How we calculate, and what we do not know.
The governing metric
Trace Margin evaluates voice agents on cost per successfully completed, policy-compliant task. Total cost includes every attempt, repeat calls, human rescue labour, monitoring, fixed fees and amortised implementation. The denominator is unique successful tasks: calls × (1 − repeat share) × (autonomous + rescued completion share). Failed attempts add cost but never add to the denominator. If no task completes, the result is “No successful outcomes”, never infinity or zero.
Cheap minutes and containment rates are reported only as context. A caller who hangs up is “contained”; a caller told “you’re booked” with no record behind it is “contained”. Neither is a completed task.
Benefits are kept apart
- Cash savings: costs that actually stop.
- Capacity released: staff hours freed, which only become cash if hours are removed or hiring avoided.
- Speculative contribution: margin from recovered jobs, until measured.
Missed-call recovery
We remove duplicates, ineligible enquiries and demand already recovered by callbacks, apply completion and conversion, then cap the result by the jobs your crews can actually take.
Provenance and unknowns
Every input is labelled Provided, Extracted, Assumed or Unknown. Extracted values come from a simulated local parser and must be confirmed before they count. Unknown charges are excluded and the total is shown as a lower bound (≥). Unknown scoring factors add uncertainty, never points.
Ownership route scoring
Each factor scores 0, 0.5 or 1 and is multiplied by its published weight. Scores within ten points produce a conditional comparison.
- tool: bounded task 25, supported integration 25, standard policies 15, limited customisation 15, operations owner ready 20.
- build: engineering capacity 25, on-call capacity 25, custom orchestration need 20, data and integration access 15, budget and time 15.
- expert: complex integration 25, limited delivery capacity 20, managed operations preference 20, funded scope 20, procurement owner 15.
Hard gates override scores. Unsafe scope or a missing fallback owner means Do not automate yet. An unknown fallback, no verifiable outcome or an unresolved outbound authorisation review means Prepare first. Under 150 eligible calls a month, the guide recommends improving routing first. No route supplies legal clearance.
Readiness scoring
Task clarity 30, demand 20, system access 25, exception burden 25. Missing fallback ownership or unsafe scope means Do not automate yet whatever the score; any blocked or unknown gate caps the result at Prepare first.
Generated tests
Trial cases and the dental test pack are synthetic. Every case starts Not executed and stays that way until someone runs it and records evidence. Passing synthetic tests is not certification, compliance or clinical approval.
Interaction conventions we borrow
Our call-review console borrows conventions buyers already know—the dual-track recording view in Vapi call logs, configurable call-history fields in Retell, call-quality views in Twilio Voice Insights and call-to-job linking in ServiceTitan—never their identity or data. Procurement framing draws on Genesys AppFoundry listings, ICMI assessment practice, CCW purchasing context and CX Today research; calling controls on the FCC’s 2024 ruling; accountability on CCXP practice. None of these organisations endorses Trace Margin.
Market context (planning only)
For 2026 planning we estimate US external spending on AI voice agents at $1–3bn, with low confidence. Segment estimates overlap and must not be added together.
| Segment | Estimated revenue | Estimated annual growth, 2026–2028 |
|---|---|---|
| Enterprise automation | $0.6–1.5bn | 25–45% |
| SMB/mid-market | $0.2–0.6bn | 40–70% |
| Developer infrastructure | $0.3–0.8bn | 30–60% |
| Implementation/services | $0.2–0.5bn | 25–50% |
Planning assumptions: $0.08–$0.35 per connected AI minute; implementation $2,000–$15,000 for a constrained SMB scope and $50,000–$250,000+ for enterprise. These are not procurement benchmarks. Expect bundling into suites, verticalisation, price compression, rapid agency entry and disputes over outcome-pricing attribution. Suppliers seek production traffic, pilot conversion and commitments; usage incentives can diverge from fewer, better contacts.
Article selection follows the buyer questions covered in the research library.
Limitations and open questions
- What share of calls is genuinely suitable for automation in each workflow?
- What do contracted costs look like once minimums, concurrency and support are negotiated?
- How much does caller variability change completion between pilots and production?
- How reliably can system outcomes be verified across integrations?
- How often should benchmarks be refreshed as models and prices change?
- Will buyers pay for ongoing assurance after go-live?
Datasets
Examples use three clearly synthetic datasets: Phoenix plumbing dispatch, a 12-location dental scheduling group and after-hours property maintenance. They include deliberate noise: unavailable slots, repeat callers, duplicate writes, unanswered queues and system timeouts. Quotes are anonymous (A/B/C) and are not real offers from any supplier.
Audio and imagery
The call recordings on the home page are synthetic speech generated for this site, and their waveforms are drawn from that audio. Illustrations are illustrative, never evidence.