01How it works
The life of a single request
Between what you say and what changes in your store, there are five checkpoints.
- 01
Understand
Breaks your request into concrete operations and parameters, and asks when unsure.
- 02
Simulate
Runs a dry-run preview of the effect on price, margin and stock.
- 03
Check
Compares against risk levels, your per-action rules and your operating standards.
- 04
Act or ask
Acts when permitted; otherwise asks you in the chat and keeps it in your approvals for 24 hours.
- 05
Record
Logs every call and the before-and-after of every field.
02Risk levels
Four levels decide how far the AI can go
Every store operation is assigned a risk level by design. You can then set individual actions to “Run automatically”, “Needs your approval” or “Forbidden”.
| Level | What it means | Examples | Agent default |
|---|---|---|---|
| L0Read | Looks at data, changes nothing | Orders, reports, stock lookups | Runs automatically |
| L1Draft | Creates things customers can’t see yet | Product drafts, copy, campaign plans | Runs automatically |
| L2Customer-facing | Changes customers will see or feel | Publishing, pricing, stock, shipping, promotions | Needs approval (automatic on Autopilot) |
| L3High risk | Involves money, access or the rules themselves | Refunds, keys, members, standards | Always needs approval |
On Growth, L2 actions always need your approval; switching them to automatic is part of Autopilot.
03Operating standards
Turn your business limits into hard rules for the AI
Operating standards are the limits the agent checks before every action. When something crosses a line, it isn’t quietly carried out; it comes to you.
Price & margin
- Margin target
- e.g. 35%
- AI price-change cap
- e.g. 10%
- AI stock-change cap
- e.g. 20%
- Protect margin at checkout
- On by default
- Minimum price
- Per variant
Inventory
- Slow-moving after
- 90 days by default
- Monthly holding cost
- % of unit cost
- Expiry warning
- 30 days by default
- Expired lots
- Write off or keep
Shipping
- Ship-by time
- Business days or hours
- Advance warning
- 4 hours by default
- Order cutoff
- e.g. 15:00
- Shipping days, holidays, make-up days
- From your business calendar
These always come to you
- A price change below the margin target
- A single price change above the cap
- Lowering a minimum price
- Repricing a product with no cost recorded
- A stock adjustment above the cap
- A promotion that drops items below the margin target
04Store playbook
It learns, but never behind your back
On Autopilot the AI gets to know your store over time. It doesn’t retrain a model; it writes what it learns into a playbook you can read and edit.
Each store’s playbook is kept separately, never used to train general-purpose AI models, and stops applying the moment you delete it.
- 01
Collect signals
What you approve and reject in the chat, the edits you make to drafts, and changes in your business data.
- 02
Find the pattern
Repeated decisions become a single rule, with its source attached.
- 03
Ask you
New rules come to you in the chat and only enter the playbook once you confirm.
- 04
Work by the playbook
The AI reads the playbook before every action; change, pause or delete entries whenever you like.
05Data & account security
Your store’s data belongs to your store alone
Layered data isolation
Access policies, scoped queries, database row-level security (RLS) and foreign-key constraints, each covered by automated tests.
Least-privilege agents
External agents authorise with OAuth 2.1 and PKCE; tokens expire and rotate every 10 minutes. Agents can never manage keys, members, settings or approval rules.
Secure staff sign-in
Password, email code, Google sign-in and TOTP two-step verification, with re-verification for sensitive actions.
Complete audit trail
The activity log records identity, channel, risk and result for every call; the change log keeps field-level diffs.
Safe to retry
Every write is idempotent, so network retries never create duplicate orders or changes.
Never used for training
Your business data serves your store only and is never used to train general-purpose AI models.
Get started
Want to see how the agent is kept in check?
In the demo we’ll set up operating standards and approval rules together, so you can see every checkpoint.