An AI bot reads low-stock alerts every 15 minutes and provisions new eSIM lines — with nobody pressing a button.
Every order claims a ready-made eSIM from stock. When stock hits zero, customers cannot buy. Restocking by hand meant someone watching email, then opening a carrier console to create each line.
A low-stock email can sit unread overnight or at weekends.
A person types ICCIDs and tariff details into a supplier portal, line by line.
Let a bot notice the alert and run the same safe provisioning steps, every 15 minutes.
Four ways in — one body of code. Whoever starts the work, the same rules, checks and logging apply.
The whole loop runs on its own. Each stage is one small, recoverable step.
Steps 1–3 happen by themselves when a customer's purchase drops stock below the limit.
Steps 4–5: the bot finds the alert, understands which product is low, and runs the skill.
Steps 6–7: audited tools create the lines through the same code a person would use.
To: stock-alerts@esim.net
Threshold: 10 · Status: Below limit
Action needed: provision new lines
Example alert for illustration. The email and its image are what the bot looks for next.
A background batch job in OpenAI reads the alert inbox and the alert images.
It picks out only the RED low-stock alerts and which product each one is for.
For each one it runs provision stock save. Nothing to find? It simply waits for the next cycle.
A saved set of instructions: read the alert, confirm the product and quantity, then call the right tool. The same steps run every time.
One of 67 business tools on the MCP server. It writes the same work-queue row a person's action would, so every call is logged and audited.
The worker picks up the queue every 30 minutes, asks the carrier for new SIMs and tariffs, and records the result.
Why it is safe: a new tool is a database row, not a code release, and the AI has no shortcut around business rules. It can only use the tools it has been given.
Batch scan of alerts, plus the provision skill.
67 live tools for orders, stock, queue, payments.
Writes code changes and opens pull requests.
Reviews every pull request before it can merge.
● Single-run switch — two runs never overlap
● SIM claimed first — a retry never double-allocates
● Unsure? Alarm. — it never guesses, it emails support
● Retry, then escalate — timeouts cost a retry, not a failure
● Full audit log — every tool call is traceable