An AI chatbot without rate limits is a public API tied to your credit card. The first person to notice will hammer it from a script, and modern frontier models can chew through hundreds of dollars in tokens in a single afternoon. Worse, the bill arrives at the end of the month, when the damage is already done and the abuse is long forgotten in your access logs.

SleekAI rate-limits before the model is called. Requests are counted per IP, per session, and per logged-in user when present, with limits stored in transients or your own table. Logged-in users from wp_users can have higher caps than anonymous traffic, and specific roles can be exempted. When a limit is hit, the bot returns a friendly throttle message instead of a 429, so legitimate visitors are not surprised.

The trickier cases are bursts versus sustained abuse and shared-IP traffic from offices or campuses. Static per-minute caps catch one without catching the other. Generic chatbots ship one global cap or no cap at all, both of which fail in production. SleekAI exposes per-window and per-day limits separately so you can shape traffic without locking out real users.

An unthrottled AI chatbot is one Hacker News post away from a five-figure surprise bill. The economics are brutal because modern reasoning models can cost cents per call, and a determined script can issue tens of thousands of calls per hour. Without a cap, the only thing standing between you and a runaway invoice is your provider's account-level spending cap, which fires after the damage.

Rate limiting moves the line of defense to your own server, where you can spot abuse early and respond before the bill grows. It also protects the user experience. When the bot is being drained by a single scraper, real visitors see slow replies or outright failures.

Per-IP and per-session caps shape traffic so that one abusive client cannot starve the rest of the audience. Trust-tier limits go further. Anonymous visitors get a sensible cap that covers normal use.

Logged-in members get a higher cap because they are identifiable and accountable. Staff get whatever they need because they are trusted. The same chatbot can serve a marketing page, a member portal, and an internal tool without redeploying.

The throttle message itself is part of the experience. A bare 429 error feels like the site broke. A polite message with a reset countdown feels like a fair queue.

That difference shows up in customer satisfaction long before anyone reads the rate-limit docs.