What the agents may do, and what waits for you

1What do the agents do?

Seven jobs run on a schedule, each named the way the product names it:

• A morning brief before service: yesterday’s numbers and today’s watch-list, by email and in the app. briefing.morning
• An end-of-day report: revenue, orders, the payment split and the day’s recipe-versus-actual variance. briefing.evening
• Low-stock alerts every hour: anything at or under its reorder level becomes a named alert. It orders nothing. inventory.stock_alerts
• A fraud scan at night: voids, discounts and reprints checked for the patterns that mean somebody is helping themselves. fraud.patterns
• Menu health every Monday: which dishes earn their place and which have quietly stopped paying. pricing.menu_health
• A device check every hour: a stalled kitchen screen or printer, raised before the cook notices. tech.device_health
• A tally of the agents’ own runs, and what they cost to run (the 1st, 5:00 am). At launch it is not shown or sent to anyone. proof.monthly_return

Two more exist and are switched off for the launch: the agent that drafts a purchase order from a shortfall (inventory.auto_orders) and weekly settlement reconciliation (reconciliation.weekly). The agents’ work comes with Growth, at ₹1,999 a month billed yearly.

2What do they read, and where does it go?

Only your outlet’s own records: sales, stock against recipes, voids and discounts, the kitchen’s timings and the devices on the counter. They work inside the same database wall as the rest of the product, so no other restaurant’s data is visible to them.

Most findings are rules our own code runs over your figures. The morning brief is written by a large language model from a third-party provider: to write it, the model is sent your name and that day’s figures, such as sales, top dishes and stock, which our code has already computed. If the model cannot be reached, the brief is written from a template instead, and the end-of-day report always is. CountStand trains no model on your data.

3What always waits for a person?

No agent moves money, changes a price or sends a purchase order without a person. The agents have no tool that pays, refunds or transfers money; for them those tools do not exist. The two that could raise a purchase order or reconcile a settlement are switched off for the launch, and everything the other seven do is read-only or a message to you.

The brief in the product says it in its own words:

4How does an agent earn more freedom?

Each agent works at one of six levels, L0 to L5. It moves up only when its own record clears the gate for the next step, and the product decides that, not a person:

• L1 to L2: an agreement rate of 98% or more, over at least 200 runs.
• L2 to L3: an acceptance rate of 85% or more, over at least 50 runs.
• L3 to L4: an undo rate under 2%, over at least 100 runs.
• L4 to L5: a reversal rate under 0.5%, over at least 500 runs.

Anything that touches money stops at L4. There is no button that raises an agent: a raise is refused by the server. The owner can lower any agent at any time, and an agent is lowered on its own if too many of its actions are undone. A change to an agent’s model, instructions or tools drops every agent one level, to be earned back.

5Can I see what an agent did?

Yes. Each finding shows what it read, and every “No” you give is recorded. The agents act through the same commands as your staff, so what they do lands in the same audit log, which the owner can export. How that log is kept is on the security page.

6Questions owners ask