Companies spend anywhere from a few hundred to a few thousand dollars per engineer per year on AI coding tools. That spend makes sense if it raises real engineering output. Below are the public numbers: current seat prices, measured productivity uplift from controlled trials, and the cost that never shows up on the invoice.
What one seat costs per year
- GitHub Copilot: Business $19 per user per month, Enterprise $39 — $228–468 per user per year (billing docs).
- Cursor: Teams Standard $40 per user per month, Premium $120 (annual billing: $32 and $96) — $384–1,440 per user per year; the top individual tier is $200 per month (cursor.com/pricing).
- Claude Code: Team Standard $20–25 per user per month, Premium $100–125; the individual Max 20x tier is $200 per month (claude.com/pricing).
- Credits and API usage above the included allowance are extra. A 10-person team on premium seats is about $1,200 per month, or close to $14,400 per year, before overage.
What that money buys, measured
- DORA 2025, nearly 5,000 respondents: more than 80% say AI increased their productivity, 59% see better code quality, but 30% still do not fully trust AI output (dora.dev).
- Black Duck's July 2026 survey: 97% of teams use AI coding assistants, 92% report improved productivity and release velocity (58% call it a major improvement), developers reclaim about 8 hours per week, and 53% grew total code volume by more than 25% (DEVOPSdigest summary).
- Google's randomised controlled trial with 96 full-time engineers: about 21% faster on one well-defined, enterprise-grade task (arXiv 2410.12944).
- METR's randomised controlled trial on experienced open-source developers: 19% slower with AI access, while those same developers believed they were 20% faster (METR, July 2025). A February 2026 update estimates an 18% speedup for returning participants, but the confidence interval still crosses zero and METR says the data is biased by developers who refuse to work without AI (update).
- METR's May 2026 survey of 349 technical workers: self-reported 1.4–2x change in the value of their work, with METR itself warning the magnitude should be treated with scepticism (METR).
The real cost starts after the code is generated
- Nearly 90% of teams hit issues with AI-generated code, concentrated in manual review (52%), security testing (51%) and code rework (48%).
- 68% say they need an automated system for tracking AI-generated code; only 30% have full governance in place for it.
- Teams with full governance are 55% more likely to report a major efficiency improvement — governance is a direct ROI multiplier, not paperwork.
- Stack Overflow 2025: 84% of developers use AI, but 46% distrust the accuracy of its output versus 33% who trust it, and only about 3% report high trust (survey).
So what do you evaluate engineers on
If the bottleneck moved from writing to checking, the metric has to move with it. Lines of code are now an input you buy by the token, not a signal of value. What is measurable sits downstream: how much of the generated code passes review without being rewritten, how long review takes, which defects reach production, and how long it takes from requirement to first working change.
- Share of code that passes review without a rewrite — the closest direct measure of judgement when using AI.
- Review time and rework rounds per change, not number of changes.
- Defects escaping to production, split by type: logic, threshold, security.
- Technical decisions a person can explain in writing: what went into the prompt, whether the output was trusted, and owning the result once it ships.
Evaluation shifts from how many lines you wrote to where your judgement was right and whether you owned the outcome. It is the same principle as in Measured: Jev, laya and an LLM: the model should only be the policy layer, thresholds and judgement belong in code and in people.