Skip to RoutePilot case study
PK® / RoutePilot

Independent product case study · AI operations

RoutePilot / AI inbound lead-routing workflow

Turn every inquiry into an explainable next action.

A human-gated workflow that preserves unknown company facts, classifies the request, models implementation scope from real delivery drivers, assigns a regional representative, and prepares—not sends—the operational artifacts a team needs.

  • 0workflow-data transmissions
  • 6visible workflow stages
  • 1required human gate
ROUTEPILOT / TRACE 01
Fixture evidenceVisible rulesNo send
Primary user
Revenue-operations and growth teams
Runtime
Browser-local deterministic planning model
Core decision
Who should review this inquiry next?
Safety boundary
No lookup, storage, submission, or send

01 / Working prototype

The complete decision trail, in one interaction.

Start with a tested scenario or enter an ordinary company domain. Every conclusion exposes its inputs, arithmetic, limitation, route, and next human responsibility.

Private browser demo

You may use a real-looking company and domain to test the workflow. Inputs stay in this browser tab and are not sent, stored, researched, or added to a CRM. Domain context is clearly separated from verified facts.

01 / Inquiry & commercial context

Describe the work, not just the company.

Headcount informs routing only. The planning range is based on scope, users, workflow volume, integrations, timeline, and regulated-context requirements.

  1. 01
    CaptureValidate browser-local inquiry
  2. 02
    ContextNormalize domain without lookup
  3. 03
    ClassifyScore visible inquiry signals
  4. 04
    EstimateModel hours, rate, and complexity
  5. 05
    RouteUse first matching owner rule
  6. 06
    PrepareBuild previews for human review
WORKFLOW / READY

No record exists yet.

Run a scenario to see the evidence, calculation assumptions, routing decision, draft-only outputs, and human-review timer.

02 / Architecture

A production-shaped system with every adapter disconnected.

The prototype proves orchestration and interface behavior without pretending a local portfolio interaction is a deployed revenue system.

  1. 01

    Inquiry intake

    Validate company, region, rollout scope, affected users, annual volume, integrations, timeline, budget, and problem context.

    Browser-local form · no transmission
  2. 02

    Domain context

    Normalize any valid company domain. Exact demo fixtures may add context; every unknown domain remains explicitly unverified.

    No web lookup · no invented research
  3. 03

    Classify + estimate

    Score competing inquiry signals, model implementation hours from operational complexity, then apply a visible blended-rate range.

    Inspectable assumptions · budget never anchors price
  4. 04

    Route

    Evaluate regulated context and complexity before technical category, then select a regional representative for the matched team.

    Evidence-first regional policy
  5. 05

    Prepare previews

    Assemble a CRM-shaped record, response draft, and team alert as local artifacts without transmitting them.

    Preview/download/copy only
  6. 06

    Human review

    Keep the inquiry in a visible review state with an SLA clock; a person remains responsible for every external action.

    No autonomous send or status update

Reading order: inquiry intake → domain context → classify and estimate → route → prepare previews → human review. Connectors represent the local decision flow, not workflow-data transmission.

03 / Walkthrough

Watch the workflow make—and explain—a decision.

The adjacent outline repeats the demonstration stages so the content remains understandable without audio.

Product walkthrough / user-controlled playback
Written walkthrough
  1. Load or edit a scenario

    Select manufacturing, healthcare, or enterprise inputs—or enter any valid company domain.

  2. Run the decision flow

    See domain context, classification evidence, hours-and-rate math, budget fit, and regional routing.

  3. Inspect every artifact

    Review the CRM JSON, response draft, alert preview, and their no-transmission labels.

  4. Exercise the human gate

    Advance the SLA simulator and mark the inquiry reviewed locally.

04 / Technology

Implemented now. Integration seams made explicit.

The distinction matters: a credible portfolio should separate working code from a proposed production adapter.

Implemented in this prototype

  • Next.js + ReactServer-rendered case study with a focused client-side workflow.
  • TypeScriptTyped form state, result contracts, and accessible interaction states.
  • Deterministic JavaScript modelDomain context, weighted classification, scope estimates, regional routing, drafting, and SLA logic shared with tests.
  • CSS ModulesRoute-scoped responsive system with no dependency on a component library.
  • Browser Clipboard + Blob APIsExplicit local copy and JSON download actions; no hidden handoff.
  • Node test runnerExecutable fixtures validate decisions, fallbacks, thresholds, and safety language.

Production integration seams · not connected

  • Structured AI classificationA production classifier could return a schema-validated category and abstain when evidence is weak.
  • Research providerLive enrichment would need licensed sources, provenance, freshness, consent, and an unknown-data fallback.
  • CRM adapterA least-privilege integration could create or update a record only after validation and idempotency checks.
  • Email + team messagingDraft and alert delivery would require explicit authorization, rate limits, audit logs, and human approval.

05 / Evaluation

Fixture-tested outcomes, not vague confidence.

These checks execute against the same deterministic decision module used by the interactive prototype.

2/2

Domain paths verified

Exact demo fixtures enrich; arbitrary valid domains preserve unverified facts without failing or fabricating.

5/5

Category paths classified

Evaluation, automation, product, advisory, and ambiguous discovery scenarios resolve from weighted evidence.

4/4

Routing teams exercised

Trust, enterprise, technical, and general teams each win a tested scenario before regional owner selection.

3/3

SLA states verified

On-track, attention-due, and exceeded thresholds remain mutually exclusive and inspectable.

0

Lead-payload transmissions

The component contains no fetch, form action, webhook, CRM, email, or team-messaging request.

06 / Safeguards

Automation stops where accountability begins.

The demo optimizes for a useful recommendation while preserving uncertainty, consent, reversibility, and a clear owner.

01

Browser-local input boundary

Ordinary work emails and domains are accepted for realistic testing, but the form has no submission endpoint and does not store or transmit the inquiry.

02

Unknown stays unknown

Unmatched domains receive a normalized label and explicit unverified fields instead of invented company research, industry, location, size, or buying signals.

03

No pricing theater

The range starts with rollout scope and adds visible user, volume, integration, urgency, regulated-context, category, and delivery-rate assumptions. Headcount is excluded from pricing.

04

Visible routing policy

Every owner rule and priority is shown, including the selected rule and the reason it won.

05

Preview-only actions

CRM, response, and alert artifacts are clearly labeled local previews. Copy and download are the only available actions.

06

Human gate + SLA

The timer measures waiting for review; marking reviewed changes browser state only and never updates an external system.

Designed and engineered by Prasiddha Karki

Route the evidence.
Keep the human.

Run RoutePilot again View source on GitHub