Building an AI copilot from 0 to $1B

The Trade-off

Every firm faces the same brutal choice:

• Outsource: Call centers answer fast but can't distinguish a $10M catastrophic injury from a fender bender. High-value cases slip through.
• In-house: More control, but management overhead, QA burden, and churn eat into margins.
• The Black Box: Partners can't calculate ROAS-settlements take 1–3 years. They optimize for "call volume" instead of "case value."

Competitive Landscape

The market had two types of players, and neither solved the real problem:

• Legacy intake platforms bolting on AI features ad hoc. Bloated UX, not AI-native. Strong data gravity and workflow lock-in kept them entrenched.
• Point solutions tackling small pieces of intake without a path from intake to settlement.

The gap - no one offered a single system that captures calls, qualifies leads, coaches staff, and provides operational visibility. All without hiring more humans.

Designing the Voice Agent

I talked to dozens of intake specialists, managers, and law firm partners. I took intake calls myself to understand the real workflow-the pacing, the empathy required, the edge cases that trip up even experienced staff.

I iterated on the voice agent prompt directly with clients, tuning tone, pacing, and question flow based on what callers actually needed. Then I evaluated the agent through simulated testing and hours of calling it myself, hunting for edge cases: confused callers, ambiguous injuries, cases that didn't fit neat categories.

The key factors that matter on an intake call:

• Demonstrating patience and empathy
• Gathering critical information
• Value-added communication
• Asking good follow up questions
• Managing expectations

Actionable, High-Value Insights

For our dashboard, I focused on the questions that kept coming up in my discovery interviews:

• What's customer service quality like at scale?
• Are our staff doing a good job with the case types we typically handle?
• Are we handling the right volume of leads?
• What is the quality of our funnel from a lead value perspective?

Every chart answers a specific operational question. No vanity metrics, no dashboards for the sake of dashboards.

Progressive Disclosure

Early versions of our lead overview showed everything up front. A summary, a full rationale for the lead score, structured data about the lead, calls. Users drowned.

They couldn't quickly ascertain whether the lead was valuable and worth attending to. So I redesigned the lead review flow around progressive disclosure:

• Level 1 - A high-level summary with a color-coded lead score. Attorneys trust "green" leads instantly.
• Level 2 - In depth rationale for the score. Liability factors, damages, case strengths.
• Level 3 - A full transcript and audio playback for deep auditing.

The Scoring UX

Firms traditionally use 1–100 scoring. But intake staff can't parse that granularity in real-time. They need to know: is this a good case or not?

I kept the 1–100 score since partners expect it, but color coded ranges: High (green), medium (yellow), and low (red). Staff get instant triage. Partners retain the familiarity they need.

Managing Real-Time Cognitive Load

Our live call coaching functionality had two major goals: Surface critical insights live on calls without distracting or overloading intake staff.

• Stack metaphor: Insights (green flags, red flags, objection responses, follow up questions) slide in one at a time, only when fully baked. Staff are never confronted by a wall of text or a barrage of UI elements.
• Typing animation: Form fields animate as they auto filling, drawing the eye only when action is needed. Staff make micro corrections without breaking their flow.

The goal: hold a natural conversation. Let the AI work in the background. Intervene only when necessary.

The Analysis

I dug into Snowflake to model unit economics across our beta cohort. The data revealed clear patterns in call behavior, token consumption, and volume distribution that shaped our pricing strategy.

Pricing Strategy

Pricing had to feel fair to small firms (~50 leads/month) while scaling profitably with large ones (500+ leads/month). I worked with Product Marketing and the CRO to inform "Per Minute" + "Per Lead" pricing that covered infrastructure costs with healthy margins across all segments.

We considered outcome based pricing, but until we own the workflow end to end fully and our voice AI agent is able to sign leads, usage based pricing was simple, transparent, and easy to implement as a starting point.}

The Competitive Edge

"The fact that we're so far ahead on intake is why we're closing all these $200K–$300K deals. Control the source of cases, you win the market."

- Matt Noe, Chief Product Officer, Eve

Customer Validation

"Are you afraid of a voice AI picking up the phone, or are you afraid of losing a $10M case? Somebody has just had a massive accident. Weigh the trade-offs. I do think Eve will be answering all of our calls very, very soon."

- Mike Morse, Founder & CEO, Mike Morse Law Firm

Realistic Data is Critical

Early code based prototypes used clean conversations, expected lengths, typical case types. When we plugged in real legal conversations, layouts broke.

The power of code-based prototyping is that you can back your designs with real data dynamically. Use that to stress-test the UI in ways traditional tools never can. I now generate adversarial test data before finalizing any layout.

Balancing the Sizzle

There are features that provide basic utility and improve quality of life. And there are features that sizzle on a sales demo.

Great products balance both. You need a product that is both desirable and usable. Testing and iteration helped us find the balance.

Owning the Last Mile

Traditionally, my design QA process involved producing Notion docs with annotated screenshots explaining design deviations.

Now I just open an MR with all my design polish items baked. It saves an incredible amount of time and significantly leveled up the polish of our product.