Pricing agents as a builder

You built a support agent and priced it like software: per seat, per month. Your buyer's CFO cannot see what it delivers, so every renewal is a negotiation about vibes. Here is the same business, re-priced on the output meter, first in business terms, then in math.

The story, in business language

You sell a ticket-resolution agent to a mid-market SaaS company. Today: 20 support seats at $99, so $1,980 a month, and at renewal the CFO asks what it did. You answer with activity: tickets touched, deflection estimates, a case study. The CFO hears a cost center.

Under quirq pricing the conversation inverts. The buyer's own team budgets a resolved ticket at $4 (their loaded human cost for the same outcome). Your agent works the queue inside the buyer's instrumented environment; every resolution is verified by state (ticket closed, reply sent, KB linked, thread untouched) and mints quirqs. You are paid an agreed share of what settled. Nothing settles, nothing is owed. At renewal there is no narrative: the ledger shows what your agent delivered, what it cost all-in, and what the buyer would have paid humans for the same quarter.

The same month, quantified

2,600 tickets attempted, 92% settle atomically at B = $4, revenue share α = 25%:

Revenue up 21%, margin intact, and the buyer sees a 4.0x return on their spend in a unit their finance team already speaks. The number that wins the deal is theirs, not yours.

The spec sheet that replaces the benchmark

Your public rating becomes two auditable numbers with a window attached, per the relativity rules: QER on the buyer's unit types, and the intervention rate with its trend. "QER 4.0x on support units, n = 2,600/mo, IR 8% falling, audit gap in tolerance" is a claim a prospect can verify in their own pilot ledger in two weeks. Benchmarks measured what your agent could do once; the ledger measures what it does in production, priced.

The math, as you scroll deeper

Revenue and floor. Your revenue on window T is R = α · ΣQ(u). Your floor comes from cost per quirq: with C_total = $333 against Q = $9,568 minted,

c_q = C_total / Q = 333 / 9,568 = 0.035

You are viable at any α above 0.035; you quoted 0.25. That 7x headroom is your margin today and your price war tomorrow.

The participation band. Both sides must prefer the deal. The buyer, whose budget B is anchored at their human baseline, saves (1 − α)·B per settled outcome, so the buyer participates for any α < 1. You participate when revenue covers cost, α ≥ c_q. The viable band is

c_q  ≤  α  <  1          here: 0.035 ≤ α < 1

Competition between builders compresses α toward c_q plus a competitive margin; tenure (the falling cost curve of how to calculate it) pushes c_q itself toward the execution floor. Both forces hand surplus to the buyer over time, which is what adoption curves are made of: your defensible earnings are the c_q gap you maintain against rival builders, not the α you charge early.

Escrow and SLA, from the machinery. For units with a terminal judgment check of weight w_j, the holdback fraction is w_j / Σw, released on verdict: escrow with no new contract language. An intervention-rate SLA prices risk the same way: post a bond and refund β · IR(T) · ΣB per window, which makes your IR trend a hard financial commitment rather than a marketing line. At IR 8% against a 10% SLA with β = 0.5, your exposure is zero; let quality slip to 14% and the month's penalty is 0.5 × 0.04 × $10,400 = $208, growing linearly with the slip.

Dynamic pricing. Quote α as a function of your own tenure curve: α(t) = (1 + m) · c_q(t) for target margin m. As memory accumulates and c_q decays, you can cut price and hold margin, or hold price and widen it: the choice is visible a quarter ahead because the cost curve is measured, not guessed.

The single-builder view. Zoom out to all builders at once and the same ledger becomes a market: reading agentic markets.