Case study
Voice AI & Sales Intelligence

A voice AI agent, and the sales workspace built around it.

Juno qualifies inbound B2B leads by voice on the Retell API. Our own AI layer turns every call into sourced sales data a team can query.
Kestrel Works' voice agent, Juno, runs on the Retell API. Our custom AI layer reads every call and finds the patterns across all of them. Sales is the use case; the same layer fits any agent that listens for a living.
26
screens
Screens and states.
Designed across the sales workspace.
10
categories
Qualification categories.
Extracted from every call.
7
stages
From call to answer.
A finished call becomes a searchable answer.
0
guesses
Fields filled in blind.
When nobody said it, it stays blank.
What we built.
Seven pieces that turn every call Juno takes into sales data a team can check, act on and query.
Every call becomes
a record you can check.
Calls arrive with the transcript, recording and metadata. The analysis pulls out needs, budget, authority, timeline, objections and buying signals, and pins each one to the moment it was said.
✓ Tagged moments on the call waveform
✓ Every finding jumps to its timestamp
✓ Findings can be verified or flagged
Leads fill themselves in.
After the call.
Each conversation is matched to a lead, company and contact. The lead page opens on Juno's summary, with every qualification field traced back to the call that produced it.
✓ Summary written from the transcript
✓ Reps verify fields with their initials
✓ Qualification kept apart from deal stage
A pipeline.
Not another CRM.
Opportunities move through stages on a simple board. Opening a deal shows its contacts, the call behind it, buying signals and next steps, without leaving the board.
✓ Qualification progress on every card
✓ Stalled deals flagged where you see them
✓ Next steps linked to call moments
Patterns across
every conversation.
Intelligence reads all calls together. Objections, buying signals and pain points are counted by how many conversations mention them, and each theme opens the calls behind it.
✓ Counts out of the total, never bare percentages
✓ Pain points quoted in prospects' words
✓ Every theme opens its supporting calls
Ask the pipeline
a question.
Ask AI answers questions about leads and conversations in plain language. Each claim cites the lead and transcript moment it came from, and a sources panel lists everything used.
✓ A citation on every claim
✓ Says plainly what the data can't answer
✓ Only searches records the user can access
Built for one agent.
Ready for more.
The AI Agent page shows Juno's instructions, knowledge sources and call outcomes. Every call is tagged with the agent that took it, so new agents plug into the same data layer. Private test calls never become leads.
✓ Instructions written in plain language
✓ Every call tagged by the agent that took it
✓ Test calls never logged as leads
Knowledge the agent
can't get lost in.
Decks, documents and FAQs go into one managed library instead of being pasted into a prompt. Each source is categorized and processed before the agent can use it, and the agent only answers from sources marked Ready.
✓ Only Ready sources reach the agent
✓ Failed files say why, with a retry
✓ Sources sorted by category and file type
How it works.
The architecture, the evidence behind every field, and the rules Juno follows on every call.
Retell runs the call.
Our AI layer runs the data.
Juno is built on the Retell API, which handles the live voice conversation. When a call ends, Retell sends it to our platform. From there a custom AI layer ingests it once, extracts each field with its evidence, links it to a lead, and powers intelligence and Ask AI across every call.
Retell APICustom AI layer · dataSales teamCustom AI layer · queryknowledge sourcescall endsRetell webhooklinks to a leadfindings · evidencereads onlyanswers · citationsJunoRetell voice agentCall recordtranscript and audioIngestionone record per call, retriesAI analysisfields, each with evidenceSales recordsleads, contacts, dealsIntelligencethemes across all callsWorkspacereview, verify, correctKnowledge baseproduct, pricing, playbookAsk AIquestions about the dataMoves or writes call dataRead only
Pipeline
Seven stages.
From call to answer.
Each stage tracks its own status. If analysis fails, the call and transcript stay saved, the failure is visible, and a retry is one click away.
01
Engagement
A prospect calls in and Juno runs the qualification
02
Capture
The voice provider sends the finished call
03
Processing
Stored once, never twice, then analyzed
04
Association
Matched to an existing lead, or a new one is created
05
Intelligence
Findings join the analysis across all calls
06
Review
The team checks, verifies and corrects in the workspace
07
Query
Ask AI answers from the records a user can access
Unknown stays unknown.
Never a guess.
Every extracted field carries the words it came from and a timestamp to jump back to. A rep can verify, flag or correct it. When a prospect never said something, the field reads Not captured instead of filling the gap with a guess.
Marcus Bell
Ledgerline · 11:24 call · 6 of 8 fields captured
Field
Value
Said on the call
Status
Budget
$20k to $25k
“Somewhere between twenty and twenty-five thousand”
04:12 · AI extracted
Decision maker
Dana Ruiz, CFO
“Anything over twenty thousand goes to Dana Ruiz”
07:38 · Verified by RK
Team size
Not captured
—
Ask on the next call
Not captured on this call: team size, competitors evaluated
Left blank, never guessed. Juno asks on the next call.
Open lead
8
fields
Structured fields.
Extracted from each qualification call.
3
states
For every field.
Extracted, verified or not captured.
1
click
From finding to moment.
Any finding jumps to where it was said.
Designed for slow, failed
and empty answers.
AI output is slow, uncertain and sometimes missing. Each screen was designed in its in-between states too: analysis still running, analysis failed, an answer streaming in, and a question the data simply can't answer.
Analysis running, analysis failed
The call is saved either way. Failures say so and offer a retry.
Loading the workspace
Skeletons keep the layout still while numbers arrive.
An answer being written
Ask AI shows which records it is reading as it works.
No supporting data
When the calls don't cover a question, Ask AI says what's missing.
Guardrails
Juno says what it is.
Every time.
The agent's rules live in its instructions in plain language, and the team checks them in private test calls before a prospect ever hears them.
Disclosure
Says it's an AI and that the call is recorded, in its first turn
Consent
A declined recording is flagged for a person to review
Pace
Asks one question at a time, in about ten minutes
Pricing
Never quotes custom pricing on the call
Escalation
Flags for review when the budget owner is unclear, and offers a callback when a prospect asks for a person
Principles
Four decisions that hold it together.
01
The data is yours, not the vendor's
Calls, analysis and insights live in your own system, so they stay your IP if the voice provider ever changes.
02
Not captured is an answer
Missing information stays missing, never filled in with a plausible guess.
03
No insight without sources
Every theme and every answer links to the calls behind it.
04
Testing never skews the data
Private test calls run on the live agent but never become leads or count toward insights.

Let's find your
next win.

In one call, we'll map where AI fits in your work, what it would take to build and what it's worth to your team.