Where should your agent’s auto-approve limit sit?

Your AI agent issues refunds, pays invoices or buys things. Say how big they usually are, how often it gets one wrong and what a person’s time costs, and see the limit where approving by hand stops paying for itself, with the rules that set it. No sign-up, and nothing leaves this page.

1Your numbers

What the agent does (each starts with example numbers: replace them with yours)
How big they are
Mistakes
actions
%
%
A person's time
min
EUR
Your limits
EUR

Counted as if a wrong one were lost in full. The suggested limit never goes above this.

EUR

Past this in a day, the rest wait for a person.

2Where the limit should sit

Let refunds up to €100.00 run on their own

Let refunds up to €100.00 run on their own

About 65 a month wait for a person: 2.2 hours of their time. Expected cost of reviews and mistakes: €263 a month, against €1,143 when every one waits.

On an average day about €1,083 runs on its own. Past the daily total of €2,000.00, the rest wait for a person.

On its own up toTotal a monthWait a monthPerson’s timeExpected loss
Nothing, all wait€1,14383527.8 hours€1,113€30
€20.00€64643514.5 hours€580€66
€50.00€3691856.2 hours€247€122
€100.00Suggested€263652.2 hours€87€176
€250.00€2531530 minutes€20€233
€1,000.00Above your cap€30000 minutes€0€300

An estimate, not a promise: each row counts at its top amount, and a bad month is worse than an average one. The lowest limit within ten percent of the cheapest is suggested.

The rules that set this limit

Add them under Rules in VariantGate, or send them to the API signed in as a workspace owner. Above the limit, and for any amount that is missing or in another currency, a person decides. Limits and daily totals need Pro or above; on Free every request waits for a person, which is stricter, never looser. Turn this into a written policy

[
  {
    "action": "refund.*",
    "effect": "allow",
    "limit_amount": "100.00",
    "limit_currency": "EUR",
    "budget_amount": "2000.00",
    "budget_currency": "EUR",
    "budget_period": "day"
  },
  {
    "action": "refund.*",
    "effect": "ask",
    "limit_amount": "100.00",
    "limit_currency": "EUR"
  }
]

How the calculator works

Every limit trades a person’s time against mistakes. Set it at zero and someone approves every refund; set it high and the agent’s mistakes go out unchecked. The calculator prices each possible limit both ways for a month and adds them up.

  • Time spent

    Everything above the limit waits for a person. Its count, times the minutes one approval takes, times what an hour costs.

  • Mistakes let through

    Everything up to the limit runs on its own. Its amounts, times how often the agent gets one wrong, times the share you would not get back.

  • Mistakes missed

    A person does not catch every wrong one either. Above the limit, the share they miss still costs money.

  • Your worst case

    The suggestion is the lowest limit within ten percent of the cheapest, never above the largest amount you let run unchecked, and never above the daily total.

What a limit alone does not cover

A limit judges one amount at a time. An agent that splits a large refund into small ones stays under it, which is what the daily total is for: past it, even small amounts wait. Some actions should not run on their own at any size, such as paying a new bank account for the first time; give those their own action name and a rule that always asks. More in spending limits for AI agents and risk tiers for agent actions. Once the agent runs, the record shows how often a person actually changed or refused what it asked, which is a better number for “one wrong in every” than any guess.

Put the limit to work

Free for one person and 100 requests a month, with no card, where every request waits for a person. Limits and daily totals come with Pro: then everything above the limit waits for a yes in email, Slack, Teams, Discord or a phone notification.

Questions

Do you see the numbers I type?

No. The calculator runs in your browser and sends nothing anywhere. A link to your numbers keeps them after the # sign, which browsers never send to a server.

Where do I get the numbers?

Start from your own records: last month's refunds, payouts or purchases, grouped by size. For how often the agent gets one wrong, use what you found when you checked a sample of its decisions by hand, or a cautious guess such as one in a hundred, and change it once the record shows real numbers.

Why not just take the cheapest limit?

Because an average hides the bad day. A limit that is cheapest on average can still let one mistake cost far more than you can accept, so the suggestion never goes above the largest single amount you are willing to let run unchecked, and a daily total caps how much can run on its own in a day.

How does it stay on the cautious side?

Every row counts at its top amount, so automatic amounts are priced as large as they can be. Past the daily total the rest is counted as waiting for a person, as the rules make it. The lowest limit within ten percent of the cheapest is suggested, and any amount that is missing or in another currency waits for a person, so leaving it out never gets an agent past the limit.

What do the rules do?

An allow rule up to the limit, with the daily total if you gave one, and an ask rule above it. In VariantGate, when rules disagree the most cautious wins, so a broader rule written later cannot let a larger amount through. Limits and daily totals need the Pro plan or above.