When a vendor tells you their AI answers calls in under 50 milliseconds, they're quoting a number you'll never experience. What you experience on a phone call is a chain of delays: the time to connect, the time to notice you stopped talking, the time to think, the time to say the first useful word. A single headline figure collapses all of that into one impressive-sounding measurement, and it tells you almost nothing about whether the caller will stay on the line. If you run a plumbing company, a dental practice, or a law office, and you're shopping for AI lead response, you need to know how to pull that claim apart. Here's how.
Key points
- One latency number hides at least seven separate delays, and callers feel each one differently.
- A greeting, a price question, and a booking confirmation need different speed budgets, so they shouldn't share one target.
- The FTC expects companies to back up claims that an AI system beats a non-AI alternative, so vague speed claims carry real risk.
- Four accounting illusions inflate AI ROI projections: treating saved time as saved cash, fictional baselines, narrow cost counts, and no counterfactual.
- Buyers should ask vendors for a per-stage latency budget, a defined evaluation window, and a kill rule before signing anything.
A phone conversation with an AI agent runs on at least seven separate clocks. The first is call setup: the time from the moment the system starts to the moment two-way audio actually flows. Carrier routing lives here, and it's outside your vendor's control, which matters when you're deciding who to blame. The second clock is end-of-turn detection. When you pause mid-sentence, the system has to guess whether you finished. Silence is ambiguous. You might be checking a document, catching your breath, or hesitating. A crude silence threshold either cuts you off or leaves dead air, and both feel terrible. The third clock is response planning: intent recognition, looking up the answer, applying rules, booking the appointment. This is where tool calls often dominate, even when the underlying model is fast. A system that can generate text instantly can still stall waiting on your calendar. The fourth clock is time to first intelligible audio. The caller doesn't care about the first token of text the model produces. They care about hearing a useful word. The fifth is turn completion: the full answer, and for a booking, confirmation that the appointment actually stuck in the calendar. The sixth clock is interruption handling. When the caller talks over the agent, how fast does the agent stop? Does the caller's speech get preserved? If the agent was mid-booking, does the unconfirmed action get cancelled safely? The seventh is recovery: what happens when the transcription mishears, a lookup times out, or the voice fails. A good system clarifies, retries, or transfers to a human. A bad one loops or hangs up. If your vendor reports one latency number, ask for the other six.
A greeting, a factual question, an appointment lookup, an address correction, and a booking confirmation don't share the same dependency path. A greeting needs nothing but a fast hello. A booking needs a calendar lookup, a slot check, and a confirmation. Treating them all under one latency target means you're either over-promising on the hard turns or under-delivering on the easy ones. The practical fix is a latency budget per turn class. For each type of exchange, the vendor should be able to tell you the target time and what happens when it's missed. The measurement boundary should also be honest: it starts at the telephony edge the vendor controls and ends at audio delivered back to that edge. Carrier latency gets reported separately, not mixed into the vendor's own numbers to make them look better or worse. This is why we treat our own promise, first reply in seconds, as a budget rather than a benchmark. The missed-call flow, where a text-back goes out the moment a call is missed, has one budget. The after-hours answering flow, where a caller asks questions and books, has another. Conflating them would be convenient and misleading.
There's a school of thought that says the goal of a voice agent is to sound human. We think that's backwards, and the latency framework explains why. If a caller spends the first thirty seconds testing whether they're talking to a person, those are thirty seconds of wasted turns. The caller interrupts less freely, repeats themselves, and structures sentences to trip up the machine. Disclosure shortens the path to the actual task. When callers know up front that they're talking to AI, and know they can reach a human, text, or book online at any point, they get to the point faster. The call becomes more interruptible and more predictable, which are latency outcomes, not just ethics outcomes. And when something goes wrong, the recovery path is explicit instead of hidden behind a bluff. There's a regulatory angle too. The FTC has advised businesses to substantiate comparative claims about AI products, including that an AI system actually outperforms a non-AI alternative. A vendor claiming their agent is indistinguishable from a human receptionist is making a claim they probably can't back up. A vendor claiming the caller always knows it's AI, and can always reach a human, is making a claim you can verify with one test call.
Speed is only half the audit. The other half is the money, and this is where most AI pitch decks fall apart. There are four accounting illusions that show up again and again. First, time saved gets entered as cash saved. If an AI receptionist saves your office manager twenty hours a week, that's capacity, not money. It only becomes money when you make an operating decision: fewer hours paid, a hire you don't make, more billable work, or more booked jobs. Until that decision happens, the savings are theoretical. Second, the baseline is fictional. Teams estimate how slow the old process was from memory, usually after seeing the shiny new pilot. Memory weights the worst cases. Rework, seasonality, and queue time get mixed with actual touch time, and suddenly the old process looks terrible. Third, the denominator is too narrow. The spreadsheet shows the software fee. It doesn't show data prep, integration, testing, training, monitoring, incident handling, or the cost of unwinding the whole thing if it fails. For a small business with no IT staff, that hidden work is often the biggest line item. Fourth, improvement gets attributed to AI with no counterfactual. If you launched the AI agent the same month you changed your ad spend and retrained your sales staff, you can't tell which one moved the number. A real ROI case names the baseline, prices the human work around the tool, and separates what was measured from what was assumed. If the sheet doesn't show those three things, it's a pitch, not an ROI.
The caution isn't paranoia. The OECD's 2025 report on AI adoption among small and mid-sized firms found real links between AI use and productivity, but the advantages shrank once you controlled for firm characteristics and supporting capabilities, and benefits often follow a J-shaped path: things get worse before they get better, while the team learns new habits. PwC's 2026 Swiss CEO survey found 67% of companies had seen no AI upside yet, and only 5% reported both cost savings and revenue gains. Meanwhile, SMEs make up 99.7% of Swiss companies and about two-thirds of employment, which means the binding constraint for most businesses isn't capital. It's management attention. You can't run five pilots at once when you're also the person quoting the HVAC jobs. That's the argument for a kill rule: decide, before you start, what result would make you stop. One process boundary, one owner, one evaluation window, one observable outcome. A month of calls answered, missed calls recovered, and appointments booked is a measurable unit. 'AI transformation' is not.
The staged-evidence logic has a practical consequence for contracts. If a vendor needs a twelve-month commitment to make their numbers work, their numbers probably don't work. A service that actually captures leads should be able to prove itself inside a month of live calls. That's why the structure matters as much as the technology: monthly billing, no contract, cancel anytime, a scoping call before you pay anything, and pricing where you only pay for minutes that actually handled a lead. Spam and robocalls screened out of the bill is the kind of detail that tells you whether a vendor has thought about your side of the ledger. Ask any prospective vendor three questions. What's your latency budget per turn class, and what happens when you miss it? What's the baseline you're measuring against, and how was it measured? And what would I have to see in the first month to keep paying you? If you get straight answers to all three, you're dealing with someone who has done the math. If you don't, you're the one running the pilot on their spreadsheet.
The speed claim on an AI answering service is easy to make and hard to check, until you know what to ask for. Seven clocks instead of one number. A budget per turn class instead of a single benchmark. A baseline, a full cost count, and a counterfactual instead of a projected ROI. None of this requires you to become a technical buyer. It just requires you to stop accepting one impressive number where seven honest ones belong. The vendor that publishes its budgets and its baseline is the one you can hold to the promise, and the one whose promise is probably worth holding.