, finance leaders did not need another list of things generative AI might someday do. They needed an operating method for deciding what should move from experiment to repeatable workflow.

McKinsey reported in November 2025 that 44% of CFO respondents were using generative AI for more than five use cases, up from 7% in the prior-year survey. Sixty-five percent said their organizations would increase generative-AI investment in 2025.

Days 1–30: Find Work, Not Technology

Start with a workflow inventory. Ask analysts and managers where they repeatedly copy information, reconcile versions, search for context, draft recurring commentary or answer the same questions.

Score each workflow on frequency, time consumed, financial materiality, data sensitivity, error cost and ease of human review. The best first use cases are frequent enough to matter, bounded enough to test and easy enough to verify.

Days 31–60: Build the Evidence Path

For each selected workflow, define the source data and what counts as an acceptable answer. If AI drafts variance commentary, the reviewer should be able to trace every material statement to financial data. If AI prepares a reconciliation, define tolerances. If AI summarizes a contract, identify clauses requiring human confirmation.

Create a small evaluation set from real historical work. Run the workflow repeatedly and record where it fails.

Days 61–90: Move From Prompt to Process

A useful finance AI workflow needs more than a good prompt. It needs an owner, approved data access, versioned instructions, review rules, an escalation path and a measurable outcome.

1. 01Level 1 — Draft — AI produces content; a person does the work.
2. 02Level 2 — Recommend — AI proposes an action; a person approves.
3. 03Level 3 — Execute with approval — AI prepares the action; a person releases it.
4. 04Level 4 — Execute by policy — AI acts inside predefined thresholds and escalates exceptions.

Not every workflow should reach Level 4.

Measure the Right Thing

Do not measure success by prompts sent or licenses activated. Measure cycle time, manual touches, rework, exception rate, reviewer time, accuracy against an agreed standard and the percentage of outputs traceable to source evidence.

The Human Review Trap

Human-in-the-loop is not automatically safe. If reviewers rubber-stamp hundreds of AI outputs, the control exists on paper but not in practice. Review should be risk-based.

What Every Finance Team Should Document

Who owns the workflow. Which data sources are approved. Which model or tool is used. What the AI may and may not do. How outputs are checked. Where actions are logged. What triggers escalation. How performance is reviewed over time.

The 90-Day Outcome

The goal is not an “AI strategy deck.” It is two or three workflows that the team can run repeatedly, measure, audit and improve. Once that foundation exists, scaling becomes an operating problem rather than a series of experiments.

Verified Facts Used

• McKinsey reported in November 2025 that 44% of surveyed CFOs used generative AI for more than five use cases, up from 7% in its prior-year survey.
• 65% said their organizations would increase generative-AI investment in 2025.
• The 90-day framework in this article is editorial guidance, not benchmark research.

Factual product claims and statistics are limited to information available by the publication date. Vendor-reported metrics are identified as vendor/company-reported rather than independent benchmarks.