The moment you open up AI chat or an internal AI platform to the whole company via API, token costs become a line item. And that line grows with headcount, and grows again every time someone reaches for a more capable model. A few hundred dollars a month gets waved through as "experimentation." Once you're into the thousands, then tens of thousands, someone upstairs starts asking questions.
Team leads and managers will be expected to explain whether the results justified the spend. If I'm being honest, token costs are a rounding error next to salaries... but if you leave things wide open, the bill balloons alarmingly fast, so some kind of brake is needed. The problem is that the obvious brake — showing each person "$xx" next to every request — makes everyone timid. People start using the tool with one eye on the meter, and adoption stalls.
So how do you keep a brake on things while still getting people to actually use it? I've been chewing on this recently, and here's where I landed.
The two standard measures of AI usage are token consumption and dollar-denominated credit burn.
Token counts tell you how much the tool is being used, but say nothing about what came out of it. Efficiency per token varies wildly between a cheap model and a frontier one, so at best it's a fuzzy total-volume number.
What about the dollar figure?
From a team lead's seat, it gives you a rough sense of how much your team leans on AI, and you can probably feel whether the results are worth it. But one or two layers up, that feeling is gone. Nobody can see how performance is scaling against cost — all they see is an add-on line item in the thousands or tens of thousands. Now look at it from the individual's side. If the real cost is hidden, AI feels like tap water. People will reach for it to tweak a comment style, to find the doc they open every day, to bump a version number from 0.0.1 to 0.0.2. But make the dollar figure visible and it's "that cost $xx?!" — and hands freeze over the keyboard. Because on a personal scale, token costs genuinely are expensive.
For reference: with a Sonnet 5 or GPT-5.6-terra class model, a single turn that fires off 20–30 tool calls runs somewhere around $1–2 (context size permitting). That adds up. You hit $50, $100 before you know it. Show the conscientious people the per-run price tag and they'll use it less. The careless ones won't blink and will keep burning. Net result: worse cost-efficiency across the board. So: token counts are informational at best, raw dollars are hard to show, and hiding them isn't good either. What's left?
The key insight is that AI output is, broadly, a substitute for labor cost — so the metric should make those two things directly, intuitively comparable.
(Someone has probably proposed this before, but) after turning it over for a while, I came up with a unit I call Human-Time. It's a relative figure anchored to the fully loaded cost of employing a mid-level person at your company, or in your industry.
Take a baseline of roughly $50K a year, fully loaded (adjust for your own market). Then:
| Cost | Human-Time |
|---|---|
| $200 | 1 human-day |
| $25 | 1 human-hour |
| $0.42 | 1 human-minute |
| $0.0069 | 1 human-second |
Now any given run can be compared, more or less head-to-head, against handing the same task to a person.
If digging through scattered internal docs and writing up a report costs 1 human-minute, you've more than made your money back. If a not-especially-complex task is eating 15 human-minutes per run because someone's prompt is bloated, well — that's something worth looking at. What members see should always be Human-Time, never dollars.
If you run a customizable in-house AI service, attach the Human-Time figure to each turn as a rough "what this just cost." If you can't, a daily usage report sent to each person works too. Either way, the goal is to give people a nudge to reflect: was that efficient for what I got, or was it needlessly expensive?
From the management side, the guidance is *not* "make this number smaller." It's "show me how much you can get done with it." Lots of usage, big results — that's the sweet spot. Little usage, good results — there's room to expand. Lots of usage with little to show for it, or little usage with little to show for it — room to improve.
The point is to hide the sting of real money and steer attention toward getting value out of the tool.
The same unit works at team level, naturally.
If a team burns **20 human-hours** a week, and you can't point to 20+ hours of work that got created as a result, the ROI is a bit off and it's worth asking whether the tool could be used more efficiently. Conversely, if a five-person team is only using **30 human-minutes** a week, there's almost certainly untapped potential.
Using Human-Time as a shared *communication* unit is what lets you roughly measure results against cost — and that's the whole reason for the unit.
I've actually built this into our platform, along with the accompanying stats, and have been testing it.
It's not fully rolled out yet, but so far it does seem to act as a gentle brake on individual usage while also giving individual and team results a convincing frame for evaluation and reporting. On balance, I'd recommend it.
So — if you're the person on the hook for rolling out AI internally, or a manager who needs to make those API bills mean something, that's the idea. Take it if it's useful.
Until next time.