Agent Jones – AI Front Desk for Local Service Businesses

An AI front desk that chats with customers, books jobs, and hands off tricky cases to humans with context.

Comment

Inspiration

Most local service businesses (like mobile detailers) don’t have a real front desk. It’s usually one owner juggling:

  • Calls and texts from customers
  • Web form submissions
  • Actually doing the work in the field

Missed calls = missed revenue, and even when a form exists, it doesn’t talk back or update the owner’s tools. We wanted an AI “front desk” that:

  • Feels like a real person, not a generic chatbot
  • Can actually update a lead’s status and revenue
  • Knows when to escalate instead of over-promising
  • Gives the owner a clean summary instead of raw logs

What it does

Judge Mode home

Landing page built for judges and owners:

  • “Start customer flow” → opens the AI chat demo
  • “Open owner dashboard” → jumps to the business view

Customer chat → lead → booking

  • Customer fills a short form (name, phone, “What do you need done?”).
  • We send that context into Claude and generate a personalized first message.

  • The AI then drives the conversation:

    • Confirms ZIP / service area
    • Explains pricing ranges
    • Offers concrete time slots (e.g., “Saturday 10am or 1pm”)

  • When the customer says “1pm works”, we:

    • Infer status = BOOKED, service type, estimated revenue
    • Persist it on the Lead so the dashboard updates automatically

Service-area & pricing guardrails

  • If the ZIP looks borderline (“outside our main service area”) or the customer is upset about price, the AI:

    • Stops short of confirming a job
    • Sets status = ESCALATE via our action contract or fallback logic

  • The owner can see exactly why it escalated from the transcript.

Owner dashboard

  • Summary strip:

    • Leads today
    • Booked revenue
    • Escalations needing attention

  • Filterable/searchable leads table:

    • Status pills (NEW / QUALIFIED / BOOKED / ESCALATE)
    • Estimated revenue per lead
    • CSV export of the filtered view

  • Clicking a row opens a drawer with:

    • Full message history
    • Service details, status, chosen slot, revenue

AI Daily Summary for owners

  • Owner clicks “Ask AI for today’s summary”.

  • Backend sends Claude a structured snapshot:

    • Counts by status, total revenue, recent leads, sample messages

  • Claude replies with a short briefing:

    • What happened today
    • Patterns (e.g. “lots of demand from 77005 for interior-only jobs”)
    • Concrete action items (who to call back, what offers to test)

AI Follow-up Coach for escalations

  • For ESCALATE leads, the drawer shows an “AI Follow-up Coach” panel.

  • Owner clicks “Draft follow-up message”:

    • Claude reads the actual conversation and drafts a ready-to-send SMS / email.
    • The draft is editable, with Copy SMS and Regenerate buttons.

  • For non-escalated leads, the panel is visible but disabled, explaining it’s for pricing / location / angry-customer cases — a gentle onboarding for the feature.

How we built it

  • Stack

    • Next.js 15 (App Router) + React + TypeScript
    • Tailwind CSS for the UI
    • Vercel for hosting
    • Anthropic Claude for all AI behavior

  • AI orchestration

    • A single conductor.ts module builds the system prompt, lead context, and user messages.
    • We use a typed ActionContract object to capture:
    • status updates
    • chosen time slot
    • estimated revenue
    • escalation flags and updated lead fields

  • Fallback inference

    • Sometimes the model forgets to fill the ActionContract.
    • We run inferUpdatedLeadFieldsFromReply() which parses the reply text to infer:
    • booking confirmations
    • revenue estimates
    • whether language like “outside our service area” should force status = ESCALATE.

  • Next.js 15 compatibility

    • /demo and /owner are server components that render <Suspense><DemoPageClient/></Suspense> and <Suspense><OwnerPageClient/></Suspense>, keeping all hooks and useSearchParams in client components.

Challenges we ran into

Keeping AI and dashboard in sync

  • Early versions: Claude would say “You’re all set for Saturday at 2pm for $150” but forget to set status: BOOKED or revenue.
  • Fix: added the structured ActionContract plus a fallback that infers missing fields from reply text so the UI always reflects reality.

Next.js 15 + useSearchParams

  • Our first UI used hooks directly in the app router pages, which broke static generation and Vercel builds.
  • Fix: refactored /demo and /owner into thin server components that wrap <DemoPageClient/> and <OwnerPageClient/> in <Suspense>.

Designing good escalation behavior

  • It’s easy for an AI to over-promise (“You’re all set!”) even when the ZIP is far away or the customer is angry about price.
  • Fix: added a dedicated “Service Area & Escalations” section in the system prompt and a guardrail in our fallback logic so phrases like “outside our main service area” force status = ESCALATE instead of booking.

Avoiding repetitive questions

  • The model occasionally re-asked for phone numbers we already had.
  • Fix: added a Phone Number Handling section in the prompt and made the lead context explicit (lead.phone). Claude is told to reference it instead of asking again.

Accomplishments that we’re proud of

  • Turned a static form into a fully AI-driven intake + booking flow.
  • Built a dashboard that stays in sync with the conversation without manual data entry.
  • Designed escalation flows where the AI knows when to hand off instead of forcing a sale.

  • Added two owner-facing AI tools:

    • Daily summary of the business
    • Follow-up Coach for tricky leads

  • Got everything working end-to-end on production (with lint + type-safe builds) within the hackathon window.

What we learned

  • AI is great at conversations but needs strong structure (contracts + fallbacks) to be reliable for operations.
  • Small prompt details (like “don’t re-ask for phone if lead.phone exists”) dramatically improve perceived intelligence.
  • Judging UX matters: a dedicated Judge Mode and keyboard shortcuts make it much easier to demo under time pressure.

What’s next for Agent Jones

  • Mobile owner app + notifications

    • iOS / Android companion app so owners get instant push notifications for:
    • New leads
    • Escalations
    • “No-response” reminders when a lead goes cold
    • One-tap actions: call, text, or mark as done.

  • Google Calendar integration

    • Sync booked jobs directly into Google Calendar.
    • Prevent double-booking by checking availability before confirming a time slot.
    • Send automatic reminders to customers (and optional follow-ups after service).

  • Multi-business + vertical templates

    • Let any local business (plumbers, cleaners, med-spa, etc.) configure their own services, pricing, and service areas without touching code.

Built With

Share this project:

Updates