Where AI helps, and how.

The safety rules (read this first)

• AI never writes to your data. Inventory, sales, customers, prescriptions — every write is a server action triggered by an explicit human click.
• AI suggestions require approval. “Should I message Mrs. Sharma?” → you click Yes → action executes. No auto-execute.
• Customer PII is masked before going to the model. Names and phones become placeholders; the real values are re-hydrated after the model responds. Phones never appear in the API request body.
• Every AI call is logged. ai_audit_log captures the prompt, response, model used, tokens spent, latency, and the calling user. If a customer ever asks “why did Opticops do this?” you have the full trail.
• Frame metadata isn't PII. Box-label photos and SKU info are sent to the API as-is — no redaction needed.

1 · Box-label scanner

On /catalog/skus/new, the 📸 panel at the top. Snap a photo of the printed label inside a frame box → Claude extracts brand, model name, model code, colour name, colour code, lens/bridge/temple measurements, MRP, barcode, and category. Form pre-fills any field you haven't already typed in.

Returns a confidence level (high / medium / low). Below high you get an amber banner asking you to verify every field. The model also writes free-text notes when something on the label was ambiguous or partly obscured.

2 · Brand / model dedup

Same form. Type a new brand (“Ray Ban”) and tab out — Claude compares against your existing brands and catches near-duplicates that simple string matching misses (hyphen vs space, regional spellings, swapped word order). Surfaces an amber pill: “Looks like ‘Ray-Ban’ already exists. Use existing?”

Same flow for model names. One click flips your form from new-mode → existing-mode and clears the dup fields.

3 · Colour-code normaliser

Type a colour name (“Matte Black”) and tab out. Claude looks at your org's existing colour-name → colour-code mappings and proposes a code that mirrors the convention. Fills the colour-code field if it's empty — never overrides a code you've typed yourself.

Why: without this, the same colour ends up encoded as BLK / BK / BLA / BLACK across SKUs and your reorder lists fragment.

4 · Customer-recall message drafts

On /customers/insights, pick a cohort (due-for-recall, lapsed, Rx-expiring, high-value, new), open a customer, click Draft message. Claude reads the cohort context + customer's purchase history and writes a 2-3 line WhatsApp-ready message in the right tone for that cohort.

Review, edit, approve. Approval copies to clipboard for manual paste (Phase 1 will queue a Meta Cloud API send). Every draft is logged to ai_audit_log with the customer ID and your approval decision.

5 · Ask

/ask. A natural-language Q&A interface over your operations data. Type a question, Claude answers with cited rows from your database.

Examples that work today:

• “How much did Tri Nagar make last week?”
• “Who's my top customer this quarter?”
• “Which sunglasses haven't sold in 90 days?”
• “Show me all orders with a balance due.”

Streams answers token-by-token. Click any cited row to drill into the source. Read-only — Ask cannot create, modify, or delete anything. If you ask it to (“cancel order X”) it refuses and explains why.

6 · Daily / weekly insights

/reports/ai-insights. A cron-driven summary of what changed in the last 24h (daily) and 7d (weekly). Examples in the Reports doc.

What this costs you

What's coming

• Hindi + regional-language voice → SQL via the Ask interface (Sarvam vs Whisper still being evaluated)
• Auto-categorisation of inbound supplier emails into purchase orders + receipts
• Frame-trend forecasting from local Instagram tags (catching the next Ray-Ban moment 2 weeks early)
• Per-org fine-tuning on your message tone — so the AI sounds like you, not generic