Unit Economics: Cost per Resolved Task

The price of an answer is not the price of a task

Token prices are only part of the cost. What matters is the cost per resolved task: model calls per task (including retrieval, retries and repair loops), plus the cost of tasks the system cannot resolve and hands to humans, plus infrastructure, evaluation and maintenance. Often the biggest lever is not a cheaper model but a higher resolution rate or fewer calls per task. Build a simple model with your own numbers, and see how sensitive it is: in the example, raising the automatic resolution rate from 50% to 75% cuts the cost per ticket from about 2.03 to 1.03, far more than any realistic token-price saving. Re-run the model when prices, volumes or quality change, and compare it with the value produced (time saved, revenue, satisfaction).

Cost per resolved ticket, run

I ran this with plain Python 3 (standard library only). With 2.4 calls per ticket at 0.012 each (about 0.029) and 38% of tickets going to a human at 4.00, the total is about 1.55 per ticket. The human hand-off dominates; raising the auto-resolve rate to 75% would bring it to about 1.03. All numbers are invented.

# Unit economics: cost per RESOLVED ticket, not cost per call.
calls_per_ticket, price_per_call = 2.4, 0.012       # example numbers
auto_resolve_rate, human_cost = 0.62, 4.00          # example: 62% solved without a human; a human handoff costs 4.00

cost_per_ticket = calls_per_ticket * price_per_call + (1 - auto_resolve_rate) * human_cost
print(f"model cost per ticket : {calls_per_ticket * price_per_call:.3f}")
print(f"human handoff share   : {(1 - auto_resolve_rate) * human_cost:.3f}")
print(f"total per ticket      : {cost_per_ticket:.3f}")
for rate in (0.50, 0.62, 0.75):
    c = calls_per_ticket * price_per_call + (1 - rate) * human_cost
    print(f"  if auto-resolve rate were {rate:.0%}: {c:.3f} per ticket")

Output:

model cost per ticket : 0.029
human handoff share   : 1.520
total per ticket      : 1.549
  if auto-resolve rate were 50%: 2.029 per ticket
  if auto-resolve rate were 62%: 1.549 per ticket
  if auto-resolve rate were 75%: 1.029 per ticket

Compare cost with value

A cost per task only means something next to the time or revenue it saves.

Quick check: In the example, what dominates the cost per ticket?

• The human hand-off for unresolved tickets
• The token price
• The font size
• The server colour