Dessa

Voice agents are the rare AI category where the value is already measurable before you sell it: call volume, missed-call rate, and handoff accuracy are numbers the client owns. That makes them easier to price than most AI work, and harder to defend once a client learns the inputs. Agencies that build the measurement layer first will own the retainer; agencies that resell a dialer will be compared on per-minute cost alone.

Residual Labor Ledger is a costing discipline: before pricing an AI voice agent retainer, count the human work that survives deployment. Every voice deployment leaves a tail of tasks (transcript review, escalation handling, consent logging, prompt repair after a bad call) and that tail, not the per-minute rate, decides whether the retainer holds margin. The framework asks agencies to log four numbers per client: calls handled without touch, escalations per 100 calls, minutes of human review per call, and cost per resolved call. Trillet's audit-trail design and identity verification reduce the review tail in regulated verticals, while Goodcall-style self-serve builds shift more repair work onto the agency. Forrester's 2027 predictions warn that AI budgets face P&L scrutiny rather than pilot credit, so a retainer priced on usage cost alone will look profitable in month one and thin by month six. Price after the ledger, not before.

Escalation fidelity is the share of calls an AI voice agent routes to the right human, with the right context, at the right moment. It is the number that decides whether an agency voice retainer survives renewal, because a booking agent that handles 80 percent of calls but dumps the remaining 20 percent into a queue without caller history creates more client complaints than a missed call ever did. Measure it before pricing: pull 50 real call recordings, tag every handoff, and count how often the agent recognized the trigger (billing dispute, clinical question, angry caller) and passed a usable summary. Trillet's identity verification and audit trail exist precisely because regulated clients treat a mis-routed handoff as a compliance event, while Goodcall-style inbound capture leans on CRM sync to make the transfer legible. Agencies that quote a flat monthly fee before running this test are pricing blind.

Consent Surface Mapping treats every call recording, transcript, and outbound dial as a regulated artifact before it is a product feature. The framework asks agencies to chart three surfaces per client: what callers are told at the start of a call, what the voice system stores afterward, and which jurisdictions the calls touch. Each surface caps what the agent is allowed to do, so scope decisions follow disclosure rules rather than the reverse. This matters because voice deployments fail audits on recording consent and outbound dialing rules far more often than on latency or voice quality. A concrete case: Trillet verifies caller identity to compliance standards and keeps a full audit trail, which suits healthcare and legal retainers where an unlogged action is itself the liability. Agencies that map consent first can price a retainer against a defined scope; those that skip it discover the ceiling after the client's counsel reads the call flow.

Do not quote a retainer until you have scored 50 to 100 real call recordings for escalation accuracy, latency, and post-call human minutes, then priced the labor that remains.

Run at least 50 real client call recordings through the candidate agent and measure escalation accuracy, latency, and consent handling before you quote a retainer.