The ROI of Copilot, Cursor or any AI copilot in your company is almost never measured well. It's measured in active licenses and in a "sense of speed" — two metrics that age badly. This week the capital markets gave a hint of why: they've started to put a price on the difference between renting capacity and retaining method. And that difference is exactly what your CFO is going to ask you about in the next budget review.
The market just repriced "bodies-in-seats"
The Indian IT sector —which for two decades has been synonymous with "software development at scale"— has seen more than 46% of its combined market value evaporate since the August 2024 peak (the five largest going from 33.7 to 18.2 trillion rupees). This isn't a technical correction: it's a re-reading of the model. When TCS opens earnings season with its AI business at a material run-rate but with margin compressed to 24.0% and flat growth, the message to the investor is uncomfortable to digest: AI shifts productivity to the customer, not to the margin of whoever bills the hours.
Translated for your leadership team: the value of AI in software development is shifting from headcount and the license toward the method. Whoever only sells bodies (or a tool's subscription) has a business problem. And whoever buys only bodies or only licenses has a symmetric ROI problem, even if they haven't named it yet.
Why "installed licenses" and "Copilot ROI" are the wrong metric
The usual way to justify investing in copilots is to count seats: "we have 80 active Copilot licenses and the team goes faster". It's the comfortable metric, and it's the wrong one, for three reasons any CFO recognizes:
- The license is a recurring cost that accumulates no asset. The day you stop paying for it, you have nothing left. It's rent, not investment.
- "Perceived speed" isn't defensible to the leadership team. Without a baseline or a quality criterion, "we go faster" is an anecdote, not a return. And speed without governance produces debt —bugs, PRs nobody reviewed well— that shows up two quarters later.
- You don't control the tool's price. If the provider raises the rate, changes the consumption model or reprices (and the AI arms race guarantees it will), your "license ROI" gets recalculated against you overnight.
Measuring AI ROI by installed licenses is like measuring a factory's ROI by the number of workers on the payroll. It counts the input, not the asset. The same holds on the output side: if throughput rises but review stops catching anything, what grows is not the asset — it is the liability.
The real ROI is the method your team retains
The return that does hold up to a leadership team —and that doesn't reprice against you— is the method that stays inside your organization. In onext's practice it's made up of three concrete things:
- Spec-Driven Development (SDD): a proven way of working that multiplies velocity without losing control of what gets built. It's not "prompt and to production"; it's specification, verification and traceability.
- Context engineering: collecting how your company really works —its rules, its domain, its criteria— and turning it into the context your agents and copilots use. That context is yours by definition; it doesn't live in anyone's license.
- Human verification at every step (human-on-the-loop): so that speed isn't paid for with quality. Governed production, not a POC in slides.
The difference is measurable. Across the 12 teams onext has transformed, the pattern is ×7 delivery velocity, 0 sprints lost during the transformation and −50% time-to-production (from a market baseline of ~16 weeks to 8 to the first feature in governed production). But the data point that matters for this conversation isn't the speed: it's that, when it's over, the capability stays in the client's team. The method isn't a license you renew; it's an asset nobody can reprice on you anymore.
A calculation framework for CFO + CTO
To take this to the budget review, change the numerator and the denominator of your ROI calculation:
Instead of:
(perceived speed) / (cost of licenses)
Calculate:
(capability retained by your team) / (total cost of ownership, with an exit date)
Concretely, three questions that order the decision:
- What's left for me if I switch the tool off tomorrow? If the answer is "nothing", you're renting. If it's "the method, the context and the criteria my team already masters", you're investing.
- Who controls the price of what I bought? License and consumption: the provider. Retained method: you.
- Do I have a baseline and a quality criterion? Without that there's no ROI, there's a feeling. With it, the conversation with the CFO stops being defensive.
The hidden cost: the "embed" with no exit date
There's an expensive variant of this mistake worth naming, because the market is pushing it hard. Half a dozen giants —from the big AI firms now setting up their own consultancies to the integrators franchising the model— offer to place certified engineers inside your house, permanently. It's convenient: they solve the ramp-up for you. But the ROI question is the same as with the license, one level up: when they leave, what do they take with them?
If the knowledge of how your AI works leaves with them, you've swapped dependence on a license for dependence on a presence. And that presence, like the license, is repriced by them. The honest contrast: one model sells you the permanent embed; the other transfers you the method and leaves with a date. For an asset you'll operate for five or ten years, the second option changes your economics more than any license discount.
Frequently asked questions
How do you calculate the real ROI of Copilot or Cursor in a company?
Not by active licenses or perceived speed, but by the capability your team retains: way of working, encoded business context and quality criteria. The useful formula is retained capability / total cost with an exit date. If nothing is left when you switch the tool off, you were renting, not investing.
Why have Indian IT stocks fallen and what does it have to do with my AI ROI?
Because the market is repricing the "renting bodies" model for building software: AI shifts productivity to the customer, not to the margin of whoever bills the hours (TCS showed it with its margin compressed to 24.0%). The read for you: value is shifting from licenses and headcount toward the retained method.
Is it better to buy copilot licenses or transform the team's method?
They're not mutually exclusive, but only one accumulates an asset. The license is a recurring cost that leaves you nothing when you cancel; the transformed method (SDD + context engineering + human verification) stays in your team and isn't repriced by the provider. The license is the tool; the method is the return.
Conclusion
The market just gave you a decision framework for free: it's putting a price on the difference between renting capacity and retaining method. Your AI ROI should be measured the same way. Copilot or Cursor licenses are the tool —useful— but the return you defend to the leadership team, and that nobody can reprice against you, is what stays inside your team.
If you want to put a number on that in your organization —what capability you retain today and what would be left if you switched the tools off— start with a diagnostic: in a few weeks you have a baseline you can defend to your leadership team and a plan to turn license spend into an asset of your own.
Jordi García is Tech Lead at onext. He works on bringing AI into governed production across development and product teams —with Spec-Driven Development, context engineering and human verification at every step— and authors onext's technical insights on the method, quality and cost of applied AI.
LinkedIn →