Framing the Challenge

‍In enterprise environments, users rarely ask simple questions. They cite invoices, reference contracts, jump between systems, and ask compound questions. Legacy tools forced them into narrow flows or returned generic results.

Sentry needed to bridge the gap between functional automation and conversations that felt intelligent, relevant, and human.

Most chatbots either oversimplify or overwhelm. They script surface-level replies, expose backend jargon, or force users to rephrase into machine-speak. That’s how trust erodes. Support gets slower. Frustration builds.
‍
‍

The Challenge

Design an interface that recognizes enterprise intent, adapts to real-world constraints, and still feels like a capable human partner—not a rule-based form with a personality.
‍

Key pain points:

1. Ambiguous & Complex Intent – Skip the jargon. Every message, prompt, and output needed to feel legible and purposeful – no mental gymnastics required.
2. Rigid Scripts – Whether it was a file name from earlier or a timestamp mentioned in passing, the agent kept track – so users didn’t have to repeat themselves.
3. Lost Context – No dead ends. Users could switch to a human, get a summary, or redirect a thread – without starting from scratch.
4. ‍Poor Handoffs – Human agents had no visibility into prior AI interaction, creating rework and frustration
   ‍
   ‍

Visually, we opted for a hybrid layout: user queries on the left, agent replies on the right. Actions like clause extraction or escalation routing were embedded directly into the thread; so users didn’t have to pause, switch context, or dig for options. The goal was continuity, not choreography.

• Item 1
• Item 2
• Item 3

What they asked for

That was the ask—short, deceptively simple.But when we listened between the lines, the actual need was less about conversation and more about coordination.

• Users weren’t “chatting” — they were triaging.
  ‍The dialogue wasn’t casual. It was clause validation, vendor status, or policy lookup—all deeply contextual, often buried in documents or tied to upstream systems.
• Legacy chatbots added friction, and obstacles to the end, desired outcome.
  Users were either asked to repeat themselves, or simplify complex requests into rigid templates. The result? Abandonment or escalation.
• Teams didn’t want a “friendlier” interface.
  They wanted one that could think in threads, understand documents on the fly, and keep the operational context alive without needing handholding.
  ‍
  ‍

In essence, they weren’t asking for a chatbot.

• Automate common intake flows—without forcing users to write prompts
• Extract contract terms on the fly—without schema training
• Respond in context—without breaking security or protocol

‍
‍
Bottom line: They weren’t asking for a chatbot.They were asking for a teammate—one who reads the brief, understands the file, and knows what to do next.
‍

• Item 1
• Item 2
• Item 3

Interface Logic & Agent Design

Designing the agent meant balancing two tensions: machine intelligence and human intuition.

The challenge wasn’t to seem smart—it had to act smart. That meant interpreting intent, handling ambiguity, and knowing when to defer. Just as important, it needed to earn trust: no uncanny mimicry, no overreach, no pretense.
‍

We anchored the agent experience around three working principles:

1. Conversational Clarity – Skip the jargon. Every message, prompt, and output needed to feel legible and purposeful—no mental gymnastics required.
2. Contextual Memory – Whether it was a file name from earlier or a timestamp mentioned in passing, the agent kept track—so users didn’t have to repeat themselves.
3. Fail-Safe Logic – No dead ends. Users could switch to a human, get a summary, or redirect a thread—without starting from scratch.
   ‍
   ‍

Visually, we opted for a hybrid layout: user queries on the left, agent replies on the right. Actions like clause extraction or escalation routing were embedded directly into the thread—so users didn’t have to pause, switch context, or dig for options. The goal was continuity, not choreography.

• Item 1
• Item 2
• Item 3

Compliance without the Coldness

In enterprise environments, compliance isn’t a feature—it’s table stakes. But most tools treat it like a tax: rigid, brittle, and bolted on after the fact.
We designed compliance into the conversational layer from the start—not to restrict users, but to give them clarity and control mid-dialogue.

Our agent respected:

• Role-based access – Users only saw and triggered what their role permitted.
• Audit traceability – Every document, prompt, and action logged and time-stamped.
• Data boundaries – No bleed across departments, no hallucinated access outside allowed scopes.
  
  ‍
  ‍

Instead of burying these mechanics deep in backend logs, we surfaced them:

• Agents quoted their sources directly in chat.
• Users could inspect logic chains mid-conversation.
• Controls existed to pause, override, or redirect sensitive queries.
  ‍

By making guardrails visible—but never disruptive—we built an agent that felt accountable, not cold.

• Item 1
• Item 2
• Item 3

System Feedback & States

AI tools often crumble under context drift—forgetting users, misplacing threads, or resetting logic mid-task.

In fast-moving enterprise settings, that’s not a UX bug—it’s an operational cost.

‍

So, we designed the agent to be state-aware and state-retentive. At any given moment, it knew:

• Who initiated the interaction
• What the ongoing task thread was
• Which file, query, or system entity was being referenced
  ‍

The interface made this visible: context summaries, embedded variables, and inline previews let users confirm what the agent was using and why.No guesswork. No memory holes. No toggling between tabs to re-educate the system.
‍

‍

We embedded live context markers across the UI—showing active variables, recall triggers, and linked task metadata. Instead of pulling users out of the moment to rephrase or reorient, the agent kept the thread warm.

This meant fewer resets, less rework, and smoother handoffs—even across departments or time zones.
The result: a thread that stayed coherent—even when the user switched topics, paused mid-task, or resumed later from another device.

• Item 1
• Item 2
• Item 3

Voice & Tone Layering

Tone isn’t cosmetic—it’s operational.

In high-stakes environments, a misfired response can trigger confusion, mistrust, or escalations. So we didn’t aim for one voice. We built a tone system—modular, adaptive, and precise.

We built a tone system that was situational, adaptive, and branded.

• Onboarding interactions were warm but direct. Just enough charm to lower friction, without feeling scripted.
• Escalation threads used clipped syntax and minimal surface language—prioritizing clarity and auditability.
• Task transitions leaned instructional, with focused verbs and no empty pleasantries.
  
  ‍
  ‍

To avoid the “chatbot glaze,” we mapped tone to intent—not just the surface domain. That meant scripting fallback responses for ambiguity, modeling confidence thresholds, and distinguishing what needed to sound human from what needed to sound official.

We also built in tone toggles—letting teams switch between formal, neutral, or friendly registers depending on their org culture or compliance needs.

This wasn’t about personality. It was about precision: sounding like a partner when appropriate, and like an SLA policy when necessary.

• Item 1
• Item 2
• Item 3