What every AI buyer should set as defaults

When you sign up for Watchpost, your account gets a set of default rules. They are calibrated for a person who has never had an AI agent buy anything for them and is rightly cautious. This post explains what each default does, why we picked the value we did, and when it's safe to relax it.

The defaults, briefly

• Per-purchase cap: $50.
• Per-day cap: $200.
• Per-month cap: $1,000.
• Ask-me-above: $25.
• Min trust score: 40.
• Max listing risk: 60.
• Allow recurring without asking: off.
• Forbidden categories: none preset.

Why $50, $200, $1,000

Three different jobs. The per-purchase cap exists to make sure no single mistake is very expensive. $50 is comfortably above what an agent buys for everyday convenience (books, cables, refills) and well under what counts as a meaningful purchase for most people. The day cap exists to make sure a series of small mistakes doesn't add up to a big one. $200 is a reasonable ceiling for "the agent had a productive day". The month cap exists because subscriptions and recurring services creep; $1,000 a month is a deliberate ceiling that matches what most people would notice on a credit-card bill.

Tighten any of these if your agent is mostly doing $5 purchases. Loosen any of them if your agent buys a lot of cloud compute or API credits.

Why ask-me-above $25

Most prompt-injection attacks try to push a single purchase past a meaningful threshold. $25 is low enough to catch them (the attacker rarely tries for less because the upside isn't worth the effort) and high enough that you don't get pinged on every $3 micro-purchase. If you find yourself approving a lot of $25-50 review verdicts, bump it to $50 once you trust the merchant set.

Why min trust score 40

Watchpost's trust score is 0-100. New merchants start at 50 (we don't know yet). A score of 40 means we have at least a few signals pointing wrong. Letting that through risks a listing-manipulation attack landing. Setting it any higher than 40 by default would block a lot of small but legitimate sellers.

Bump this to 60 if you only buy from established merchants. Drop it to 30 if you're okay with the agent exploring small shops. Both are fine, just be honest with yourself about which you are.

Why max listing risk 60

Listing risk is also 0-100. A clean listing scores 0. Mechanical pre-checks score medium-severity issues at 55 and high-severity at 85. Setting the cutoff at 60 means medium-severity flags pass (they could be false positives) but high-severity flags block (they're almost never false positives).

If you've ever had a confusing-but-real listing get blocked, raise this to 70. If you've ever had a sketchy listing get through, drop it to 50.

Recurring is off by default

This is the rule we are most opinionated about. AI agents are not great at noticing when a "trial" enrolls them in a $40-a-month subscription. Watchpost defaults to review-on-recurring-charge so you see the auto-renew before it sticks. Turn this off only if your agent's whole job is managing subscriptions for you.

What we don't preset

Forbidden categories. They're personal: alcohol, gambling, kids, pets, automotive, donations, etc. Add any you don't want your agent buying in. The full taxonomy lives on the rules page in your dashboard.

The honest version

The defaults are a starting point. Spend a week with them and read your verdict log; you'll find at least one rule you want to tighten and one you want to relax. The point is that you're starting with sensible numbers, not zeroes.