Engineering leaders evaluate agent projects with a familiar spreadsheet: tasks automated × cost per task = savings. The spreadsheet almost always omits a line item that quietly determines whether the project succeeds: the recurring cost of failures that never get fixed.
An agent failure that produces no system change isn't a one-time cost — it's a subscription. You will pay for that same mistake next week, and the week after, for as long as the agent runs.
Anatomy of one failure, fully costed
Take a support agent that mishandles partial-refund requests — a real pattern I've seen more than once. Each occurrence costs:
- The redo: a human agent handles the ticket anyway — the cost the agent was supposed to eliminate, plus the agent's own inference spend on top.
- The cleanup: the customer is now on their second contact and measurably angrier; handle time is longer than if no agent had touched it. Some percentage escalate further or churn.
- The investigation tax: every few weeks, someone senior notices "the refund thing" again, pulls traces for an hour, concludes it's the same issue, and moves on — because fixing it properly means touching a 4,000-token prompt nobody dares refactor without a safety net.
Now the multiplication. If partial-refund requests are 3% of 2,000 weekly tasks, that's 60 occurrences a week. At a conservative $15 fully-loaded cost each, this single failure pattern costs about $47,000 a year. A typical production agent carries five to fifteen such patterns simultaneously. That's the invisible line item — often the same order of magnitude as the savings on the visible ones.
The two cost curves
Plot cumulative failure cost over time and every agent falls on one of two curves. Open-loop: a straight line climbing forever — same failure rate, growing volume, cost compounding with scale. Scaling the agent scales the waste. Closed-loop: a curve that steps downward each time a failure cluster is identified, fixed, and gated against recurrence — each class of mistake is paid for roughly once, and new volume arrives on top of a shrinking failure rate.
The gap between those curves is the entire business case for closing the loop, and it widens with every week of operation and every point of scale. It's also why "our agent works fine" and "our agent is bleeding money" are frequently the same agent described by people looking at different dashboards — the first is looking at task volume (see why dashboards miss quality), the second at where the tickets actually went.
The one-question audit
You don't need a consultant to price this for your own system. Pull last month's worst fifty agent failures and ask one question of each: has this exact mistake happened before? In open-loop systems the honest answer is "yes" for the majority — teams that run this exercise typically find most failures are repeats of a known pattern. Whatever fraction you find, that's the share of your failure budget you're paying as a subscription. Then the investment question inverts: not "can we afford to build a feedback loop?" but "how long can we afford to keep re-buying the same mistakes?"