By the Onsites AI team · Last updated · 5-minute read
Per-seat pricing survived four decades because seats were a reasonable proxy for work: one human, one license, roughly one share of the tickets. In 2026 the proxy breaks. AI now drafts the replies, reads the tracking, dates the quotes and summarizes the sagas — a desk of five people with good tooling handles what a desk of fifteen buckled under, and the fifteen-seat price those five would have paid is pure margin the vendor no longer earns. The industry's quiet crisis isn't that customers can't pay; it's that the meter counts the wrong thing. This piece maps the three forces killing the license model, where support pricing is actually heading, and how to buy during the transition without signing a contract that the dying meter still holds.
The strongest force is arithmetic. When a copilot drafts at published credit prices — 3 for a draft, 6 for a suggested reply — one person's hourly output roughly doubles to triples on the routine half of the queue, and the tool-calling tier absorbs whole categories (WISMO, stock checks, status confirmations) without a head at all. A team that would have hired seat number seven instead buys $40 of credits. The vendor's answer, historically, was to move the AI behind its own paywall — a per-seat license *plus* an AI subscription *plus* per-resolution fees in the aggressive cases, the stack the ten-person math prices at ≈$2,300–$3,100/mo against ≈$255 on a flat meter — but stacking meters on top of a broken one doesn't fix the proxy, it just triples the invoices. The desk feels this as the gap that widens with every step: volume multiplies the outcome lines while seats merely add.
The second force is competitive. Software that's genuinely free at small scale — full product, three seats, published AI credit prices, no per-resolution fees — resets what the next vendor must beat. A five-person team's rational default is now "free until it isn't," and the growth ceiling is knowable in advance (seat seven at $15, the catches checked up front). Legacy pricing can't follow vendors into $0; margins built on $19–$99/seat/mo don't survive a floor of zero, so the incumbents respond with the only lever left: metering outcomes — per-resolution fees in the $0.99 era, AI interaction counts, "flexible" credits — which converts the death of the seat meter into a livelier meter that compounds with your success instead. The buyer's tell is simple: any vendor whose price list gets shorter is competing on value; any whose list gets longer is defending old margins.
The third force is literacy. A decade of subscription fatigue taught finance teams to model support software like infrastructure, and the artifacts of that literacy are now standard diligence: the bill under a peak month (the CFO's worked example), the total-cost curve across hiring scenarios (the three sizes, both meters), the exit cost of the data (the export ritual). Metered pricing submits badly to this light: per-resolution and per-interaction lines are exactly the ones that scale with the business's good months, and diligence finds them. The 2026 buyer asks one question the 2019 buyer didn't: *what does this cost when we grow?* — and the honest answers sort the market into flat meters (seats as a headcount fee, work included) and metered ones (success taxed). Procurement now reads the difference in the demo's first fifteen minutes, because the vendor who can't show you the bill under growth is hiding it.
Three shapes are converging, and they're not mysterious. Flat meters with published extras: a seat count (or a free band under it) plus published unit prices for scarce inputs — AI credits at roughly a cent, storage at $1/GB/mo — the shape this desk prices at (3 seats free, $15/seat/mo self-hosted, credits prepaid and never expiring). The lines are few, the derivations are checkable, nothing compounds with volume. Outcome-sharing at the enterprise ceiling: above a few hundred seats, contracts increasingly reference verified outcomes (with satisfaction floors, the honesty condition) rather than raw resolution counts — slower to spread than per-resolution fees precisely because it's harder to game. Self-hosted capacity pricing: the self-hosting renaissance pushes a fourth shape — pay per seat for the license, bring your own AI model and storage, marginal cost of growth near zero. The common thread: charge for what costs the vendor something (seats, compute, storage), stop charging for what costs the customer something (their own success).
Concrete, because abstractions hide the death. Take the ten-person desk from the ten-person math: 4,000 conversations a month, 20% AI-assisted, one hire planned next quarter. Seat-as-proxy meter (2020): ten licenses, maybe $59/seat → ≈$590/mo, and the hire takes it to ≈$649 — the model was simple because the proxy still held. Stacked meter (2024): the same ten seats repriced with an AI add-on at $30/seat plus per-resolution on the AI share — ≈$1,200 seats + ≈$794 AI line → ≈$2,300–$3,100/mo, and the hire pushes it higher while doing less of the work themselves. Flat meter (2026): $105 for ten seats (4 beyond free at $15) + ≈$100 published credits + $0 storage → ≈$255/mo, the hire's seat $15, the credits flexing with usage and never compounding. Same workload, same desk, three invoices — the middle one is the transition's lesson: vendors don't die quietly, they reprice. The worked bill under a peak month is where the stacked meter confesses; the flat meter's bill is the same on the worst month of the year. The desk that models all three before renewing doesn't just save $2,800/mo — it avoids signing a two-year contract with the meter that's already dying.
Practical terms, since contracts outlive marketing. One: model your bill under growth *before* signing — three scenarios (flat, 2×, 4× volume) through every meter the vendor names; the worked sizes are a template. Two: reject per-resolution and per-interaction lines, or cap them hard in the contract — a meter that taxes good months is a tax on your success, per the peak-month math. Three: demand AI's price be published, not "contact sales" — a credit table (the published rates) is the difference between a line item and a negotiation. Four: confirm the exit on day one — full export, the day you leave, no ransom in fidelity. Five: prefer annual billing only where the meter is flat; monthly on any outcome meter so a bad quarter's bill can't surprise you. The transition favors buyers who read meters, and the reading is now the easy part — the vendors still selling seats-as-proxy are pricing for an economy that AI already ended. Choose the meter that wants you to grow; it's the only one that still makes sense in 2026.
Why is per-seat pricing dying?
Because seats stopped tracking work. AI handles 20–40% of queue volume without a head, one person with a copilot does the work of several, and a meter that counts licenses while the real driver is AI share leaves money on the table — for the vendor, not you.
What are vendors doing to defend per-seat revenue?
Stacking meters: an AI subscription layered on seats, per-resolution fees, interaction counts and bundles. It converts the broken seat meter into outcome meters that compound with volume — the ten-person stack that prices ≈$2,300–$3,100/mo vs ≈$255 on a flat meter.
What replaces per-seat pricing?
Flat meters with published extras: a seat count or free band plus unit prices for scarce inputs (AI credits at about a cent, storage at $1/GB/mo). Above a few hundred seats, outcome-sharing with satisfaction floors; on the self-hosting edge, capacity pricing with BYO AI and storage.
How do I avoid overpaying during the transition?
Model three growth scenarios through every meter before signing; reject or cap per-resolution/per-interaction lines; require published AI credit prices; confirm full-fidelity export on day one; bill monthly on any outcome meter. Prefer vendors whose price list is getting shorter.
Is any version of per-seat pricing defensible?
Yes — as a plain headcount fee on a flat meter with work included and no outcome lines. What's indefensible is seats-as-proxy for workload, and any stack where a line compounds when your volume grows. If the bill under a peak month surprises you, the meter is the problem.