You pay for conversations.
That’s the whole meter.

How the meter works

One shape, both sides of the market: a fixed amount each month, plus an amount for each conversation you run. Nothing is charged up front — we bill at the end of the month for what actually happened during it.

The per-conversation rate is graduated: it steps down as your volume rises, and each band prices only the conversations that fall inside it. If the rate drops after the first thousand conversations, your first thousand still price at the first rate and only the ones above it price at the lower one. That’s deliberate. The alternative — one cheaper rate applied to everything once you cross a line — produces a bill that goes down when your usage goes up, and a bill nobody can check is a bill nobody should sign.

How many conversations does an evaluation actually take?

This is the question the plan cards can’t answer for you, and without it the rate above is unusable. A run is three numbers multiplied together — how many vendors, how many situations, how many times each — so here are the sizes real evaluations come out at:

• A quick first look: 2 vendors × 5 situations × 3 repeats = 30 conversations. Enough to see which one falls over, and honest about the rest — at that size a grade usually publishes as a range rather than a single letter.
• A typical shortlist run: 3 vendors × 8 situations × 5 repeats = 120 conversations. This is where most buyers start, and it is normally enough to separate a good agent from a bad one.
• A decision you’re putting a year of budget behind: 3 vendors × 12 situations × 10 repeats = 360 conversations. This is the size at which two close vendors usually stop being “too close to call”.

The repeats are not padding. One call tells you what happened once; five tell you whether it happens reliably. And the number of repeats is exactly what decides whether we can honestly name a winner at the end or have to tell you the two vendors are indistinguishable — which is why we’d rather you ran eight situations five times than forty situations once.

Multiply your own three numbers, then read the rate off the plan cards above. We deliberately don’t print a total on this page: the plans are served live from our billing system, and a number typed into the page would drift from the one on your invoice.

Working out what the vendors will charge you is a different exercise, and there’s a page for it: how much an AI receptionist costs.

What counts as a billable conversation

One conversation with a vendor’s agent — one phone call, one chat session, one email thread, one SMS exchange — that ran to a real ending and produced a result we could score. That’s the unit, and it’s the same unit whether the result was good or bad for the vendor: a conversation where the agent failed the task still cost us a call and still tells you something, so it bills.

Three things are not conversations and never bill: a run we place for our own public grading, an attempt that never reached the vendor at all, and — the one worth reading twice below — anything our own platform broke.

If our own equipment broke the call, you don’t pay for it

Every test run depends on our own equipment: the phone system, the synthetic caller, the connection that carries the audio. When that equipment is what failed, the conversation is marked as our fault, it is excluded from the results, and it is never billed to anyone — not to the buyer who commissioned it, not to the vendor being tested.

The same branch in our code does both jobs, and that is the whole point. A platform fault that quietly billed a customer would be a platform fault that quietly counted against a vendor’s grade. We refuse to charge for our own mistakes for exactly the reason we refuse to publish them as somebody else’s.

A card before the first billable test

We ask for a card on file before the first conversation that can bill — even when the fixed monthly amount on your plan is zero. Billing in arrears without a card means discovering at month end that a month of testing was never collectable, and the version of that story where we chase you for it is worse for both of us than the version where we ask first.

Some plans carry a monthly conversation cap. Where one does, we refuse to start a run that would exceed it, before any calls are placed — you are never billed for an overage you couldn’t see coming.

Changing or cancelling

Cancel whenever you like; it takes effect on your plan, not on your history. Conversations that already ran still bill. You’ll receive one final invoice covering the tests placed before you cancelled, and nothing after it.

The price you agreed to when a run started is the price that run is billed at. If we publish new plan pricing while your test is in flight, your test keeps the tariff you approved — a new number can never reach an invoice without having passed a preview and your explicit go-ahead first.