Velocity, efficiency, and operational excellence all have new definitions.
This week I’m sharing six metrics I have our company brain track and surface to me each week.
PS: since I launched my shortform video agency, JOLT!, last month, we’ve seen founders we work with go from nearly 0 to over 10k followers (like the one below).
Want me to do the same for you? Let’s chat.
5 AI-Native Metrics to Track
Revenue per AI Dollar
Trailing 4-week revenue / AI spend (aka model APIs, agent tooling, subscriptions)
Tokenmaxxing is irresponsible.
There, I said it.
It’s not incoherent to say startups should be AI-native, while still advocating for efficiency.
Purposeless spending will quickly cost you runway.
Founders lie to themselves, and saying that your company is actually doing better just because it’s spending more on AI tooling is one of those ways.
That’s the opposite of creating leverage.
Revenue per Human Dollar
Trailing 4-week revenue / total people cost
If we’re tracking efficiency for AI spend, and we believe AI can augment human spend, we should also track efficiency for human spend.
It’s not just for payroll expenses either. Consider things like how much you’re spending on external agencies and partners as well.
You can increase this either by growing revenue or reducing human spend, so this isn’t just about minimizing human-related costs.
And be careful that any reductions to this doesn’t increase your revenue per AI dollar spent by the same amount. Just moving expenses around is pointless.
Hours Returned per Week
# of tasks completed by agents * avg minutes previously spent per task by humans
Everyone wants AI to help them get more done in less time. So, track how much time you’re getting back.
This one requires some estimation about the average number of minutes the tasks that are currently being automated by AI would have taken humans.
Keep yourself honest by:
Only counting actually completed work
Keeping a record of your assumptions about how long specific things take, and having the calculation specifically reference them
Re-checking your assumptions every month, or quarter
Intervention Rate
% of agent outputs that require human correction
An autonomous system should operate, well, autonomously.
No system is going to be perfect, but you should continuously be trying to move it in the direction of perfect.
Easier said than done, since you can build faster than ever now.
But without bringing this down over time, you’ll likely still see the amount of work required by humans to maintain core systems still scale linearly alongside growth.
Capture Rate
% of meetings, decisions, and key context that actually made it into your company’s brain
A well-run AI-native company is constantly farming for context to add to its knowledge base.
That means recording every meeting transcript, logging every decision in a ledger, and keeping tabs on everything else too.
That would’ve been tedious before LLMs, but now it’s a potential advantage.
This is the one that’s hardest to track because, if something isn’t logged, the system won’t know.
That’s why each week I have our system run a script that looks at how many meetings were held and whether a Granola transcript is associated with it in our logs.
I also have it run a diff of our decision log relative to a scan of all public comms (Slack channels, my emails, etc).
It’s not perfect, but it gives me a sense of whether we’re doing a good job collecting context.