Six layers. One governed GenAI platform.

The interesting thing about large language models, from where we sit, is not that they can talk. It is that talking is only the last five percent of what needs to happen for a conversation to matter in healthcare.

The other ninety-five percent — understanding intent, reasoning under policy, calling the right systems, staying inside compliance, and closing the loop on outcomes — is what actually determines whether a patient's care moves forward. When we set out to build Interactly, we didn't want a chatbot with a phone number attached. We wanted a stack where every one of those layers was a first-class part of the platform.

We call it the six-layer architecture. Nothing about the number is sacred; it just happens to be the smallest set of layers where each one earns its keep and none of them collapses into the next.

Layer one: interaction

The interaction layer is what the patient actually experiences. Voice, SMS, chat, email, and — because healthcare still runs on it — fax. What matters here is not that we support these channels individually, but that a single conversation can begin on one and continue on another without losing state. A patient who receives a text about a scheduling conflict at 8:14 in the morning should be able to answer a callback at 8:47 and pick up mid-sentence.

Getting this right is boring engineering. Session identifiers, retry logic, callback windows, and quiet-hours rules. But if the interaction layer leaks, no amount of cleverness above it can recover.

Layer two: understanding

Above interaction sits understanding — the layer that turns a stream of tokens into structured intent. In healthcare, this is harder than it looks. The same sentence can mean "I want to reschedule," "I want to cancel," or "I want to complain, but I'm being polite about it," and the right downstream action for each is different.

Our understanding layer produces two things for every turn of conversation: a normalized intent with confidence, and a set of extracted entities tied to the patient's record. When either is uncertain, the system asks — it does not guess. Guessing is what makes voice bots feel like voice bots.

Layer three: reasoning

The reasoning layer is where policy meets the model. It decides what to do next given the intent, the patient's history, the caller's authority, the time of day, and the workflow's rules. Sometimes that means answering directly. Sometimes it means opening a ticket. Sometimes it means transferring to a nurse triage line because the patient just described symptoms that fall inside a clinical escalation trigger.

"A model that reasons without policy is a liability. A policy engine that runs without a model is a decision tree. The reasoning layer is where those two things become one thing. — from an internal design doc, 2024

Reasoning is also where we do most of our tool selection. The model does not have unbounded access to the world — it can call a defined set of actions, each one with a schema, a scope, and an audit trail. That constraint is a feature, not a limitation.

Layer four: action

This is the layer that most conversational-AI stacks leave out, and it is where Interactly earns its keep. The action layer writes back to the systems of record — Epic, Cerner, eClinicalWorks, Salesforce Health Cloud, Availity, Change Healthcare, and whatever telephony platform is carrying the call.

If the referral isn't closed in the EHR, it isn't closed. If the prior auth isn't submitted through the payer portal, it isn't submitted. The action layer treats every conversation as an atomic unit of work: either the outcome landed in the system of record, or it didn't, and we can prove which one.

Layer five: governance

Governance runs alongside every other layer, not on top of them. It's the layer that redacts PHI before it leaves the runtime, checks that the caller is who they say they are, records the consent that lets the conversation continue, and — when something goes wrong — surfaces a trace that a compliance officer can actually read.

Healthcare AI without governance is a demo. Healthcare AI with governance added later is a rewrite waiting to happen. We built this layer first and pointed the other five at it, which sounds obvious in retrospect and cost us a year to internalize.

Layer six: outcomes

The final layer closes the loop. Every conversation produces a set of outcome signals: was the appointment kept, did the care gap close, did the referral get scheduled inside the SLA. Those signals feed back into understanding and reasoning, so the platform gets sharper on your specific workflows over time — not on a generic benchmark.

This is also the layer that pricing rides on. We charge per completed care action, not per minute or per call. That single decision, more than any model choice, is what forces the rest of the stack to be honest.

Why six, and not five, or seven

Whenever we've tried to collapse two of these layers, one of them starts to leak into the other's job. Understanding starts making policy decisions. Action starts inferring intent. Governance becomes a review meeting instead of a runtime.

Whenever we've tried to split a layer, we've ended up with a second layer that never quite justified its own interface. Six is the smallest number that lets each layer be replaced independently without anything else breaking — which is the only definition of "modular" that matters when you're operating a healthcare platform in production.

‍