Measuring AI's real impact
From token maxing to value maxing
AI made writing code faster. The hard part now is turning that speed into value for the business and the customer.
The shift
Q2 2026
Token maxing
Leaderboards and bragging about who burned the most tokens. Consumption was treated as success, and it turned out to be the wrong leading metric.
“How much AI are we using?”
Now
Value maxing
Adoption is effectively 100%, and code generation is largely solved. What matters is the return on every token, for the business and for the customer.
“What are we getting out of it?”
By the numbers
2×
70 → 140 features per quarter
The right reaction isn't to celebrate but to ask: could we triple it? What does each feature cost, tokens included?
4.5×
longer in code review
How long AI-generated code can sit in review compared with before. The bottleneck moved downstream.
+15–18%
more security vulnerabilities
In code shipped with AI assistance. Without system-level governance, more problems reach production.
10–100×
more code generated
AI is the great amplifier, but humans still review and ship every line.
100%
adoption
Virtually every engineering team now uses AI coding assistants. Measuring adoption no longer tells you anything.
Read with care
The 4.5×, +15–18%, 10–100× and 100% figures are industry estimates with no published methodology. Treat them as indicative.
The bottleneck didn't disappear. It moved.
Writing code (the inner loop) got very fast. The outer loop didn't keep up: review, validation, security, deployment and, finally, customer adoption. A feature is part of a product, which is part of a system, which is part of a platform.
Code
AI-accelerated
Review
up to 4.5× longer
Test & validate
system-level
Security
validate auto-fixes
Deploy
into the platform
Customer adoption
still slow
The surprise: features shipped faster than ever, but customers still run their own validation before rolling anything out, even when configuration is automated. More features isn't always better, so the pace of delivery needs rethinking.
Six lessons for engineering leaders
1. Govern alongside the rollout
Governance should ship with the tools, not after training and usage tracking. Retrofitting it later costs more.
2. Make learning organizational
Set up a central hub of effective prompts, examples and outcomes. Ask “has someone solved this already?” before starting from scratch. A bonus: teams collaborate more.
3. Keep governance invisible
Build governance, security and compliance into the developer's flow so it isn't one more pipeline step. A small central team looks after it so developers don't have to.
4. Measure value delivered, not tool usage
AI is part of the SDLC, not a tool. Compare customer adoption of features built AI-first end to end with features built the legacy way.
5. Mine the ship-to-adopt gap
Customer value takes time to show up. Meanwhile, measure simplicity, efficiency, quality, accuracy and security. The gap is rich data that few teams look at.
6. Put tokens on the P&L
Track token spend per product, as you already do for people and cloud. A product nearing end of life shouldn't burn as much as a growth bet. Forecast it before Finance caps it.
How the metrics evolved
Phase 1 · Usage
Who uses which tools · Which products use them · Time spent with AI on code gen, quality, validation
Phase 2 · Value
Cost per feature, tokens included · Customer adoption of AI-built features · End-to-end flow, not lines of code · Token spend per product
Questions to take back to your org
- Do we measure cost per feature, tokens included, or only speed and usage?
- Where is our bottleneck today: review, validation or deployment? Do we measure end-to-end lead time?
- Is AI governance moving in step with the rollout? Who owns it so developers don't have to?
- Do we have a shared hub of prompts, patterns and solved problems?
- Can we attribute token spend per product or team and forecast it before Finance does?
- Do we track customer adoption of what we ship, and how long it takes to get there?