Why voice-AI agency margins die in month three.

The mistake is hiding in plain sight in the original pitch. You quoted the client a flat per-minute rate — let's say thirty cents — because that's what their bookkeeper wants on the invoice. Underneath, your infrastructure is metered: per-token LLM, per-minute voice engine, per-minute telephony. So your cost moves at call time, but your price doesn't. Day one, that's fine; you priced with a comfortable spread. Day ninety, you're staring at a margin alert because the agent took a 14-minute call from a confused customer who wanted to dictate their entire HVAC service history.

The three things that always go wrong

If you've already shipped one voice-AI client, you can probably name them. If you haven't, here they are:

1. Call duration drifts up. Once your agent starts handling real volume, average call length goes from your 90-second pilot to 4–6 minutes. The agent gets more conversational. Callers get more comfortable. The economics quietly degrade.
2. Token footprint grows. You add a knowledge base entry. The client wants the agent to handle three new objection patterns. The system prompt gets longer because you're chasing edge cases. Every one of those tokens is multiplied by every LLM call per minute.
3. One caller blows the average. A robocall, a confused customer, a competitor probing your agent — one minute can land at four to five times your average per-minute cost on a metered plan. You're absorbing it because you already quoted the client a flat rate.

Each of these on its own is survivable. Stacked, they're the reason most voice-AI agencies have a "what just happened" conversation with their accountant in month three.

What the math actually looks like

Let's run a typical small client. 1,500 minutes per month of voice agent activity. You quoted $0.30/min, so you're billing them $450/mo. Your published infrastructure cost on a metered platform, at the time you signed, looked like $0.15/min — $225/mo — so you projected a $225 margin per account. Beautiful.

Now month three hits. The agent's average per-minute cost has crept to $0.21 because the knowledge base grew and call durations lengthened. That's $315/mo of cost on a $450/mo invoice. Margin: $135. You also have two heavy days where a robocaller and a confused upsell loop pushed effective rates north of $0.40/min for a few hours. That month, you eat $50 of unexpected cost. Net margin on that one account: $85.

Stack twelve clients like that. You projected $2,700/mo of voice margin from the cohort. You're banking $1,020. You're still profitable, but you priced a $32K/yr line item like it was a $32K/yr line item, and now it's $12K. That delta is the gap that breaks agency hiring plans.

The structural fix isn't "negotiate better with the platform"

The platform's economics aren't broken. Metered pricing is rational from their side — they pay LLM vendors per token. The problem is that you, the agency, are sitting between a metered cost and a flat-priced revenue line, and that's a structurally fragile position.

The fix is to pull the LLM cost into a flat per-minute line on your side too. Callibre's pricing exists because of this exact pattern. Voice AI is a flat $0.07/min. The LLM is a flat per-minute line item — you pick the model at build time ($0.006/min for Gemini Flash, $0.025 for Claude Haiku, $0.060 for Sonnet 4.6) and the rate is locked. Telephony is pass-through at $0.010/min, no markup. Your cost moves only if you change the agent, not if the caller does.

What that gives the agency: a margin floor you can defend in a kickoff meeting. If you sell at $0.30/min and your cost is $0.105/min (Haiku default), your margin is $0.195/min regardless of who calls, how long they talk, or what the agent does inside the minute. Multiply that by 1,500 minutes and you're banking $293/mo per account — and that number doesn't move when the knowledge base grows or when a robocaller hits.

The retention follow-on

The other thing that dies in month three is your client's trust. They priced their service into the market based on the rate you quoted. If you go back to them with "we need to renegotiate because our costs are higher than projected", you've turned a profitable account into a difficult conversation. The agency that priced flat and held it for the full term is the agency they renew.

You don't have to take this as gospel — drop your real numbers into the calculator and see what the spread looks like at your client's volume. The point isn't that one platform is good and one is bad. The point is that flat-cost infrastructure is a different posture for an agency to operate from, and most agencies don't realize the difference matters until they're already locked into the other side of it.