Production AI, grounded in backend and search engineering.

The floor is not a number you pick. It is one you solve for. Set EV to zero and read off the business value that gets you there.

EV = p·bv − cost − risk. The router costs $0.0017 a request and carries a $0.05 risk reserve, so $0.0517 leaves before any value arrives. Retrieval answers 67.0% of requests, so each dollar of business value returns only 67 cents of expected value.

bv = ($0.0017 + $0.05) / 0.67 = $0.0772

Below that, nothing ships. Above it you are technically positive and practically exposed: at $0.08 the router clears $0.002 a request — $76 a month at 40,000 requests, less than the cost of the meeting where you argue about it.

So break-even is not a target. It is the line under which the conversation should stop. Anything within a few cents of it gets erased by the first assumption you got wrong — a lower success rate, one more retry per request, a support ticket nobody counted.

Notice which term dominates. At the floor the $0.05 risk reserve is 97% of the cost side and inference is a rounding error. Cheaper tokens will not move this line. Only a more valuable request will, which makes “what is a solved request actually worth?” the first question, not the last.

$0.42 is the measured value of a solved request in the EV+ experiment — a support answer that would otherwise cost a person a few minutes. At one step, that is the whole model.

EV = 0.67 × $0.42 − $0.0017 − $0.05 = +$0.230

Every retrieval-backed request returns 23 cents. At 40,000 requests a month that is $9,188. The no-retrieval baseline stays negative at −$0.052, because a strategy that answers nothing correctly still pays inference and still carries risk.

The number that matters is not the margin, it is the payback. Fine-tuning cost $20 and lifted p(success) by 0.45, so each request gains 0.45 × $0.42 = $0.189, and $20 / $0.189 = 106 requests. The investment is repaid before lunch on day one; everything after is surplus.

This is the shape most projects actually have — unremarkable per request, decisive at volume, and robust enough that halving the success rate leaves it positive.

The cost side barely matters at this value: $0.0517 of cost against $0.281 of expected value. The levers are p(success) and business value, not the price of tokens — which is why the router beats RAG by a hundredth of a cent and nobody should spend a sprint on it. It will not, however, survive a human gate: one $0.50 approval takes this to −$0.270.

High-value work tends to be multi-step work, which is where the step model starts to matter. A request worth $1.00 — a contract question, a reconciliation, something that replaces twenty minutes of an analyst — runs 12 steps rather than one.

At 99.5% per step, the 11 steps after the answer survive at 0.995 ¹¹ = 94.6%, so p(task) falls from 67.0% to 63.4%.

EV = 0.634 × $1.00 − $0.0017 − $0.05 = +$0.582

That is $23,294 a month, and the $20 one-off repays after 45 requests.

Note what did the work. Value per request rose 2.4× and the margin rose 2.5×, because reliability cost almost nothing — 99.5% over 12 steps is still 94.6%. Run the same 12 steps at 95.0% and you get 56.9%, p(task) 38.1%, and a margin of $0.329 — a 43% cut from half a point of per-step reliability. At high value and many steps, reliability is worth more than the model.

Margin this wide also buys the one thing the cases above cannot afford: a human in the loop. A single approval at $60/hour for 30 seconds costs $0.50 a run, and this work still clears $0.082 after paying it. The $0.42 case does not survive that gate at all. If you want a person on the irreversible steps, the way to afford them is to raise the value of a request, not to shave the inference bill.