Conni Operations

AI Insights

The customer-intelligence layer Ray asked for on the call.

Phase 2 · scoping in proposal

"Throw it at the LLM and ask it to analyse."

You said it directly: "we can analyse more data — we just collect as many as we want and throw them into the large language model and ask it to analyse that amount of data." Once Customer 360 unifies the signals, this is the layer that reads across the data and tells your team what to act on.

Runs on Cloudflare Workers AI (edge inference, ~10× cheaper than OpenAI direct, no data leaves Australia) with optional GPT-4 / Claude routing for higher-stakes queries.

Example queries (illustrative)

What your team would actually type in

> Which customers haven't ordered in 90 days but were chatting last week?
I found 23 customers. The 6 highest-LTV are listed below, with their last chat summary and last order date.
> Summarise this customer for me before I call them back.
Margaret Wilson — 67 yo, ordered 14 times since 2022 ($1,840 LTV). Last chat 4 days ago asking about NDIS funding. Has left a 4.7-star review. Prefers email contact in the morning.
> What's the most common reason for returns this month?
38% of returns this month cite "sizing — too small". Concentrated on Chantilly Ladies Briefs in S. Recommended action: update size guide on PDP, flag at checkout.
> Which wholesale accounts are at risk of churning?
4 wholesale accounts haven't ordered in 60+ days but were ordering monthly before. The two highest-revenue ones (Capital Chemist, Whiddon Group) are flagged with their account manager.

Customer auto-summary

Pulls Unleashed orders, Klaviyo activity, Review.io reviews and LiveChat history into a one-paragraph briefing the moment the phone rings.

Natural-language search

"Who bought Bed Pads in NSW in March and chatted about delivery?" — written like a question, executed like a SQL query.

Daily morning digest

Auto-emailed summary of yesterday's standout signals: spike in returns by SKU, customer at risk, new high-LTV customer signup, etc.

Pattern detection

Surfaces things no one asked about: customers buying Chair Pads but not Bed Pads (cross-sell), Aged-Care accounts ordering below historical average (early warning).

Data residency + cost

  • Default model: open-source Llama running on Cloudflare's Sydney GPUs — your data never leaves Australia.
  • Optional escalation to GPT-4 / Claude for complex queries — routed through Cloudflare AI Gateway with full audit log.
  • Token budget controls per user role — no surprise bills.
  • At Conni's expected query volume (a few hundred queries per day) total LLM cost is in the $20–$50 / month range.