Most agentic AI pitches land in the CFO's inbox the same way: a slick demo, a vague ROI claim, and a request for budget. The questions that actually decide whether a project gets funded almost never get asked on stage.
After dozens of finance-led conversations this year, the same four come up every time.
1. Where does the agent sit in our operating model?
A finance team is a control environment. Before anyone cares what the agent does, they care where it fits — who owns it, who reviews its output, and how it shows up in the org chart. "It's just a tool" is not an answer; agents take actions, and actions need an accountable owner.
The cleanest model we've seen: treat the agent like a junior analyst with a named manager, a defined remit, and a weekly review. Boring on purpose.
2. What is the unit of value?
Time saved is not an investment thesis. CFOs want one of three things:
- Cycle time — close in 4 days instead of 7
- Capacity — same team, more entities
- Quality — fewer post-close adjustments, fewer audit findings
Pick one. Measure it before and after. If you can't measure it, you don't have a project — you have a pilot.
3. How do we know it's still working?
Agents drift. Models change. Vendors update prompts. The control question is not "did it work in the demo?" but "how will I know in six months if it's stopped working?"
This is where evaluation harnesses earn their keep — a small, repeatable set of test cases the agent has to pass before each release, owned by finance, not engineering.
4. What's the cost of being wrong?
Every agent decision sits somewhere on a spectrum from "annotates a journal" to "posts a journal." The further right you push, the more the failure mode matters. Most successful finance deployments we've seen stay firmly on the left for the first two quarters — agents draft, humans approve — and only move right once the eval scores are boring.
None of this is glamorous. But these are the questions that turn an interesting demo into a funded programme. If your vendor can't answer them on a whiteboard, the demo isn't the problem.