Most people budget for AI the way they budget for a gym membership: a monthly subscription they hope to use enough to justify. That framing hides the number that actually decides whether automation makes sense. The metric you want is ai cost per task, meaning what does it cost to summarize one document, draft one reply, or classify one support ticket. Once you can answer that, the whole ROI question becomes arithmetic instead of a leap of faith.
Subscriptions blur this on purpose. A flat fee per seat feels predictable, but it tells you nothing about the unit economics underneath. When you switch to the API layer and pay per token, the cost per task drops into view, and it is usually far smaller than people expect. That small number is what shifts the frontier of what is worth handing to a machine.
Let me work three real examples at current API prices, then compare each against the labor it replaces. The point is not the exact rupiah, which keeps falling. The point is the method, so you can rerun it whenever prices drop again.
The Method: Tokens In, Tokens Out, Minutes Saved
Every AI task has three cost inputs you can estimate:
- Input tokens, roughly the length of the document or prompt you feed in. A token is about four characters of English, so 1,000 words is near 1,300 tokens.
- Output tokens, the length of what the model writes back.
- The human minutes it replaces, the honest comparison, priced at your actual labor rate.
For an Indonesian SME, a competent admin or junior staffer costs somewhere around Rp 4,000,000 to Rp 6,000,000 per month. Call it Rp 30,000 per hour, or Rp 500 per minute, as a working figure. Now every task has a machine cost and a human cost, and you can compare them directly.
Three Worked Examples
Summarize a five-page report
A five-page report is roughly 2,500 words in, or about 3,300 input tokens. A tight summary is maybe 300 words out, near 400 output tokens. At mid-tier model prices in late 2023, that lands around Rp 100 to Rp 300 per summary depending on the model you choose.
A human doing the same read-and-summarize honestly takes 15 to 25 minutes. At Rp 500 per minute, that is Rp 7,500 to Rp 12,500. The machine is cheaper by a factor of thirty or more, and it returns the result in seconds.
Draft a customer reply
A support reply needs the incoming message plus a bit of context, maybe 500 input tokens, and produces a 150-word answer, around 200 output tokens. Cost per drafted reply sits near Rp 50 to Rp 150.
A staffer writing that reply from scratch spends 5 to 10 minutes, or Rp 2,500 to Rp 5,000. Even after a human reviews and edits the draft, which they should, the net time saved is real. The ai cost per task here is close to a rounding error against the labor.
Classify a support ticket
Routing a ticket into a category needs the ticket text in, around 300 tokens, and a one-word label out. This is the cheapest task of the three, often under Rp 30 per ticket, and it runs in a fraction of a second.
A human triaging tickets spends maybe 1 to 2 minutes each, Rp 500 to Rp 1,000. The ratio is smaller here, but at volume, say 2,000 tickets a month, the difference is Rp 60,000 in machine cost against Rp 1,000,000 to Rp 2,000,000 in staff time.
Where the Frontier Actually Sits
Here is the pattern across all three: the machine wins hardest on high-volume, low-judgment, repetitive tasks. It wins least when the task needs real judgment, taste, or accountability, because those still need a human in the loop, and a reviewing human eats into the savings.
So the ai cost per task frontier looks like this:
| Task type | Automate fully | Automate with review | Keep human |
|---|---|---|---|
| Classify, tag, route | Yes | ||
| Draft, summarize, translate | Yes | ||
| Decide, negotiate, apologize | Yes |
Falling prices keep pushing tasks up this table. A summary that cost Rp 1,000 two years ago costs a fraction of that now, which means tasks that were not worth automating suddenly are. This is why the frontier is not fixed. You should rerun the arithmetic every few months, because the line between "too expensive to bother" and "obviously worth it" keeps moving in one direction.
None of this replaces judgment about whether AI belongs in a process at all. Cheap does not mean appropriate. I have written separately about when not to use AI, and cost per task does not override those cautions. A task that is cheap to automate but wrong to automate is still wrong.
What This Means for Your Budget
Stop asking "can we afford an AI subscription" and start asking "which of our repetitive tasks cost more in staff minutes than in tokens." The second question is answerable with a spreadsheet and an afternoon.
A practical way to start:
- List your five most repetitive text tasks.
- Estimate the human minutes each consumes per month, times your labor rate.
- Estimate the ai cost per task using the token method above, times the volume.
- Compare. The gaps will be obvious, and usually larger than you guessed.
The reason unit economics matter more than subscription pricing is that they tell you where to point the automation, not just whether you can afford it. A seat license makes you feel like you should use AI everywhere to get your money's worth. Cost per task tells you exactly where it pays and where it does not, which is a much healthier way to spend.
The Practical Takeaway
Budget AI the way you budget any input cost: per unit of work, against the labor it replaces. When you price a summary at a few hundred rupiah against fifteen minutes of staff time, the decision stops being a matter of hype or fear and becomes plain arithmetic. Run the numbers on your own tasks, and rerun them as prices fall, because the set of tasks worth automating only grows. If you want a sharper read on which of these actually survive contact with production, the demo-to-production gap is where most of the surprises live.