We published a post on 2026-07-23 that named exactly one open gate. Three days later that gate was gone — and the sentence was still on our site.
The post was our write-up of an AI agent built to sit on insurance carrier hold queues for a storm-restoration contractor. It carried this claim: counsel sign-off on the consent language is the only go-live gate still open. On 2026-07-26 the counsel gate was removed, and the published sentence stayed on the page describing a world that no longer existed. We are correcting it in public rather than by silent edit, because a claim that was true when written and false three days later is exactly what an agency quietly fixes and nobody hears about.
The correction is the smaller half. The larger half is a buyer's guide, and it earns its place for one reason: our own build failed two of the acceptance tests we are about to hand you. A question only a vendor's marketing can pass is not a question. These have already caught us.
The correction, stated plainly
The counsel gate was removed on 2026-07-26. That did not mean the consent question stopped mattering. It means the requirement stopped being a sign-off and became code.
The vendor's interface exposes no opening-message field for custom-model agents — no box in which to type the sentence a caller must hear first. So the deterministic disclosure moved into the bridge and is spoken as proactive response zero: emitted before any model output, on every call, without the model being consulted about whether to say it. A dedicated test pins that ordering, because the value of a deterministic disclosure is that it cannot be skipped by a model having an unusual day. That is a better outcome than a sign-off, and a harder thing to have claimed in a blog post.
Two more corrections on the same build
Bilingual mirroring was aspirational when first written down. An earlier ledger entry described the disclosure as mirrored across both languages; the persona in fact carried only a language rule — respond in the caller's language — and no Spanish disclosure text existed until a later session wrote one. A language rule is not a script.
An English 911 instruction described as "proven" was retracted to unverified, because the vendor setting its proof depended on was later disabled. It was published as a regression rather than left standing as passed. A claim whose evidence has been withdrawn reverts to unverified.
Build It Dormant
Four gates, one rule: fail a gate and the system stays dormant. Not throttled, not "in beta," not "soft-launched to a small group." Built, tested, unplugged.
Gate 1 — A red-team harness against the real persona, with zero dials
Seven adversarial scenarios: hold-music pitch, DTMF tones, pressure against the appointment rules, social engineering for identifiers, a do-not-call request, a Spanish-speaking representative, and prompt injection. All seven run against the production persona and place zero dials.
Worked example. The hold-music scenario produced the least obvious engineering finding in the build. The agent had been responding to hold music: the audio pipeline saw sound, treated it as a turn, and the model answered it. The bridge now keeps a local call-state mirror and emits nothing unless that state reads "with human," and it no longer bumps the response identifier, which had been cancelling in-flight turns. In the ledger's words: the agent can no longer talk to hold music.
Gate 2 — A legal and consent gate that survives the model
The disclosure is response zero, spoken before any model output, pinned by a test. It is a gate rather than a checklist item because every alternative makes the disclosure conditional on model behaviour.
Worked example. With no opening-message field available, the options were: instruct the model to open with the sentence, or emit the sentence outside the model. The first is a prompt, and a prompt is a request. The second is a guarantee. Only the second satisfies the gate — which is why the counsel sign-off could be retired without the consent question loosening at all.
Gate 3 — A kill switch with a human fallback, failing closed
Every feature flag ships false, and the code fails closed: only an explicit true or 1 enables anything. An unreadable config, a typo, a missing environment variable and a deploy race all resolve to off.
Worked example, and it is the gate nobody thinks of. Before the flag can be flipped, the queued-call count must equal the intended pilot set — three or fewer. The sweep dials every queued row the moment the flag flips, so a stray row left from a test three weeks earlier is a stray live call to a real person. "The system is off" and "the input queue is empty" are two different facts, and only one shows on a dashboard.
The human fallback carries a rule appearing in four separate documents across this work: an AI answer never counts as a live answer. If the fallback is the agent, there is no fallback.
Gate 4 — Vendor redundancy, including redundancy against the vendor's own claims
Worked example. The vendor's AI product has no configuration API. No endpoint returns current settings, so there is no way to diff a configuration, restore one programmatically, or detect drift. The consequence, recorded verbatim: screenshots are the only rollback baseline that will ever exist for it. Not "for now" — that is the permanent condition of that surface until the vendor ships an API, and a build has to be designed around it rather than waiting.
The second half of this gate is that a vendor's written statement is not verification. That one cost us a reopened item.
The two we failed, start to finish
Before a single real caller could reach the system built for a commercial property-and-casualty brokerage, fourteen scripted acceptance calls were placed against it by hand, each against a written scenario with a stated pass condition. Twelve passed.
A10 — Spanish emergency — FAILED. The scripted caller said "Tengo una emergencia." The system returned an offer to transfer to reception, with no 911 instruction in Spanish. A mitigation was attempted in the same session and the retest also failed. The item was escalated to a vendor case.
A12 — force-fail — FAILED. A caller reaching a disabled extension heard "that extension is not recognized" followed by a busy tone: a dead end, with no voicemail chain behind it.
Both were published as failures on the client-visible dashboard rather than fixed quietly. Go-live proceeded as a recorded, accepted-risk carve-out, with the matrix explicitly not declared passed. "We went live with two known open items, named, in front of the client" and "we went live" are different sentences, and only the first survives being read back six months later.
Then A10 was closed on the vendor's word — and that was wrong. The vendor supplied a written assertion that the behaviour was corrected. The item was accepted on that assertion, with verification waived, and — the part that made recovery possible — the waiver was recorded at the time as not meeting the standing closure bar. The shortcut was written down as a shortcut.
So it was reopened. Against the house bar, a vendor assertion is not evidence about a live system; it is evidence about what the vendor believes. A10 was reopened, tested again, and closed on live proof, with the earlier waiver explicitly superseded rather than left sitting underneath the closure as its apparent reason.
A12 closed differently. It was ruled a consequence of a routing loop, structurally closed on 2026-08-10 — a better class of fix than patching the symptom, because it removes the dead end for every extension rather than the one the test happened to dial.
Failed → published as failed → briefly closed on a vendor's word → reopened against the house bar → closed on proof. Five steps where most write-ups have one, and the middle three are visible only because the waiver was logged instead of assumed.
Three vendor-surface findings that will cost you a day each
The dialog that discards what you just did. The vendor console's exclusion-list dialog silently discards a completed selection if dismissed without saving — proven directly: seventeen selections, dismissed, re-opened, zero. No warning, no confirmation, no draft state. The house rule that followed is stated as a negative because that is how it gets remembered: never screenshot the dialog you just filled in. Screenshot the saved state, from a re-opened dialog, after the write. A screenshot of your own unsaved input is a picture of your intentions.
Fourteen "per-person" queues that do not contain the people they are named after. A queue named for the owner's service line contains one agent, who is not the owner. Another contains two different people. No owner appears in any of the fourteen. Every routing decision and every conversation about "his queue" had been running on a name that described nothing.
The buyer's guide
The best question in this category is already in the post this one corrects, and it has not been improved on:
Ask whoever is pitching you an AI caller: "Say the exact sentence your agent uses to disclose recording — in both languages my customers speak." Then count the seconds of silence.
The acceptance matrix supplies the rest. Each is a question we can answer only because we were on the wrong end of it first.
- What happens when a caller says they have an emergency — in each language you serve? Get the exact wording, then ask whether it is a scripted string or model output. Model output is not an instruction; it is a tendency.
- What happens when a caller reaches a disabled or misconfigured extension? A busy tone with no voicemail chain behind it is a dead end no dashboard reports, because from the system's side the call was handled.
- How many scripted acceptance calls run before the first real caller reaches this — and what is the pass bar?
- Where do failed acceptance items get published? If the answer is "internally," the buyer is not on the distribution list for their own system's known defects.
- Does a vendor's written assertion close an item in your process? The correct answer is no; the second-best is "only with the waiver recorded as a waiver."
- What is the rollback baseline for each vendor surface — data or screenshots — and when was it last refreshed?
- What is in the input queue right now, and what happens to every row the moment the system is enabled?
- Does the assistant identify as automated? It should, on every call, and it should never be given a human name or allowed to claim personhood.
What we will not claim
We do not publish numbers we have not measured, and this build produced several numbers people would like us to publish. So, explicitly:
- No hours-saved figure and no outcome claim. Our internal model defines time saved as thirty minutes multiplied by human-reached calls — a baseline assumption written down before the build, not a measurement. Multiplying an assumption by a real count produces an assumption with a decimal point on it.
- Hold survival above 85–90% is a target inside an acceptance gate, not a result. Quoting the bar as achieved performance would be quoting our own homework as our own grade.
- An AI answer never counts as a live answer. Any answer-rate figure folding AI-handled calls into "answered" measures something other than what the client cares about.
- We do not publish a concurrency ceiling for this vendor. A figure circulated for one; fetched directly, the cited source contained no such claim. It was a summarisation artifact, and no first-party ceiling exists.
What would prove this wrong
- A live-proof retest of the Spanish emergency case failing after we recorded it closed. The closure would be wrong and the house bar would not have done its job. We would publish that as we published the original failure.
- Vendor written assertions matching live behaviour every time, across many items. Then reopening A10 cost time and produced nothing, and the bar is over-engineered rather than load-bearing.
- A real call in which model output precedes the disclosure. Response zero claims that cannot happen; one recording falsifies the gate.
- A queued row that does not dial when the flag flips. Then the pre-flip count gate guards against nothing, and it is ceremony.
- A measured hold-survival read below the gate's target after the call-state mirror shipped. We have not published one, because we have not taken one.
Dormant is the hard part
An AI voice agent is not hard to build and is very hard to build dormant — adversarially tested, every flag false, the disclosure unskippable, the queue empty, and a human as the fallback rather than another model. A system that fails one of the four gates stays unplugged until it does not. If you are being pitched one, start with the disclosure sentence in both languages and work down the list above. If you would rather have someone go through it with you, we will.
Get the kit, not just the theory.
We'll send the build checklist behind this post — and the next pillar when it ships. One email, no drip sequence. Unsubscribe in one click.