You wired up an AI agent, traffic picked up, and the invoice did something you did not expect: it grew. Not because you added seats or features, but simply because more people talked to your support. That is the moment most founders stop and ask the obvious question. Why is AI customer support this expensive, and why does it get worse the better my product does?
The short answer is the billing model. Most AI support tools do not charge you for the software. They charge you for every outcome the AI produces, and they call it a resolution or a conversation. It sounds fair until you do the arithmetic. This piece breaks the economics down in plain language, shows where the real cost actually lives, and explains the one pricing change that makes AI support cheap again.
What "per-resolution" actually means
The dominant way to bill AI support right now is per resolution or per conversation. Intercom's Fin charges per resolution. Tidio's Lyro sells you a bucket of conversations and meters against it. The label changes, the shape does not: you pay a fixed-ish fee every time the AI handles an interaction.
On the surface that feels reasonable. You only pay when the bot does something, right? The problem is what counts as "doing something." A resolution is usually defined as a conversation the AI handled without a human stepping in, whether or not the customer actually got their problem solved. The meter runs on activity, not on success. Read the comments under any video about these tools and the same complaint shows up: people describe a chatbot that can burn $60 on one interaction and still not solve the problem. You paid for the outcome. You did not get one.
Why the cost scales up exactly when you do not want it to
Here is the part that catches everyone. Per-resolution billing ties your support cost directly to your traffic. The more visitors you get, the more conversations happen, the more resolutions you are billed for. Your support bill is now a function of your growth.
Think about what that means at the worst possible moment. You get featured somewhere, a launch goes well, a post takes off. Traffic spikes. Those are the days you want to celebrate. Instead, your AI support meter is spinning, and a busy week turns into an invoice you did not budget for. Success literally punishes you.
Normal software does not work this way. You pay a flat fee for your CRM whether you send ten emails or ten thousand. But per-resolution AI support behaves like a tax on every customer interaction. Scale your product and you scale the bill in lockstep, with no ceiling and no way to plan around it.
- →A slow month is fine. A great month is expensive.
- →Going viral, the thing you want most, is the thing that costs the most.
- →Your worst-case bill is unbounded, because it is tied to traffic you cannot fully predict.
- →You end up quietly hoping fewer people use your support, which is a strange thing to want.
So where does the cost really come from?
To see why per-resolution pricing is so marked up, you have to look at what an AI answer actually costs to produce. When the AI replies to a visitor, it makes a call to a large language model. That call has a real, measurable cost: the provider charges for the tokens going in and the tokens coming out. A typical grounded support answer, a question plus the relevant docs plus a short reply, is a few thousand tokens. At provider rates that is a fraction of a cent to a few cents.
Read that again. The raw cost of one good AI support answer is usually pennies. So when a tool bills you a dollar or more per resolution, the gap between the pennies it costs and the dollars you pay is pure markup. You are not paying for compute. You are paying a margin on top of compute, multiplied by every conversation you have.
The fix: bring your own key (BYOK)
Once you see that the model itself is cheap and the markup is the problem, the fix is obvious. Cut out the markup. That is exactly what bring-your-own-key does. You paste in your own Anthropic key, and the AI runs on your account instead of someone else's resold capacity.
When you BYOK, every AI answer bills directly to your own provider account at cost. No per-resolution fee. No per-conversation fee. No multiplier. You pay Anthropic the few cents the tokens actually cost, and that is the whole bill. There is no meter sitting between you and the model taking a cut on every reply.
That single change flips the economics. Instead of a cost that scales up with your traffic and punishes your best months, you get unlimited AI at the provider's raw rate. A viral week costs you what the tokens cost, full stop. You can finally want more people to use your support, because more usage no longer means a scary invoice. It just means you are getting value out of a key you already control.
How muro prices AI, honestly
We built muro around this on purpose, because per-resolution billing is the thing founders tell us they are most afraid of. So muro never charges per resolution or per conversation. There is no meter that punishes a busy week.
There are two ways to run the AI, and we will be straight about both:
- →BYOK. Paste your own Anthropic key and the AI Answer Agent runs unlimited, at cost, with zero markup. You pay Anthropic directly for tokens and nothing extra to us for AI usage. This is the cheapest possible way to run AI support, and it is the one we point heavy users to.
- →Managed credits. Do not want to manage a key? Every plan includes a monthly pool of AI credits, where one credit is one action. If you run out in a big month, you can buy a top-up pack. It is metered, but it is a predictable pool you choose, not an open-ended per-resolution charge that grows with your traffic.
And the rest of muro is flat. Solo is $19/mo for two projects, Fleet is $59/mo for unlimited projects, both with a 14-day free trial and no card required. We will also be honest that there is no free-forever plan. The point is that none of those prices move because your support got busy. Your bill is a function of the plan you picked, not the traffic you got.
One more thing: don't pay for answers that don't help
The other half of why AI support feels expensive is paying for bad answers. The $60-on-one-interaction story is not just about price, it is about a bot that burned the money and still failed. A grounded agent helps here too.
muro's AI answers only from your real help docs, or a URL you point it at, and never invents facts, prices, or features. When it is not confident, or the visitor is asking about a refund, a cancellation, or anything it cannot truly handle, it hands off to a human instead of bluffing through ten more billable exchanges. Cheaper and better, because you are not paying for the bot to dig a hole and then argue its way out of it.
AI customer support is not expensive because the AI is expensive. It is expensive because per-resolution and per-conversation billing add a markup on top of cheap tokens, then charge it again on every interaction, so the bill grows exactly when your product does. The way out is to pay the provider at cost. With muro you can bring your own Anthropic key and run AI unlimited at zero markup, or use a predictable credit pool with no per-resolution fee either way. Start a free trial, or read how the AI agent answers from your docs before you turn it on.