Calculating AI tool ROI for small teams (without lying to yourself)
A spreadsheet-grade method for figuring out whether your AI subscriptions are actually paying off — or just feel like they are.
Engineering managers love to claim AI tools save 'so much time'. Most cannot show the math. Here is a defensible framework you can run in an afternoon and update quarterly.
The honest baseline
Before you can measure savings you need a baseline. Pick four metrics that already exist: median time-to-PR, PR cycle time, CI first-pass rate, and incidents per month. Pull the last three months of numbers.
Resist the urge to invent new metrics. Existing telemetry is more trustworthy than anything you build for the purpose of justifying the tool.
What to count as savings
- Reduced time-to-PR for similar tickets.
- Reduced reviewer comments per PR (a proxy for first-draft quality).
- Reduced CI failures per PR.
- Reduced repeated questions in team chat (the rubber-duck effect).
What not to count
Do not count 'engineer happiness' in the financial model. It is real and it matters; it just does not survive a CFO conversation.
Do not count incidents avoided. You cannot prove a counterfactual, and any number you pick will be wrong.
A worked example
Eight engineers, fully-loaded cost $180/hour. AI subscriptions at $40/seat/month: $3,840/year. Measured time-to-PR drop of 18% over six months on tickets of similar size. Assume engineers spend 25% of time on PR-shaped work. That is roughly 90 hours saved per engineer per year, or $16,200 of recovered time per engineer.
Even discounting heavily for measurement noise, the tools clear the bar by an order of magnitude. The point is not the number — it is that you ran the math.
Re-run it every quarter
Tool quality changes. Team composition changes. Pricing changes. A subscription that paid for itself in Q1 may be a wash by Q4. Make this a recurring exercise, not a one-time defense.