PreviewYou're previewing CFO AI Ledger. Sign in to use it.Sign in
← All lessons
L2 · Proving value: business cases, ROI and value tracking9 min read

NPV, IRR, payback and sensitivity for AI projects

The classic tools work for AI, but they need care because AI projects have uneven cash flows, uncertain adoption and fast-changing costs.

NPV first

Net present value discounts each year's net cash flow at your hurdle rate. It is the best single measure because it handles timing and size. A positive NPV at your hurdle rate means the case creates value on its assumptions.

IRR, carefully

The internal rate of return is the rate at which NPV is zero. It is easy to communicate but can mislead:

  • If cash flows never change sign (no upfront investment, or never positive), there is no IRR.
  • If they change sign more than once (for example a big re-platforming cost in year 3), there can be several IRRs. Use NPV.
  • A very high IRR on a tiny investment says little about value. Look at NPV too.

Payback

Payback (in months) answers "when do we get our money back?" It ignores what happens after, but it is a good guard against long, speculative cases. Many finance AI pilots should pay back within the horizon you would give any process improvement.

Sensitivity: the tornado

Move each key driver by, say, ±20% and see how NPV changes. For AI cases the usual suspects are volume, time saved per item, adoption and run cost. If one driver can flip the NPV negative, that is the assumption to test first in the pilot.

Break-even adoption

Ask: at what share of planned adoption does NPV reach zero? If the answer is 30%, the case is robust; if it is 120%, it cannot work as scoped.

Risk haircut

Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. A haircut on benefits is a simple, honest way to reflect delivery risk.

What this means for the CFO

Insist on NPV, payback and a tornado for every AI case, and on break-even adoption for anything that depends on behavior change. Treat IRR as a communication aid, not a decision rule, especially when cash flows change sign more than once.

Ask the sponsor to name the single driver that most moves NPV and to make the pilot test that driver.

Agree a standard template across the company, so every AI case uses the same discount rate, horizon and risk haircut by risk class. Comparable cases make portfolio decisions faster and stop sponsors from shopping for flattering assumptions.

Real example

Gartner (June 25, 2025) predicted that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. Gartner also warned of "agent washing" and estimated that only about 130 of the thousands of agentic AI vendors are real. Source: Gartner — Over 40% of agentic AI projects will be canceled by end of 2027 (Jun 25, 2025) ↗

Use it

Quick check · 3 questions

  1. 1.When can an IRR not be calculated?

  2. 2.What does a tornado chart show?

  3. 3.Break-even adoption is 130% of plan. What does that mean?

0 of 3 answered

Sources

Last reviewed 2026-10-01.