Engineering management in the generative AI era
𝗝𝗲𝗻𝘀𝗲𝗻 𝗛𝘂𝗮𝗻𝗴 𝘀𝗮𝘆𝘀 𝘁𝗼𝗽 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝘀 𝗺𝗶𝗴𝗵𝘁 𝗯𝘂𝗿𝗻 $250k in AI tokens per year.
𝗕𝘂𝘁 token burn ≠ productivity 📉—here’s why it could backfire ⬇️
Recently, 𝗝𝗲𝗻𝘀𝗲𝗻 𝗛𝘂𝗮𝗻𝗴 argued that if engineers aren’t using significant AI compute, it’s like designing chips with “paper and pencil.”
He even suggested top engineers might reasonably use hundreds of thousands of dollars worth of tokens annually.
That framing is powerful.
But it also points to a potential trap:
👉 When does a useful signal become the wrong metric?
Let me ground this in a real scenario.
I’ve seen this pattern play out across multiple waves of tooling—
from code metrics to cloud adoption to AI.
Two engineers:
Engineer A
• Thinks deeply about the problem
• Designs intentionally
• Uses AI surgically
• Ships something that actually moves the needle
Engineer B
• Starts prompting immediately
• Hits dead ends
• Retries constantly
• Burns 10× the tokens to ship something… average
If token usage becomes the signal,
Engineer B starts to look more “productive.”
But Engineer A is the one creating real value.
We’ve seen this movie before:
• Lines of code
• Hours logged
• Commits per week
None of them measured what actually mattered.
Here’s where it gets subtle.
Token usage is a useful signal—at first.
Early on, it helps answer:
• Are teams adopting AI?
• Are they unblocked?
• Are they experimenting?
I’ve seen teams (including mine) use it exactly this way:
👉 as a way to drive behavior change and adoption
And it works.
For a while.
Then the shift happens.
Signal → Target → Game
This is exactly what Goodhart’s Law warns about:
When a measure becomes a target, it ceases to be a good measure.
If token usage becomes a performance metric, we’ll get:
“Tokenmaxxing”
📈 More tokens spent
🔊 More noise generated
⚠️ No guarantee of better outcomes
In other words: productivity theater
What high-performing teams actually do
Once adoption is there, the best teams move on:
👉 They stop measuring inputs
And go back to measuring outcomes
📦 What shipped?
✅ Did it work?
💡Did it create customer value?
Token usage is a system metric.
It becomes dangerous when it turns into a people metric.
The bigger shift happening
We’re entering a world where:
→ Compute is abundant
→ AI is embedded in every workflow
→ “Using AI” is table stakes
So what differentiates engineers?
Not how many tokens they consume—
but how effectively they turn compute into impact.
Are we driving further—
or just burning more fuel?
Curious how others are handling this:
Are you tracking AI usage at the individual level?
Or focusing purely on outcomes?
What’s actually working in your teams?
👇