Is your AI feature losing money on every call? The 20-minute check.
Here’s a quiet way to lose money with AI: ship a feature where one use costs you more than it brings in, then treat rising usage as good news. Every new user makes the loss bigger. It happens more than you’d think, because almost nobody checks the one number that reveals it.
Most teams watch total AI spend. Very few know what a single use of a given feature costs versus what it earns. That gap is exactly where an unprofitable feature hides — and if you scale it, you scale the loss. Before you pour traffic on an AI feature, spend twenty minutes on its unit economics.
1. Cost per use (10 min)
Pick one feature — the AI summary, the auto-reply, the “ask our docs” box. Get two numbers:
- Its model spend for last month. Group your provider bill by the key or tag tied to this feature. (If everything runs through one key and you can’t isolate it, that’s the first thing to fix — you’re flying blind.)
- How many times it ran — from your own logs or analytics.
Divide. Now you have cost per use: “one AI summary costs us $0.11,” “one support auto-reply costs $0.34.” That sentence is the whole game, and most teams have never said it out loud.
2. Revenue per use (5 min)
Now the other side. What does one use actually bring in?
- Paid, metered feature: what the user pays for that use. Easy.
- Part of a paid plan: the slice of the subscription this feature is responsible for — rough is fine, but be honest.
- Free / onboarding feature: its contribution to conversion or retention. This is where people fool themselves. If a free AI feature doesn’t measurably move conversion, its revenue per use is close to zero — and any cost per use is pure loss.
3. Compare, and be honest (5 min)
Put them side by side. You’re not just looking for cost above revenue — you’re looking for cost anywhere near revenue, because:
- Model prices drift upward, not down, so today’s thin margin is tomorrow’s negative one.
- A few heavy users can pull the average cost per use far above what the typical user pays — the average hides them. Check the per-user distribution, not just the mean.
Watch cost per use, not just total spend.
If cost per use is close to or above revenue per use, growth is not traction. It’s a faster meter.
The three fixes when it’s upside down
You usually don’t have to kill the feature. In order of least disruption:
- Send the work down a tier. Most of what a feature does doesn’t need the priciest model — route the easy calls to a cheaper one, or replace the deterministic parts with a function so they cost nothing per run. This alone often flips the sign.
- Cap the heavy users. If a handful of users drive most of the cost, a per-user cap protects your margin without touching everyone else.
- Gate the expensive version. Put the costly AI feature behind the paid plan, so the people generating the cost are the people paying for it.
The one-line habit
Add cost-per-use to how you think about every AI feature, next to the metrics you already track. “This feature costs $0.11 a use and earns $0.40” is a business. “We spent $9,000 on AI last month” is not a sentence you can act on. The first tells you what to scale; the second just tells you to worry.
If you’re not sure whether an AI feature is making or losing you money — or you want the per-feature numbers pulled without wiring up tracking yourself — that’s a quick job for me. It’s just me, so it comes straight to me and I reply fast. Send me the feature and your rough numbers and I’ll tell you the cost per use, whether it’s underwater, and the cheapest way to flip it — free, within a business day.