How the calculation works
Every number above comes out of the formula below. There is no hidden multiplier and no backend. If you can read it, you can check it.
steady_annual_gain = engineers × loaded_cost × gain% × ai_assistable%
ramp_factor = average of min(1, m / ramp_months) over months 1..12
gross_gain = steady_annual_gain × ramp_factor
rework_cost = gross_gain × rework_drag%
tool_cost = engineers × seat_cost_mo × 12
net_first_year_roi = gross_gain − rework_cost − tool_cost
roi_ratio = net_first_year_roi / tool_cost
breakeven_month = first month where cumulative (monthly gain after rework and ramp
− monthly tool cost) turns positive
Two choices in there are worth calling out, because they are where honest and dishonest calculators part ways.
The ramp factor. A tool bought in January is not at full productivity in January. Adoption climbs over a quarter or two, so year one captures only part of the steady-state gain. We model that ramp as a straight line to full use over your ramp-months input, then average it across the twelve months. A calculator that skips this quietly assumes full productivity from day one, which nobody gets.
The rework term. This is the one vendor math leaves out. Some of the time an assistant appears to save comes back as time spent reviewing, correcting, and re-testing what it produced. It is real work and it is not free. We discount the gain by your rework percentage before calling anything a return. Set the slider to zero and you will see the rosier number the vendors quote; the report shows both side by side.
Where the defaults come from
Every default is a starting point you should replace with your own data. Here is the source for each, so you can decide whether ours is reasonable for your team.
| Input | Default | Why that number |
|---|---|---|
| Loaded cost / engineer | $180,000 | A mid-band fully loaded cost (salary plus roughly 1.25 to 1.4x for benefits, overhead, and taxes) for a US senior engineer. Use your own comp data if you have it. |
| Seat cost / month | $39 | Between GitHub Copilot Business ($19) and the Cursor / Claude Code tier (~$40 to $60+). Metered usage can run higher, so track actual spend rather than the sticker. |
| Productivity gain | 10% | Vendor studies claim 20 to 50%. Independent field data (DX, across 184 companies) clusters at a 5 to 15% median. METR's 2025 randomized study found experienced open-source developers were about 19% slower with AI on their own mature repositories. Ten percent is a cautious middle that assumes the tool helps, modestly. |
| AI-assistable share | 50% | Engineers spend well under half their time hands-on-keyboard writing code; the rest is meetings, design, review, and waiting. The gain only applies to the code-writing slice. |
| Adoption ramp | 6 months | Vendor and Microsoft rollout guides run adoption in phased quarters, not a flip of a switch. |
| Rework drag | 15% | Studies put the share of apparent time savings respent on debugging and review at 15 to 25%, and find AI-authored code churns 1.8 to 2.5x more than human-written code. Fifteen percent is the low end. |
Sources are named rather than hot-linked because vendor and study URLs rot; each is searchable by the description given (METR 2025 developer productivity study; DX Core 4 / DX field research; Microsoft and GitHub Copilot rollout guidance). This tool takes no position on any vendor and earns nothing from your choice.
Questions people ask
What ROI should I expect from GitHub Copilot or a similar tool?
Honestly, a range that includes zero. The credible independent numbers land far below vendor claims, and a minority of measured cases go negative, especially for senior engineers on code they already know well. Run this calculator at 10% and again at 30% and treat the span between them as your real forecast, not a point estimate. Then instrument the actual gain rather than trusting any of these figures, including ours. Our guide on proving ROI to a skeptical VP walks through that.
Why did AI make my team slower?
It is a documented outcome, not a fluke. METR's 2025 randomized controlled trial found experienced developers took about 19% longer with AI assistance on their own mature codebases, even while believing they were faster. The usual mechanisms are review and correction overhead (our rework term), context-switching between writing and vetting, and confident-but-wrong output in code where the human already had deep context. On unfamiliar or boilerplate-heavy work the tool tends to help more.
How do I measure the real productivity gain?
Not with a vendor dashboard or a lines-of-code count. Measure adoption and outcomes without turning it into surveillance: see the playbook on measuring AI adoption without a leaderboard.
What is the cost per developer, really?
Often more than the seat price, because metered usage is invisible until you instrument it. The playbook on tracking AI spend across providers covers how to reconcile actual spend back to the invoice.
This calculator answers "does it pay?". The prior question, whether a machine should be doing this particular piece of work at all, has its own tool: what should AI do here?, with the rubric published in full.