Before an AI bot answers anybody, somebody has to decide who owns the model key and who receives the model bill. There are two answers, they are both fully supported, and the choice is reversible — but they behave differently once you are live, in ways that are easy to miss while you are still comparing prices.
This is what actually differs.
Managed: we supply the key
On a managed plan, MyLiveChat holds the provider key and bundles a reply allowance into the plan. You never open a provider console, never paste a key, and get one bill from one supplier. Plans are priced per site per month and come in three tiers, with more included replies as you go up.
The reply allowance is a real counter, not an estimate. What you have left is your included replies plus any additional replies you have purchased, minus what has been used. It is visible on the usage screen and summarised on the AI overview, and it is the number to watch.
It is also the number that stops the bot. When a managed allowance reaches zero, AI replies stop until capacity is added — the setup stays exactly as it was, the bot simply stops answering. Conversations do not vanish when this happens; they behave like a site with AI off, which means they go to your agents or to your offline message form depending on who is around. That is a survivable failure rather than a broken widget, but it is a failure your visitors can see, so the warning on the usage page is worth an alert rather than a glance.
The trial counter is separate
If the site is in an AI trial, the remaining figure you see is the trial allowance rather than the plan allowance — a separate limit with its own used-count, displayed as trial replies remaining so you can tell the two apart.
The reason to know this is the transition. Running a trial to zero and then buying a plan are two different events, and the screen tells you which one you are looking at. A trial that has run out reads as a capacity block in exactly the same way a plan that has run out does, and the fix is the same: choose a paid path, or move to your own key.
Own key: you supply the key
The alternative is to paste your own provider key into bot setup. Model usage is then billed by your provider, directly to you, at whatever rate your account is on. You pay MyLiveChat for the AI agent seat rather than for replies, priced per agent per month.
Two things follow, one good and one worth planning for.
The good one: we do not meter you. An own-key site is not counted against a managed reply budget, and there is no allowance for it to exhaust on our side. Your ceiling is whatever your provider account allows.
The one to plan for is the same fact from the other direction. Because your spend sits in your provider account, we cannot warn you about it. There is no early warning on our usage screen for a provider account running low, hitting a rate limit or having its key rotated by somebody doing housekeeping, because none of that is visible to us. The readiness panel will happily show a fully-configured site while the provider account behind it has stopped accepting requests. If you go this route, whoever owns the provider account needs to be someone who will notice.
Choosing, and changing your mind
The decision usually comes down to procurement rather than technology. If the answer to who is paying is finance, and nobody on the team wants a provider account to look after, managed is the shorter path and the reply warnings are worth having. If your organisation already has a provider account, an approved spend line and somebody who owns it, own-key keeps the model bill where your existing controls already are.
Volume matters too, but less than people expect at the start. Per site per month against per agent per month means the crossover depends on how many agents you have as much as on how many replies you send, so it is worth doing the arithmetic with your own numbers rather than reasoning from the headline price. What drives AI cost covers the volume side.
Neither choice is a one-way door. The mode screen shows which one is active, and a saved key can be removed later from bot setup without having to paste a replacement first — a small detail that matters, because the alternative would be needing a working key in order to stop using keys at all.
How the screen decides what mode you are in
One behaviour is worth knowing if the label ever surprises you. If no billing mode has been recorded for the site but a provider key has been saved, the screen treats that as own-key mode rather than as nothing chosen. That is almost always what you want: somebody who pasted a key has clearly chosen, whether or not they also clicked through a checkout.
Managed mode, by contrast, is only ever set by activating a managed plan. So the mode label is a fair summary of the state, but the definitive answer to who pays for a given reply is the billing status line, not the card that happens to be highlighted.
What to do either way
Whichever you pick, the setup after it is identical: enable replies, add real knowledge, and check what the bot says before you send it traffic. Billing decides who gets the invoice; it has no effect at all on answer quality, which is entirely a function of what you have taught the bot. If you want the AI to be good, that work happens on the knowledge base, not on this screen.
Whichever arrangement you settle on, the models screen reads your saved keys to decide what it shows, and one key silently outranks the other. Why no model is marked active on your AI models page covers that precedence, which is worth knowing before you read a missing badge as a broken setup.
Whichever mode you land in, the usage caps sit alongside it as an early-warning layer rather than as a second billing mechanism. What your AI spend caps do when you reach them explains the three of them and how to choose a number for each.
If you go the managed route and later need more than the plan includes, what happens after you buy an AI reply pack covers the top-up path and the one timing detail worth knowing before you pay.