Mila — Chatbot Design Principles

01 — The Organizing Principle

Definition

Mila earns her place when she does something the UI cannot do — not something the UI already does, faster.

A chatbot that wraps existing functionality without adding intelligence is worse than the original. It adds a layer of indirection without adding value. Users learn that Mila is the slower, less reliable path to something they could do faster another way. That's not a trust signal going in the right direction.

The distinction

Capability extension means Mila does things the UI can't do. Convenience layer means Mila does things the UI already does, differently.

Inline intelligence — insights surfaced directly on the page, contextual to where the user already is — is almost always better than a chatbot for predictable needs. The user doesn't have to ask. The information appears where it's relevant. Mila belongs where the user's need is too variable or open-ended to surface proactively. When you know what the user needs, surface it. When you don't, Mila earns her place.

02 — Capability Extension vs. Convenience Layer

Capability extension — Mila belongs here

+ The input space is genuinely open-ended and can't be anticipated or pre-surfaced e.g. loan scenario suggestions — every borrower situation is different
+ The user doesn't know what they're looking for yet and needs to reason through it conversationally e.g. "what should I be looking at?" before a pairing decision
+ The task requires synthesizing across data sources or turning a natural-language request into a configured artifact that would otherwise require navigating multiple UI steps e.g. a report card, loan comparison, debt-consolidation analysis, or flyer.
+ Behavior signals genuine friction and Mila can surface a contextual response e.g. repeated failed actions, extended time on page without progress
+ The output requires reasoning over variable inputs specific to the user's situation e.g. tailored guidance, not templated responses

Convenience layer — Mila doesn't belong here

— The feature is navigation or lookup with no inference e.g. searching for an agent and returning a profile link — just navigate there
— A better version of this already exists in the product e.g. replacing a working Zendesk chatbot with Mila doing the same thing worse
— Inline intelligence would surface the same information in context without requiring a prompt e.g. agent insights, buyer activity overviews — these belong on the page
— The output is deterministic and rule-based — no model needed e.g. scheduled sends, status updates, templated outputs
— The real problem is a broken flow or unclear IA e.g. users can't find something — fix the navigation, don't add a chatbot on top of it

03 — Feature Examples Applied

Feature	What it does now	What it would need to do	Verdict
Loan scenario suggestion	LLM generates customized loan scenarios based on user inputs	Reason over variable borrower inputs to suggest options the Loan Comparison tool can then evaluate. Input space is too open-ended to pre-surface.	Capability extension
Agent search	Returns a profile link for a searched agent	To belong here, Mila would need to return intelligent information — insights about the agent, whether they're a good match, signals from transaction history. A link is not intelligence.	Convenience layerFix the search UX or surface insights inline
Zendesk / product help	Searches Zendesk articles — replaces a chatbot that did the same thing better	To earn her place, Mila would need to do more than retrieve — recommend next actions, surface relevant reports to create, synthesize across help content and product context.	ConditionalOnly if Mila adds reasoning, not just retrieval
Report & artifact creation	Creates predefined reports and artifacts, but is limited by existing templates and workflows	Earns its place when Mila can translate intent into a configured, useful output — not simply expose an existing report-generation flow through chat. The value is reducing the work required to specify, assemble, and produce the artifact.	ConditionalIntent → configured artifact earns it. UI shortcut doesn't.
Behavioral friction response	Surfaces open-ended situational prompts when behavior signals genuine friction	Already doing this correctly — prompt earns its appearance rather than assuming it. Mila responds to context, not availability.	Capability extension

04 — The Test — Three Questions Before Scoping a Mila Feature

01

Does the UI already do this — or could it?

If yes, Mila is a convenience layer. Inline intelligence surfaced in context is almost always better than a chatbot for predictable needs. Build the UI feature instead.

02

Is the input space variable enough that it can't be anticipated?

If yes, Mila has a legitimate role. The open-ended, user-specific nature of the request is what a conversational interface is actually good at. If no, surface it proactively.

03

Is Mila adding reasoning, or just retrieval?

Retrieval without inference is a worse navigation path. Mila earns her place when the output requires genuine reasoning over variable inputs — not when she's wrapping a lookup in a chat interface.