Academy

Layered intelligence: when conversational AI meets background AI

A good agent has to do two things well at once: fluent conversation with a human, and deep cognitive work behind the scenes. These are two different layers of intelligence — and we believe separating them is the key.

The Felesh team 5 min

When we think about an intelligent assistant, we usually picture something that talks to us. But the real work is rarely just talking. Behind every fluent conversation, deep cognitive work has to happen: understanding intent, extracting details, and turning a vague request into a clear action. We believe these two jobs are two different layers of intelligence — and separating them, not blending them, is the key to building something that genuinely works.

Two layers, two responsibilities

The first layer is conversational intelligence: the interface that faces the human, understands natural language, guides the interaction, and asks for clarification when something is ambiguous. This layer must be fluent and patient. It deliberately avoids the complex logic of the domain; its sole job is to clarify intent and gather enough context.

The second layer is background intelligence: the specialised services that handle the heavy cognitive work. This layer never talks to the human. Its job is deep processing — structuring the request, inferring priority, and executing domain logic. This is where the actual expertise lives.

Why separation matters

The easy temptation is to pour everything into a single layer: an all-in-one agent that both talks and executes the heavy work. But these two jobs have fundamentally different operating requirements. Conversation must be fast and fluent; deep cognitive work is often slow and computationally expensive. If you blend the two, either the conversation blocks waiting for execution, or the cognitive work is kept deliberately shallow to maintain response times. Separating the layers lets each be the best version of itself.

One agent against two layers over timeTwo timelines of the same late-order message. In one blended agent, it reads the message, then does deep work for a long stretch during which the person sees nothing, then replies; or it cuts the work short, replies sooner and gives up depth. With two layers, the conversation layer reads, says it is checking, asks and reassures while the background layer does the same full deep work, then the reply arrives.one agent does bothtwo layersit blocksor shortcutconversationbackgroundreadsdeep workreplythe person waits and sees nothingreadsshallowreplydepth given upreadschecking…asks, reassuresreplydeep work, nothing cutsame deep work, nobody left waiting
Blended into one agent, the conversation either blocks on the deep work or the work is cut shallow; split into two layers, the person keeps talking while the full work runs.

A simple example

Imagine an end-user sends a message: “My order hasn’t arrived and nobody is answering.” The conversation layer intercepts this, recognises it as a complaint about a delay, and prompts for an order number if one is missing. It then hands the payload to the background layer, which structures the request, determines the issue type, assigns priority, and defines the next action. Finally, the conversation layer provides the user with a clear, natural reply. The user experiences a simple conversation, but behind it, two distinct layers of intelligence work in concert.

One late-order message through both layersIn the conversation layer, the customer writes: My order hasn’t arrived and nobody is answering. The layer understands a late-order complaint, asks for the order number and gets 3095. It hands the request down to the background layer, which structures it into a case: type delivery delay, order 3095, priority high, next action trace the parcel. The case goes back up and the conversation layer replies that the parcel is with the courier and has been flagged.conversation layer: talks to the personbackground layer: never talkscustomerMy order hasn’tarrived and nobodyis answering.understandsintent: late orderasks: order no.?gets: #3095repliesIt’s with thecourier; we’veflagged it.structures itinfers priorityruns domain logictypedelivery delayorder#3095priorityhighnexttrace the parcel
A late-order message passes through the conversation layer, becomes a structured case in the background layer, and comes back as a plain reply.

This isn’t a chatbot plus an API

To be precise: this is not merely a chatbot calling a backend service. The critical difference is that both layers are intelligent. The conversation layer is not a static interface; it understands intent and guides the conversation itself. The background layer is not a fixed function; it reasons and decides itself. This is cognitive layering: a conversational presence alongside specialised depth, neither of which is sufficient alone.

Chatbot plus API against cognitive layeringA table compares a chatbot plus an API with cognitive layering on the same late-order complaint. The chatbot is a static interface with no sense of intent, over a fixed function get_status(order_id) that runs the same steps every time, and it returns in transit, the lookup and not the complaint. In cognitive layering the conversation layer understands intent, guides the talk and asks when unclear, the background layer reasons that the order is late and ignored and so escalates at high priority, and the reply fits the actual complaint. Both layers are intelligent; neither is enough alone.chatbot + APIcognitive layeringtalks to youworks outyou geta static interface:menu, fixed form,no sense of intenta fixed function:get_status(order_id)same steps every time“in transit”the lookup, notthe complaintunderstands intent,guides the talk,asks when unclearreasons and decides:late and ignored,so escalate, higha reply that fitsthe actualcomplaintboth layers are intelligent; neither is enough alone
Both layers reason: a chatbot over a fixed function can answer the lookup, but not the complaint.

Putting it together

We believe the future of intelligent assistants isn’t one do-everything agent, but layers of intelligence that each do their own job excellently and work in concert. The conversation layer stays with the human; the background layer brings the depth. And it’s precisely this division of labour that lets a simple message become the right action — without sacrificing either the fluency of the conversation or the depth of the expertise.