Tiered approval: the autonomy dial nobody talks about

Full automation sounds like the goal. One system, no friction, everything handled. And when it works, it works beautifully.

But there's a version of that story nobody tells — where the AI sends a half-finished response to a high-value client, or fires off a reply to a complaint before anyone had a chance to read it. The system performed exactly as designed. The outcome was still a problem.

At Fortis, the way we avoid this isn't by limiting what the AI can do. It's by being deliberate about what it's allowed to do without asking first. We call this tiered approval, and it's one of the most important structural decisions we make on every build.

When we built the AI operations system for JMJ Homes — a construction and development business with a high volume of inbound comms across email, text, and enquiry forms — the first thing we mapped wasn't the workflows. It was the risk profile of every message type.

Not all messages are equal. A subcontractor confirming a delivery time is not the same as a lead enquiring about a $2M knock-down-rebuild. An admin question is not the same as a client flagging dissatisfaction. Treating them identically is where automation earns a bad reputation.

So we built three tiers.

The first tier is full autonomy. Routine, low-stakes messages — general enquiries, scheduling confirmations, simple information requests — get classified, drafted, and sent without any human in the loop. The AI handles it start to finish. This is the volume layer, and it's where most inbound messages actually land.

The second tier is draft and review. When the AI identifies a message as higher significance — a new lead, a decision that needs to be made, something with commercial weight — it drafts a reply and holds it. The draft surfaces for a human to read, approve, and push. The AI did the work. The human makes the call to send.

The third tier is flag before drafting. For anything the classifier reads as sensitive — a dissatisfied client, an overdue invoice alert, something that requires context the AI doesn't have — it doesn't draft at all. It flags the message, summarises what came in, and waits. The human decides how to respond. The AI is in listening mode, not action mode.

The temptation when deploying any new AI system is to unlock everything immediately. The demos work. The logic holds. Why slow it down?

Because trust between a human team and an AI agent is earned through repetition, not promised upfront. The first weeks of a deployment are when you discover the edge cases your discovery process didn't surface. The caller who phrases a routine question in a way that sounds like a complaint. The message that looks like a lead but is actually a follow-up on an existing project. The client whose relationship history means they need to hear from the principal, not an automated reply.

At launch with JMJ Homes, we set the tier thresholds conservatively. More messages sat in the review tier than strictly needed to. The team saw more drafts than they'd eventually want to approve. That was intentional.

What it created was a feedback loop. Every time a draft was approved unchanged, that was a data point. Every time a human edited before sending, that was a signal — about tone, about threshold calibration, about a category of message the AI was misreading. Over time, the tiers shifted. Things that started in review moved to auto. Things that were flagged too aggressively got recalibrated.

The system didn't get smarter on its own. It got smarter because humans were watching it closely at the start, and the architecture made that watching easy.

The AI reads every incoming message before it routes it. It's not keyword matching. It's reading for intent, urgency, and category — understanding whether a message is a new lead or a returning client, whether the tone signals satisfaction or frustration, whether the content involves a financial matter, a time-sensitive request, or something that requires a judgement call.

The categories we mapped for JMJ Homes included lead enquiries, subcontractor communications, project updates, payment-related messages, and client dissatisfaction signals. Each category had a default tier assignment. Some had override conditions — a payment message from a new contact goes to tier two; from an established client in good standing, it might resolve in tier one.

That mapping came entirely from discovery. We spent the first days of the build asking questions that felt peripheral: what does a difficult message look like for your team? Who should never get an automated reply? What's the kind of email that, if handled badly, could cost you a client relationship? The answers to those questions are what the tier thresholds are built from.

What tiered approval really gives you is a dial. At one end, everything requires human approval — the AI is just a drafting tool. At the other end, everything is automated — the AI is running operations. Neither extreme is right at launch.

The goal is to start closer to the conservative end and move the dial deliberately as trust is established. Not based on a timeline, but based on evidence. You move tier two items to tier one when you've seen enough approved drafts to know the AI is getting it right. You tighten tier three when you've calibrated what "sensitive" actually means for your business.

This is also why clients who understand the system tend to stick with it. The retainer model we run means we're watching the system alongside them. When something should move tiers, we adjust it. When a new message category appears that doesn't fit the existing map, we handle it. The dial never gets set and forgotten.

Full automation is still the goal. It's just a destination you arrive at through earned trust — not a setting you flip on day one.

If you're evaluating AI ops for your business and want to understand how tiered approval would apply to your workflows, book a discovery call. We'll map your message types and show you where the dial should sit at launch.