A server that goes down announces itself. The site is blank, somebody notices before lunch, and it gets fixed the same day. The failures that cost real money are the quiet ones — the system keeps reporting success while some piece of it has stopped doing its job, and nobody learns until the month's numbers look strange.
The difference between down and wrong
Down is loud. Wrong is quiet, and wrong is where the money goes. An automated text-back that stops delivering doesn't show the owner an error; it logs "sent" and moves on. Almost every quiet failure lives in a seam between two components that each believe they did their job: the messaging platform handed the text to the carrier, and the carrier dropped it; neither considers itself broken. So the first question to ask about any automation is not what it does when it works. It is what it does when it half-works, and which human being finds out, and how soon. The rest of this page is that question applied to the systems we build.
Texts that report as sent and never arrive
Since February 2025, US carriers have blocked business text traffic that is not registered under a verified legal business identity — a regime called A2P 10DLC. The dangerous part is not the paperwork. It is what carriers do with unregistered or sloppily registered traffic: much of it gets filtered silently. The sending platform records the message as sent. The customer receives nothing. Weeks can pass before anyone connects quiet phones to missing texts. There is a second version of the same failure that has nothing to do with registration: landlines cannot receive text messages, and a text to a landline does not bounce — it evaporates. A system that skips the line-type lookup before sending will mark a landline caller as contacted when they never were. How you'd know: compare delivery confirmations against send counts weekly, and once a month, call your own line from an unfamiliar phone and check whether the text actually lands.
The reply nobody read
The most common collapse in customer messaging is not technical at all. The automatic text goes out in fifteen seconds — the part everyone tested — and then the customer replies with a real question, and that reply lands in a web dashboard nobody has opened since training day. The customer waits an hour, then calls a competitor who answered. The design rule that prevents this: replies must arrive where the owner already looks, which usually means their own phone, not a portal with a password nobody remembers. The check that catches it: count the conversations in which the customer spoke last. That number should sit near zero, and it is the single most revealing figure in the whole system.
The answer that was right in 2019
Systems that answer questions from your documents fail silently in their own way: they cannot tell current from stale. A price sheet from 2019 and this year's price sheet cover the same subjects, so a retrieval system sees them as near-twins. Ask a pricing question and it may serve the 2019 numbers with full confidence and a page citation. Procedures are worse. When version 3 and version 7 of a checklist both sit in the index, the system can blend them into instructions that were never true in either version — and a blend is harder to spot than an outright error. The fix is unglamorous: effective dates attached when a document enters the system, superseded versions retired at the door, and one named person who owns the document set the way someone owns the truck inventory.
The scan that inverted a sentence
Paper records enter these systems through optical character recognition, and OCR fails in a uniquely dangerous way: it produces readable text that is wrong. A smudged "not permitted" that comes through as "now permitted" reverses the meaning of a rule, and every downstream component treats the reversal as authoritative — retrieved and cited, page number attached. Photographed paper is worst of all; a manufacturer bulletin shot at an angle in a truck cab can extract into fluent nonsense. The countermeasure is human and cheap: extracted text from scans gets reviewed before it can be cited, and any answer built on a scan should carry the original image so a person can compare the two in ten seconds.
Counting the thing that flatters the system
Vendors of AI phone answering advertise containment — the share of calls that never reached a human. Containment can look excellent while the underlying problem worsens, because a caller who was deflected without being helped calls back two days later, less patient than before. The measure that doesn't flatter anyone is the repeat-contact rate: how many callers come back within about 72 hours on the same issue. The same discipline applies to missed-call recovery. Texts sent is a measure of activity. Jobs booked that trace back to a recovered call is the result, and producing it requires writing every lead into a record with its source attached. A system that cannot produce the second number is asking to be judged on effort.
What checking cannot catch
Monitoring finds the failures you thought to measure. It will not flag a fluent, polite, wrong answer, because wrongness has no error code. The tooling that claims to detect AI mistakes automatically is itself unreliable, so a share of quality control stays stubbornly manual: a person reads a sample of transcripts every week, the way a good shop owner walks the floor. Budget that time or the audit will not happen. And none of this repairs an operational hole. If replies reach the owner's phone and go unanswered anyway, the failure is not in the software, and no dashboard will fix it. A system nobody reviews is not running well. It is running unobserved, which is a different thing.
Where to start
This week, before buying anything from anyone: call your own business after hours from a phone your systems have never seen. Leave the message a customer would leave, reply to whatever comes back, and write down every point where the thread went quiet. That one page of notes — who answered what, and when the silence started — is the honest map of what is already breaking, and it costs one evening.