AI Agents in FP&A: Where Judgment Still Matters

Hypothetical: An agent pulls actuals, compares them to plan, and drafts variance commentary. It ranks the largest variances first, but a large timing item may be expected while a smaller recurring-cost change affects the forecast. The workflow is fast; deciding what matters still requires context.

What agents change—and what they don’t

Finance teams already automate fixed steps with formulas, templates, and BI. Agents can connect retrieval, calculations, and draft commentary across a less rigid workflow. That increases throughput; it does not decide which exceptions matter to the forecast or which assumptions a finance owner should approve.

Pilot a bounded workflow

Start with a recurring internal task whose output is easy to verify, such as drafting first-pass budget-versus-actual notes for cost-center owners. Specify which sources the agent may read, how it should rank exceptions, and which actions always require approval. Keep the output reversible, show the source values and assumptions, and leave prioritization with the finance owner. Do not begin with board commentary or external guidance.

RACI before the agent runs

Name the roles before tooling expands:

Decision	Responsible	Accountable	Consulted	Informed
Which sources the agent may read	FP&A systems owner	Finance controller	IT / data steward	Analysts
Exception ranking rules	FP&A lead	FP&A lead	Controllership	Cost-center owners
Approve narrative for leadership	Finance owner	Finance owner	Peer reviewer	Stakeholders
Change ranking or writeback rules	FP&A + systems	Controller	Risk / IT	Team

If the table is empty, the pilot is a demo with a longer memory.

Approval gates that stay human

Require human approval for:

1. Anything that leaves the team (leadership packs, external guidance).
2. Any writeback into a system of record.
3. Any change to ranking rules, materiality thresholds, or source allow-lists.

Allow agent-only drafts for internal first-pass notes when the source values and assumptions are visible and the draft can be discarded without side effects.

Reversible outputs

A useful agent output is a draft with citations to the numbers it used—not a committed journal, forecast lock, or email to a controller. The variance workbench shows a related pattern without a model: review and escalate stay in browser memory, and nothing writes back. See the walkthrough for what to inspect in thirty minutes.

The effect on junior-analyst development is a separate problem; see How AI Changes the FP&A Learning Path.

---

For a guided path through related writing, see Start Here.