A guided conversation that still feels like a conversation.

Empathy → Define → Ideate → Prototype → Test & iterate.

The problem: three good tools, no shared concept.

Frontline teams already had the pieces: a flow engine that stepped agents through predefined tasks, small focused AI answers for specific actions, and static instructional cards for reference. Each one worked on its own, but nothing connected them into a single coherent experience for the conversations that actually matter — sales, discovery, support — so agents were still swiveling between tools mid-conversation, and customers felt the seams.

Naming the concept, and drawing the line around it.

The core definition had to be precise enough to design against: a Conversational Path is the guided journey a user takes through a conversation with an AI assistant. The user talks (or types), and the assistant responds in a way that feels natural — like a helpful guide walking them through a process every step of the way. It blends a generative, guided response style with a step-by-step adaptive flow that stays non-linear, driven by the user's actual intent and input.

The fastest way to lose a new concept is to let it blur into the tools people already know. Two comparisons kept it distinct:

The framework was built to support both ends of a conversation from day one: assisted, where an employee stays in the loop and the AI coaches them in real time, best for complex or sensitive conversations; and unassisted, where the customer talks to the AI directly for routine tasks, escalating to a person only if things get complex or emotional.

Principles first, then the decision logic, then the options.

Before any screen, the concept was grounded in a short list of operating principles: data-driven with customer segmentation, AI-powered offer eligibility built from customer intent plus business intelligence, low-code so the logic stays easy to extend, and — the point of the whole exercise — removing the cognitive load from the agent.

• Data-driven, with customer segmentation.
• AI-powered offer eligibility, built from customer intent plus business intelligence.
• Low-code, no-code — dynamic agentic assists, and paths built from guided, intelligent questions that pinpoint customized offers.
• Remove the cognitive load for the agent.

Discovery isn't a script — it's a loop. Based on intent, the assistant runs a discovery path for the right line of business, pulls in what it already knows about the customer, generates guided questions, narrows those answers down to a persona and offer set, and arms the agent with talking points, rebuttals, and promos. From there it can route straight into checkout, refine the offer if needed, surface a likelihood-to-close score to prioritize effort, and loop back with leading questions for other lines of business.

The last step loops back to the first — closing out one line of business becomes the opening question for the next.

Rather than design one generic flow, the concept work scoped five distinct directions up front — because a new prospect and a returning customer need genuinely different conversations, not the same flow with different copy.

A review with stakeholders validated the direction and pushed two specific changes into the next pass: build in a way to tie a new sale back to the conversation that originated it, and make sure the experience reads as valuable for every agent — not only the ones who asked for "a guided experience" — since the flexible questions were a deliberate design choice, not a gap.

Discovery, built out as a working concept.

Of the five directions, the prospect-facing discovery flow got built out as a concept prototype: an agent opens a suggested action on a customer's account, the assistant asks what the customer cares about most using quick-pick chips rather than open text, and as the agent narrows it down, the assistant hands over real-time talking points — competitive comparisons, promotions, trust-building answers — instead of leaving the agent to improvise.

The recommendation step doesn't just hand over an offer — it scores it, so the agent knows how hard to push, and it stays reversible: "Not this one" and "more questions" are first-class options, not dead ends. A cross-sell variant applies the same pattern to mobile — "save money by adding mobile" — using the same chip-based question style rather than inventing a new pattern per use case.

The framework's tone rules — personalize and empathetic, trusted and transparent, accessible and inclusive, collaborative and evolving, clear and efficient, dialogue over monologue — weren't just a slide. Applied to a real billing question, they're the difference between a wall of text and an answer that explains the increase, shows the math, and offers a next step in one short exchange.

The framework grew up — literally changed its name.

The clearest sign this was tested and iterated on, not just shipped as a deck: the concept itself changed. A second round of work reframed "Conversational Paths" as "Agentic Paths" — the same core definition, the same building blocks, the same assisted/unassisted split — but with an added layer of rigor, breaking every path down into a repeatable anatomy (acknowledge the request, surface the right content, offer the right prompts) and mapping how a single path scales from one question into a full branching conversation tree.

What's still open, stated plainly: whether the framework still needs its own flow engine once other technologies mature, how it should integrate with scripted automation, and how deep "dynamic" paths — ones that reconfigure themselves mid-conversation — should go. None of that blocked the discovery concept from getting built and reviewed, but it's real, unresolved scope, not an implied "done."