Nate DouglasSenior Product Designer, home

Verdigence: designing an AI intake agent law firms could trust

I founded Verdigence and was its only designer: an AI voice agent and follow-up service for personal injury law firms.

  • Conversation design
  • Human review and handoff
  • AI trust and failure recovery
  • AI evaluation
Role
Founder and sole designer, a side venture alongside my full-time role at Conduent
Team
A white-label engineering partner, Client Acquisition, built it. I designed it, wrote the specs, set priorities and reviewed every build.
Timeline
September 2025 to March 2026
Built with
Retell AI, ElevenLabs, GoHighLevel, Make and two Anthropic API agents on Vercel
100 casesin the test every change had to pass
3 tiersof guarantee in the calculator, from 4 to 28 signed cases in 90 days
+32%booked consultations projected: 50 → 66 per 150 inquiries

Also projected: incorrect routing 18% to 6%, missed follow-ups 22% to 7%. For scale: 26% of law firms never reply to an online lead (Hennessey Digital, 2025) and only 40% answer the phone (Clio, 2024).

Where the projections come from

The consultation, routing and follow-up figures come from the Verdigence calculator, for a firm with 150 inquiries a month. A booking means an accepted appointment time and a confirmed calendar entry. A promised follow-up has seven days to happen.

The Verdigence identity

The first version answered calls in seconds. Then I found a record marked “follow-up complete” for a caller nobody had reached. The automation had finished. The inquiry hadn’t.

Verdigence was an AI voice agent and follow-up service for personal injury law firms. I founded it and was its only designer. A white-label engineering partner built it from my specs.

Comparison of one inquiry as the first version recorded it and as it actually stood.
What I designed
  • The intake conversationThe intake conversation
  • Staff review and handoffStaff review and handoff
  • Accurate inquiry statesAccurate inquiry states
  • A 100-case evaluation setA 100-case evaluation set
The design problem

A finished call isn't a handled inquiry

I called firms when their prospects would. After hours, inquiries went to voicemail or a form. Weekends were mostly silent.

The caller needed to be heard and to know what happens next. The firm needed records its staff could act on. If staff didn’t trust the records, they wouldn’t use it.

Intake System A's post-intake review: the package is marked complete and ready while treatment, witnesses, employer involvement and insurance are unknown.
Intake System AMarked ready while the details that set a case's value were still unknown.
Intake System B's dashboard: at the end of the call the caller says he was driving a company truck and has surgery scheduled, and the score, priority and follow-up timing stay the same.
Intake System BA company truck and a scheduled surgery come up at the end of the call, and the priority doesn't change.
Intake System C's platform overview, stating that accuracy is measured at capture and that follow-up is left to the firm's own process.
Intake System CAccuracy is measured at the moment of capture, and follow-up is left to someone else.

1 / 3Click a screen to see it full size.

I tested 15 AI intake systems the way real prospects use them: after hours, with incomplete information and high emotion. Most answered fast and gave a clean summary. These three show the gaps I kept running into.

Hear it. A 1:27 demo call from December 2025. I’m the caller. The agent never gives legal advice: it says an attorney should look at it and hands off. Its questions screen the case, so the firm doesn’t take on ones it can’t help with.

0:00Agent: Thanks for calling Verdigence. How can I help today?

0:04Caller: Yeah, I got rear-ended last Friday, and I wanted to see if I could talk to someone.

0:09Agent: Okay. Were you hurt at all?

0:11Caller: Yeah, my neck and lower back. Um, mostly my neck.

0:16Agent: Got it. With your neck, are you feeling any tingling or numbness in your arms or hands?

It follows the injury instead of jumping to the next field.

0:22Caller: Yeah, actually, there's like some tingling in my right hand.

0:26Agent: Okay. Any weakness there, like trouble gripping things?

0:29Caller: A little, yeah.

0:32Agent: Quick check. Were you driving your own car or a vehicle from work?

A work vehicle changes the priority. Intake System B heard the same detail and changed nothing.

0:38Caller: Oh, it was my work van.

0:42Agent: All right. Were you heading straight home from the job site, or had your shift already ended?

0:48Caller: I was driving home from the site, and I was still on the clock.

0:52Agent: Okay. And has anyone talked to you yet about whether that tingling could be nerve-related?

0:58Caller: No, they just said it was probably a strain.

1:02Agent: All right. Based on what you're describing, this is something an attorney should take a look at.

An attorney at the firm they called. A person read every transcript before the firm took a case.

1:07Agent: What's the best number to reach you? And is it okay if we text you updates?

1:12Caller: Um, yeah. 555-123-4567. And texting me is fine.

Texting needs its own clear yes, and it gets one.

1:22Agent: Perfect. We'll follow up shortly.

Most AI intake systems optimize for completion. An inquiry is handled when it has reached the right next step, with an owner, and with what’s unknown recorded honestly.

Conversation design

Designing the conversation

The agent asked one open question first: what happened, and were you injured? It took that account, then asked only about what was missing, one question per gap. It never guessed. If the answer was still unclear, it stayed unclear. I wasn't going to press an injured person for certainty on a recorded line.

A call might go like this:

Agent: Thanks. Can you briefly tell me what happened and whether or not you were injured?

Caller: I got rear-ended on the highway Tuesday. My neck's been killing me since.

Agent: Got it. Have you received any medical treatment yet, or have you scheduled an appointment?

Caller: I think my sister took me to urgent care? Honestly that whole day is a blur.

Agent: Okay, I'll note that. Would it be alright if someone from our team contacts you about this?

Caller: I guess, as long as it's not a sales thing.

Agent: Just to confirm, I have permission for our team to contact you?

There are two decisions in that call. The treatment answer is neither yes nor no, so it stays unresolved and goes to a person. And "I guess" isn't permission, so the agent asks once more before anyone calls back. The agent gave no legal advice, case evaluations or fee estimates; those questions were noted for the team. The flow went through nine published versions in two days.

Retell editor showing the intake flow's capture steps, missing-information checks and the agent's global instructions.
I separated capture from the missing-information checks: take the account first, then ask one question per gap.

My job was deciding what the system should do, what it must never do, and how I'd know whether it was working. I used AI tools and templates to build faster than I could have alone, and ran the service on Retell AI, ElevenLabs, GoHighLevel, Make and two Anthropic API agents on Vercel.

  • Intake clarification. Checks the caller's account for missing or contradictory information and suggests the next clarification question. It preserves "unclear" rather than inventing an answer and flags unresolved information for firm staff. It does not assess case merit or give legal advice.
  • Follow-up coordination. Reviews the inquiry's current state and recommends the next permitted action: callback, scheduling confirmation, another attempt or staff review. Make/GoHighLevel execute approved actions; application rules enforce consent, attempt limits and valid state changes.

The distinction: one helps understand what the caller said; the other helps determine what should happen next. Firm staff retain human review and case decisions.

Human review and handoff

Where a person steps in

The agent did what needed no judgment: answering, capturing the account, asking permission, calling back and scheduling. Staff handled what did. “Verdigence informs decisions. Licensed attorneys make them.”

I left live transfer out on purpose. After hours it would ring an empty office. Anything needing judgment went to a review stage with the reason, an owner and a response window.

GoHighLevel intake pipeline with names blurred. The Needs Review stage holds five inquiries, each with its review reason: treatment unclear, permission unclear, contact method unclear, booking did not save, attempts used up. Below, the AI intake call note quotes the caller, and a task to confirm medical treatment is assigned to intake staff with a due time.
Unresolved details stay open, with the caller's words, an assigned staff member and a follow-up deadline.
The open opportunity in GoHighLevel with personal details blurred: stage Needs Review, status Open, review reason Treatment unclear, treatment Unclear, the caller's statement, next action Confirm medical treatment and a response window.
Treatment stays unclear until staff confirms it. The open record shows who owns the next action.

My first queue flagged every gap, and staff couldn’t keep up. I narrowed review to the gaps that changed the next step.

Trust and diagnosis

Fast, but quietly wrong

In a weekly review with a firm, we pulled up an inquiry the CRM showed as followed up. The callback had gone out. Nobody had answered. Nothing had happened since. I went back through recent inquiries and found four patterns:

  • Unanswered callbacks recorded as completed follow-ups.
  • Offered appointments counted as booked. "Thursday might work" looked identical to a confirmed booking.
  • Unclear answers collapsed into "no," because the CRM field only had yes and no.
  • "I guess" treated as permission in one step and not in another.

Every step reported its own completion. None of the automations was broken on its own. The problem was what the product counted as done.

Failure recovery

Rebuilding the inquiry states so they were accurate

Three rules. Attempts keep an inquiry open. Only an outcome moves it forward or closes it, with a reason. Every open inquiry has one next action and one owner.

Diagram of inquiry states. Unresolved information goes to staff review. A callback attempt becomes unanswered or connected; connected leads to offered times and only a confirmed appointment hands the inquiry to the firm. Unanswered attempts continue until an agreed limit, then go to staff escalation; declined contact, withdrawn consent or being unreachable under the policy closes the inquiry with a recorded reason.
Attempted actions, completed outcomes, legitimate closes and who acts next.
  • A booking counts only when the caller confirms it. Offered times start a confirmation follow-up.
  • Permission has to be an actual yes, and texts keep their own consent.
  • "Unclear" is a real value everywhere downstream, not just in the conversation.
  • Failures don't disappear. Any step that fails reopens the inquiry with an owner and raises an alert.
GoHighLevel contact with personal details blurred. The text thread asks the caller to confirm Thursday at 2:00 PM, with STOP opt-out language, and the caller replies yes. The activity timeline moves from appointment offered, not confirmed, to confirmation text sent, caller confirmed and appointment booked.
A suggested time stays offered until the caller confirms. Only then is the appointment booked.
Evaluation

A 100-case test every change had to pass

Before callers saw a change, it had to pass 100 cases. Each case set the expected state, next action and owner, and what the system must never do. One must-not violation failed the case.

Evaluation case: situation, expected state, next action, owner and must-not rules, with the case groups.
Evaluation case structure.

The set reproduced the problem before any fix shipped: the conversation can go right and the record can still be wrong. The result below is that failure, from before the rebuild.

Evaluation result for the case Unclear treatment, after hours. The conversation checks pass: one clarifying question, the answer kept as unclear. The record checks fail: treatment saved as No, no next action, no owner. The must-not rule against recording unclear as no fails the case.
The failure the evaluation caught: the conversation preserved "unclear," but the record saved "no" and had no next action or owner. The case failed.

After that, every change had to pass the set, and each new kind of miss became a new case.

Beyond the agent

The same rule in the sales calculator

The same rule shaped the calculator firms saw before they signed: it entered their figures as stated and marked anything estimated.

Calculator showing Entered, Estimated and Required states next to the firm's operating inputs.
I designed this calculator. Entered, Estimated and Required states mean a guess never looks like a fact: the same rule as "unclear" in the intake record.
Projected outcome

What it was built to change

The calculator projected confirmed consultations rising from 50 to 66 out of 150 inquiries, a booking rate of 33.3% to 44%. Speed is the headline, but it only counts if the record is honest.

One number to prove it works, two to make sure it isn’t quietly wrong: confirmed consultations, with incorrect routing (18% to 6%) and missed follow-ups (22% to 7%) as guardrails.

Looking back

What I'd design next

The measures stop at the booked consultation. Next I’d close the loop from consultation back to intake: did the caller attend, did the firm offer to take the case, did the client sign.

If I started over, I’d write the must-not rules before the first version went live. On every AI product since, the evaluation set comes first.