Perspective · AI Provenance
A musician just asked the question every CTO will face in the next twelve months. When AI generates most of the work, how do you prove what humans contributed, and what it’s worth?
“How can you validate human sweat to make something?”will.i.am · Founder & CEO, FYI.AI · Bloomberg, August 2026
He was talking about music. In an interview with Bloomberg this week, will.i.am argued that in the AI era, human creativity needs to be valued above raw compute output, and that the industry has no mechanism to do it. When a machine can generate a thousand tracks an hour, the sweat behind a song becomes invisible. Unprovable. And what you can’t prove, you can’t price.
Replace “song” with “pull request” and he’s describing software engineering in 2026.
More than half of merged code across the industry is now AI-generated, up from roughly a third just one quarter ago. Adoption exceeds 90%. The commit log, the record engineering has trusted for two decades to answer “who built this,” no longer answers that question. A commit today might be 90% agent output with ten minutes of human review, or 20% agent scaffolding wrapped in hours of human architecture and judgment. The log shows both identically.
That’s not an academic gap. Everything downstream depends on knowing where human contribution lives:
Compensation and performance assume you can see individual contribution. Output metrics that a $20-per-month agent inflates can’t carry a promotion decision. Budget defense assumes you can answer the CFO’s new favorite question: if AI writes half the code, what exactly are we paying 500 engineers for? Audit and liability assume you can show a regulator or a court which code was human-authored, who verified the machine-generated part, and when. In the regulated industries we serve at Waydev, banks, insurers, and enterprises with real compliance exposure, “we don’t know who wrote what” is not an acceptable answer. It’s a finding.
Human sweat in engineering is no longer typing. It’s judgment: the review that caught the flaw, the prompt that steered the agent, the confidence to ship.
This is exactly why we rebuilt Waydev as the measurement layer for AI-written code. Two modules do the heavy lifting: AI Checkpoints establish provenance, and AI Impact proves whether the output created value. Together, they’re engineering’s answer to will.i.am’s question.
AI Checkpoints give you commit-level attribution for AI-assisted work. Every checkpoint captures which agent wrote the change, how many tokens it consumed, what it cost, and, critically, the split between AI-generated and human-edited code. Per commit, per repository, per team, per vendor.
Look at what those two checkpoints actually record. The first is an agent doing its job well: high AI share, low cost, one human review, straight to production. The second is human sweat made visible: an engineer who took mediocre agent output and rewrote most of it across three cycles before it was safe to ship. In git blame, these two commits look the same. In Waydev, they tell opposite stories, and the second engineer’s judgment finally shows up in the data.
That’s the provenance layer. It answers who and what. But provenance alone doesn’t answer the CFO. For that, you need to know whether any of it mattered.
The AI Impact module connects AI usage to delivery outcomes. It tracks generated code across the pipeline: how much survives review, how much passes CI, how much reaches production, and how token consumption and cost map against what actually ships. It’s the difference between measuring activity and measuring value.
Illustrative team view. The gap between generated and deployed, 51% in this example, is where tokens burn without value: rejected suggestions, rewrites, failed CI runs. That gap is your real cost of AI, and most orgs have never seen it.
This is where will.i.am’s second point lands. Human creativity, he argued, should be worth more than raw compute output. AI Impact is how you operationalize that in engineering. Compute output is the top of the funnel: tokens, generations, suggestions. Value is the bottom: code that a human judged worthy, that survived the gates, that runs in production serving customers. The funnel between them is the exchange rate between compute and value, and every vendor, team, and workflow has a different one.
Once you can see that exchange rate, decisions get simple. Compare Copilot, Cursor, and Claude Code on cost per shipped PR instead of vibes. Spot the team whose acceptance rate says the tool fights their codebase. Catch the repo where AI code passes CI but drives rework two sprints later. Kill the spend that never reaches production before renewal, not after.
Three forces are converging. AI spend is compounding, with median quarterly budgets up roughly 28x among large tech companies year over year. Quality signals are diverging, with PR sizes doubling while change confidence falls. And accountability is arriving, from CFOs interrogating budgets to regulators drafting AI provenance requirements for critical software.
Every one of those pressures resolves to the same primitive: a trustworthy, auditable record of who and what produced each unit of work, and what happened to it. Music doesn’t have that system yet. That’s why will.i.am is asking the question. Engineering does:
The organizations that build this record now will spend the next decade making confident decisions about people, tools, and budgets. The ones that don’t will be arguing from anecdotes while their most valuable engineers, the ones whose judgment quietly saves every release, stay invisible in the data.
Human sweat should be worth more than compute output. In engineering, it finally can be measured that way.
See AI Checkpoints and AI Impact on your own repositories. Track every agent, every token, and every line from commit to production.
Book a demowill.i.am quote from his Bloomberg interview with Mishal Husain, August 2026. Industry figures from DX’s State of AI Impact in Engineering, Q2 2026 (500+ organizations). Checkpoint and funnel data shown are illustrative.
Ready to unlock your SDLC productivity?
