AI receptionist · self-monitoring, self-repairing
Thousands of small businesses, each with its own agent, its own hours, its own rules. AllHours keeps every one of them good without a human retraining it: each call is scored, failures are clustered and diagnosed, the fix is replayed against real past calls before it ships, and the result is measured in jobs booked that would otherwise have gone to voicemail.
The loop
Most voice-agent products ship a prompt and hope. Ours treats every call as a test case. The six stages below run continuously across the whole fleet, and no stage can widen what an agent is allowed to do.
Transcript, tool calls, latency and outcome are captured per turn. Seven checks run on each call: did it reach a booking or the correct escalation, was intake complete, did it state only real availability and real prices, did it handle interruptions, how long were the pauses, did the caller hang up early, did it follow the business’s policy.
A single bad call is noise. Four calls across three businesses where the caller asked about Saturday and the agent said no is a signal. Failures are grouped by what went wrong, so one fix covers every business with the same problem.
Each turn is linked to the prompt section, knowledge entry or tool result that produced it. The diagnosis names the line, not the vibe: “hours knowledge was stale after the owner edited Saturday hours; the sync job skipped the agent’s cached copy.”
A candidate change is generated against the diagnosis: a prompt revision, a tool fix, a knowledge resync. It is scoped to the businesses in the cluster and may never relax an escalation rule, a booking cap, or a protected window.
The patched agent re-runs the failing calls and a regression set of calls that were already passing, with a simulated caller playing the transcript. It ships only if the failures pass and nothing that passed before now fails.
The fix goes to a tenth of affected agents. If the live score holds for a shift, it is promoted and the owner reads one line in the morning digest: what was wrong, when it was fixed, which calls it touched.
What a score looks like
Scoring is specific enough to be actionable. A failing check points at a turn in the transcript.
One incident, start to finish
Why it matters
The reason to keep agents good is not an abstract score. A small business that misses its after-hours calls loses the jobs to whoever answers. Here is the same fleet, same month, with the quality score and the booked value on one chart.
What is hard about this
Benchmarks don’t have a furnace out at 2 AM. Our scoring runs on production calls, inside the tenant, so recordings and transcripts never leave the business’s account to be graded.
A simulated caller plays the original transcript back to the patched agent, adapting when the agent’s turn changes. It is what lets a fix be proven on 212 calls at 3 AM instead of spot-checked on five.
Every business is its own agent: own hours, intake, knowledge, escalation rules. Clustering failures by shape across the fleet means a problem found at one HVAC shop is fixed at three hundred.
Safety boundary
No automated patch may relax an escalation rule, raise a booking cap, open a protected window, or let the agent quote a price the owner didn’t set. Changes in those classes are proposed to a person and shipped only by hand. A hazard call is escalated the same way on the first day and the thousandth.
Auto-shippable
Human-gated
The demo line at (218) 774-4492 runs the same pipeline as every customer agent. Ask us for the scorecard of your own call afterwards.