"Full automation" is the pitch everyone leads with. Let the system handle everything. No humans in the loop. Lights-out operation. It's a compelling idea and it's mostly wrong — at least for the first few months of a build.
The JMJ Homes build was the clearest lesson we've had on this. James runs a growing residential construction company. He wanted to automate his inbox, his DMs, and his CRM logging. The temptation was to set everything to auto and let it run. We didn't.
Full auto sounds great until the bot books your day off. Or sends a price quote to a competitor who emailed in.
Instead, we built a three-tier approval system. And it changed how we approach every operations build since.
The three tiers
The routing decision — which tier an email or message falls into — is made by the AI based on a classification prompt. It considers the sender, the content, the implied action, and a set of rules we defined during discovery. James added his own rules: his top five clients always go to tier 3, regardless of content. Any email mentioning a specific subcontractor goes to tier 2.
Why starting conservative matters
We deliberately launched JMJ Homes at tier 1 for most actions. Not because the system wasn't capable of more — it was — but because trust in automation is earned, not assumed.
In the first two weeks, James reviewed everything the system wanted to send. He edited a handful of things. He flagged two cases where the classification was wrong — a legal enquiry that the system had rated tier 2 instead of tier 3. We updated the rules.
By week four, he was barely looking at tier 1 actions. By week eight, he'd moved several categories from tier 2 to tier 1 because the drafts were consistently good enough to send as-is.
That progression — conservative launch, earned autonomy, gradual expansion — is the right way to do this. Launching everything at full auto means you find the edge cases the hard way, when they've already sent something wrong.
The Telegram interface
The entire approval flow runs through Telegram. A tier 2 action sends James a message with the full draft, the context (who the email is from, what they asked, what the system thinks the right response is), and two buttons: Approve or Edit.
If he taps Approve, it sends. If he taps Edit, it opens a thread where he can rewrite the message before it goes. If he does nothing — he's on a job site, phone's in his pocket — it waits. It doesn't send. It doesn't expire and send anyway. It waits until he acts.
That last part is important. A lot of automation tools have "auto-approve after N hours" settings. We don't use them. An action that's been sitting in approval for six hours probably needs to wait for the right person — not send itself because a timer ran out.
Adjusting the dial over time
The tier system isn't static. We review it monthly with clients in the first three months, quarterly after that. As the system proves itself reliable in a category, the threshold moves. As the business changes — new team members, new services, new sensitivities — the rules update.
The goal is to get to a point where the vast majority of routine work runs on auto, and the human's attention is reserved for what actually needs it. Not everything — just the things where judgment, relationship, or risk genuinely requires a person.
Full automation is a destination, not a starting point. The dial gets you there safely.