The agent does not decide what reaches a customer

It answers from the business’s own information, remembers the conversation, and looks up the customer’s record. Then a separate check — plain code, not the AI — decides whether that answer is allowed out, or whether a person takes over.

Working build, 65 of 65 automated checks passing against the deployed code, plus a live run on 1 Oct 2026 that confirmed an empty message never reaches the AI.

6Conditions required before any auto-reply
65/65Automated checks passing on the live code
25Steps in the finished build
0.8Confidence floor, set by the business

How it works

Five steps, from the customer’s question to either an answer or a person’s inbox.

The workflow on the n8n canvas, left to right
  1. A customer asks something

    From a chat, a form or an inbox. An empty message stops here.

  2. The agent looks things up

    It searches the business’s own answers and, when it needs to, the customer’s record.

  3. It drafts an answer

    With the articles it used, how confident it is, and whether any part of the question went unanswered.

    Asked
    “What are your opening hours?”
    Draft
    Answered from the opening-hours article, high confidence
  4. A separate check decides

    Plain code, six conditions, all required. It can only ever hand a conversation to a person — never the other way round.

    Asked
    “I’ve been charged twice… also what are your opening hours?”
    Decision
    A person takes it, even though the answer was right
  5. Sent, or handed to a person

    If it passes, the reply goes out. If not, your team gets the conversation with the reason, the sources and a draft they can edit.

How it’s built

Three decisions, each one the answer to a way this kind of system goes wrong.

The AI is not its own safety net.

The agent is asked to flag anything it should not handle — but the thing that might misread a dispute as a simple question is the AI, so the backstop cannot be the AI too. A separate word list is checked against what the customer actually wrote.

An auto-reply needs all six: the agent did not ask for a human, no escalation word appears in the customer’s message, confidence is at least 0.8, nothing was left unanswered, there is an answer at all, and a policy answer names at least one source.

“Not found” beats a near miss.

The demo information deliberately has holes in it: business invoicing, unusual shipping destinations, warranty length. Those come back as not found, with the topic named, and that overrides any article that happens to look similar.

It is there because of a real bug caught in testing: “can I pay by invoice” matched the card-payments article on the word “pay”, and the agent confidently answered a question nobody had asked. A weak match has to read as a miss.

The handover carries its own context.

A conversation handed over arrives with why it escalated, the customer’s own words, which articles the agent used, how confident it was, what it could not answer, and a draft reply to edit and send. Nobody reads the thread from the top.

Urgency is not left to the AI either: anything money-, legal- or safety-shaped is marked high priority on the words in the message, not on the category the model chose.

The case that makes the point

A customer writes: “I’ve been charged twice for my order — also, what are your opening hours?” The agent answers the opening-hours half correctly, cites the right article and is confident about it. All of that is good work.

It is still handed to a person — because the words “charged twice” are in the message the customer sent.

A check that trusted the AI’s own judgement would have auto-answered the opening-hours part of a chargeback threat and closed the conversation. That is the single worst outcome this build exists to prevent.

What was verified

And, just as importantly, what was not.

The build validated clean: 25 steps, no errors, four expected warnings — two for the switched-off sending steps that stand in for the real ones, two for lookups that take a single plain question, which is correct here.

The check in front of the customer is tested. 65 automated checks pass, run against the code taken straight out of the deployed build, so they cannot drift from what is running. They cover a normal question, an empty message, several at once and a malformed one, plus every one of the six conditions failing on its own, the near-miss bug above, and the structural guarantees — all three outcomes reaching the log, both sending steps switched off.

It ran live on 1 October 2026. An empty message went in; the guard caught it, no AI was called, no reply was drafted, and the conversation was logged as ignored. Counts read from the run itself: every step one in, one out.

Not verified: the five demo conversations have not been run through the live agent. They pass against simulated agent answers, which proves the check, not the AI’s judgement. Driving the real agent needs the server to register a freshly switched-on entry point, and this one does not until its container is restarted — confirmed by building a minimal two-step test from scratch, which failed the same way, while an older registered one still answers. So: the gate around the AI is tested; the AI’s own real-world judgement is not yet, and this page will not claim otherwise.

By design rather than by omission: the information search and the customer lookup are demo versions, not switched-off real ones. A tool attached to an agent still gets called even when switched off, so a disabled connection to a client’s live system would have hit their real data during a demo. The real ones drop into the same slots unchanged.