StellaAI-driven CRMWe built BOLD an AI that knows when to reach out.
- Watches nine buying triggers across the entire customer base
- Drafts the message, and the reasoning behind it, per customer
- A human approves every send: that is the default state, not a setting
Client: BOLD Precious Metals · bullion e-commerce, USA · in production
The signal was always there. Acting on it was the problem.
Stella reads the customer base daily and picks the best next action per person.
Read-only in, human-approved out
Four problems every lifecycle programme runs into
Each maps to a mechanism in the product, not a promise.
Signal you can't extract by hand
The triggers sit across orders, holdings, watchlists and clickstream. Nine builders compute each audience live, on demand.
Batch-and-blast doesn't scale
Claude drafts per customer, not per segment. Identical decisions collapse into one call: 500 lapsed VIPs, about 12 calls.
Timing windows last hours
A 2% dip is live for hours, not weeks. That scenario runs intraday, and sends nothing if nothing moved.
Nobody automates blind
The fear is an AI mailing the whole list something wrong. So review is the default path, with four limits behind it.
Nine reasons to reach out, watched every day
Real thresholds, computed live. Each has its own cooldown, so nobody is pestered.
Lapsed VIP
Spent over $5,000. Nothing for a year.
Watchlist price drop
They saved an item. Spot has fallen since.
Dormant holders
Still holds metal. No order in 90 to 365 days.
Replenishment due
Overdue against their own average gap between orders.
Post-purchase
Ordered in the last 14 days. Follow up as service.
Anonymous on-site
No account, five product views in a week. Gets an on-site offer.
Abandoned cart
Added to cart, never ordered. Three reminders, then stop.
2% spot drop
A metal falls 2% below its 30-day high. Alerts proven dip-buyers.
Payment recovery
Payment failed, never bought. Support helps, no discount.
What's nextin build
Auto-dispatch stays off until a holdout proves incremental lift. SMS and bulk email are in build.
Stella never sends anything you haven't seen
That's not a setting. Out of the box, the system cannot send at all.
-
A human approves everything, by default
Auto-dispatch and sending both ship off. Every recommendation waits in a queue, with its draft, reasoning and confidence score, for a marketer to approve, edit or skip.
-
It never learns your customers' names
The model sees a hashed profile, CUST-4A3F91, and drafts with a [First name] placeholder. Real details are looked up only at send time. Stella's database holds no name, email or phone.
-
It cannot damage your production data
A query layer rejects any write or DDL statement before it reaches the server, on top of a read-only login. Everything Stella writes goes to one separate database.
-
Four independent brakes
Opt-outs and STOP. A 24-hour gap between messages. A 72-hour dedupe and a cap of three a week. Plus a per-scenario cooldown before anything is even drafted. Every block is logged with its reason.
A quarter of the "visitors" in the raw analytics were bots.
Before Stella can act on behaviour, it has to know which behaviour was human.
Only 1.4% of them ever identify themselves, which is why the on-site channel exists. Snapshot: 2026-06-17.
Three moments Stella catches that a weekly campaign can't
-
The $80,000 customer who quietly left
Over $5,000 spent, then nothing for a year. Claude drafts a personal win-back that leads with service, not a discount.
-
The 2% drop at 11:40 on a Tuesday
Silver falls below its 30-day high. Stella already knows who buys the dips, from their own order history, and alerts them while it's live.
-
The 1.6 million people you can't email
98.6% of real visitors never identify themselves. View five products in a week and Stella writes an offer the website shows you.
Applied AI with the unglamorous parts done right
Built on the Claude API, with the engineering that makes a model dependable in production.
- Schema-validated output
- 12 actions, two escape hatches
- De-identified profiles only
- Cost stays flat as the list grows
- Every call logged in USD
- No hype, no fake scarcity
- Holdings privacy in the prompt
- Prompts versioned in the UI
- Reasoning written for reviewers
Want this for your business?
Stella was designed, built and is operated by Realisier Labs. If you're sitting on customer data you're not acting on, let's talk.
Talk to Sachin