AI Features

Overview

SalonOS has multiple AI-powered capabilities built on an internal AI gateway. The gateway routes requests to the appropriate Anthropic Claude model (Haiku, Sonnet, or Opus) based on cost and latency requirements, enforces tenant scope, handles rate limiting and cost tracking, and logs all AI interactions for audit and quality review.

AI Gateway

All AI capabilities route through a centralised gateway:

Responsibility	Description
Model routing	Selects Haiku (fast/cheap), Sonnet (balanced), or Opus (complex reasoning) per use case
Tenant scope	All prompts include the tenant's context; the AI cannot access another tenant's data
Tool management	Manages which tools (functions) the AI can call for each capability
Guardrails	Validates outputs before they reach the product
Logging	Every AI call is logged: prompt, model, tokens, cost, response
Cost tracking	Per-tenant token and cost accumulation
Idempotency	Duplicate calls return cached responses where appropriate

Model Selection by Capability

Capability	Model	Rationale
Booking agent (chat)	Claude Haiku	Low latency for real-time conversation
Voice receptionist	Claude Haiku	Sub-second response needed
Marketing copy	Claude Sonnet	Quality matters for brand voice
Review replies	Claude Sonnet	Nuanced, brand-appropriate tone
AI Accountant	Claude Sonnet	Accurate reasoning on financial data
Ask Your Data	Claude Sonnet	SQL generation quality
Style analysis	Claude Opus	Complex vision + reasoning
Forecasting insights	Claude Haiku	Summaries of statistical output

AI Booking Agent

The AI booking agent is a conversational interface for clients to book appointments via WhatsApp or the website chat widget.

Conversation Flow

1. Client sends a message (e.g. "I'd like a haircut this Saturday")
2. The AI identifies the service, checks availability, and asks for clarifications (preferred stylist, time preference)
3. The AI confirms the slot and asks for the client's name/phone (if not recognised)
4. On confirmation, the AI creates the appointment in the scheduling engine
5. A confirmation message is sent to the client

Capabilities

• Handles multi-service bookings in one conversation
• Understands natural language requests (no rigid command structure)
• Looks up existing clients by phone number
• Respects real-time availability (cannot book a slot that is taken)
• Handles reschedules and cancellations
• Gracefully hands off to a human when it cannot fulfil the request

Human Handoff

The agent recognises when it should escalate:

• Client complaints or unhappy messages
• Special requests outside standard booking
• Payment disputes
• When it fails to understand the request after two attempts

On handoff, the conversation is flagged in the inbox for a front-desk response.

Guardrails

• Cannot make bookings outside the location's operating hours
• Cannot override capacity limits or conflict rules
• Cannot access client financial data
• Cannot modify or cancel an existing booking without explicit client confirmation
• Responses are validated before sending to ensure they are coherent and safe

AI Voice Receptionist

An AI-powered phone receptionist that handles inbound calls to the salon when the front desk is busy or the salon is closed.

Call Handling

1. Caller dials the salon's number; call is routed to Dograh (voice AI provider) or Twilio
2. The AI greets the caller, identifies the call purpose (booking, cancellation, information)
3. For bookings: collects service, preferred date/time, and caller's phone number; creates the appointment
4. For cancellations: verifies the caller's identity and cancels the appointment
5. For other queries: provides salon hours, address, and standard information; offers a callback

Voice Capabilities

• Supports natural, flowing conversation (not a DTMF phone tree)
• Handles interruptions and corrections gracefully
• Recognises when to offer a callback (cannot resolve the request)
• Leaves a voicemail summary in the staff inbox when a message is left
• Call recordings are stored per tenant (configurable retention)

Integration

The voice system integrates with Twilio for call routing and optionally with Dograh's voice AI platform for more sophisticated call handling. All booking actions taken by the voice agent write to the same scheduling engine as the calendar.

AI Style Advisor

The style advisor analyses client photos and consultation data to suggest services, colours, and styles.

How It Works

1. Client uploads one or more photos (face, hair, inspiration images) through the booking flow or portal
2. The AI analyses: face shape, current hair condition, skin tone, and (if provided) the style inspiration
3. The AI generates style recommendations with explanations: suggested services, colour approaches, maintenance expectations
4. Recommendations are shown to the client and stored on their client record for the stylist to review

Use Cases

• Pre-appointment consultation: client gets recommendations before arriving
• Stylist briefing: the assigned stylist sees the AI analysis alongside the consultation notes
• Self-discovery: clients can explore options without committing to a booking

Models Used

• Vision analysis: Claude Opus (complex visual reasoning)
• Recommendation generation: Claude Sonnet (quality explanations)

AI Accountant

The AI accountant provides financial Q&A and analysis on the salon's accounting data.

Capabilities

• Answer questions like "What was my profit margin last quarter?" or "Which service category has the lowest COGS?"
• Identify unusual patterns: "Your supply costs in February were 30% higher than the 6-month average"
• Generate narrative summaries of financial periods: "Here is a summary of Q1 2026..."
• Suggest actions: "Your AR aging shows ₹45,000 overdue — here are the top 5 invoices to chase"

How It Works

The AI Accountant operates on a structured view of the brand's financial data (P&L, expense breakdown, AR/AP summary). It does not have direct database access — it receives a structured data snapshot and generates analysis from it.

All outputs are prefaced with a disclaimer that the AI is providing analysis assistance, not certified accounting advice.

Marketing Copy Generation

Within the marketing campaign builder, AI can draft message content:

• Provide the offer and tone; AI writes the WhatsApp message or email
• Suggests multiple variants (formal, casual, playful)
• Adapts to the brand's configured voice guidelines
• Draft is always reviewed and approved by a human before sending

Review Response Drafts

In the reputation module, AI drafts responses to Google reviews:

• Positive reviews: warm, personal thank-you
• Negative reviews: empathetic acknowledgement, invitation to resolve offline
• Neutral reviews: balanced, friendly
• All drafts reviewed and approved before posting

AI Interaction Audit Log

Every AI interaction is stored in the audit log:

• Timestamp, user/session, capability used
• Prompt tokens, completion tokens, model used, latency
• Whether the output was used or discarded
• Any guardrail triggers

This log supports quality review, cost management, and debugging.

Evaluation and Quality

AI outputs are monitored continuously:

• Offline evaluation: golden test sets per capability run nightly; results alert if quality drops
• Online monitoring: production samples reviewed regularly
• User feedback: thumbs up/down on AI-generated content is collected
• Regression detection: new prompt versions are A/B tested before full rollout