What AI Support Actually Costs You
Jennifer Tee7 min read
On this page
Most AI chat support services use bundled pricing. You pay per resolution, conversation, or ticket, and the AI is included. You never think about tokens or which model ran. Published rates run from $0.75 per resolution at Help Scout through $0.99 per outcome for Intercom's Fin to $1.50 per automated resolution at Zendesk. Simple, and that simplicity is the whole pitch.
Finding those numbers is harder than it should be. Zendesk's own pricing page says only that AI agents are "included in every Suite and Support plan, with pricing based on the successful outcomes they deliver". The $1.50 appears in a blog FAQ instead. Before you compare anything, budget for the cost of finding out the cost.
But that simplicity isn't free, and the more volume you run through it, the more you overpay. We think there's a better way to do this: paying your AI provider directly, from an account you hold. You connect your own OpenAI, Anthropic, or Gemini account and pay the provider directly for what you actually use. Optlo handles everything around that: the agent, the inbox, routing and workflows, the tools your team needs to run support day to day. We don't sit in the middle of the token bill.
What the model actually costs
A support conversation is a handful of messages with your help content attached, so call it 10,000 tokens: eight messages at roughly 1,000 tokens in and 300 out. At today's prices that works out to somewhere between $0.02 on Claude Haiku 4.5 and $0.10 on Claude Opus 5.
That is a deliberately cautious figure. Anthropic's own worked example for support tickets puts a conversation at about 3,700 tokens and 10,000 tickets at roughly $37 on Haiku, which is under half a cent each. Our number is higher because it assumes more context goes into every prompt.
Support prompts also repeat themselves, which is the cheapest discount on offer and one a bundled price never passes back to you. The instructions setting your agent's tone and the help content it quotes are identical on every ticket. Both OpenAI and Anthropic bill that repeated opening at a tenth of the normal input rate once it's cached. On a two cent conversation the saving sounds trivial. It is the difference between paying for the same system prompt a thousand times a month and paying for it roughly once. You only collect that saving when the token bill is yours.
That's the real cost of the AI doing the work. Set it next to a bundled price of $0.75 to $1.50 and most of what you're paying clearly isn't the model. On a single ticket the gap is easy to ignore. Across thousands of resolutions a month it's most of the bill.
Put some rough numbers on it and say the AI costs $0.10 per resolution and a bundled tool charges $1.00:
| Monthly resolutions | Bundled | BYO AI | You save |
|---|---|---|---|
| 1,000 | $1,000 | $100 | $900 |
| 5,000 | $5,000 | $500 | $4,500 |
| 10,000 | $10,000 | $1,000 | $9,000 |
| 20,000 | $20,000 | $2,000 | $18,000 |
Your own numbers will be different, since resolution rate, model choice, and context size all move it around. Model choice moves it most, and the OpenAI key walkthrough and Gemini key walkthrough both chart what the tiers cost before you commit to one.
The problem compounds
Two things drive the gap, and one of them grows with time.
The first of them is margin, because a per-resolution price has the model cost baked in plus a multiple on top. That's reasonable for the vendor, but it means you're paying for intelligence at retail when you could buy it at cost.
The second is the one that stings later, because model prices keep falling while bundled prices do not. When a provider cuts token costs in half, that saving flows to whoever holds the provider relationship, and in a bundled setup that isn't you. Your per-resolution price stays put while the thing it's priced against gets cheaper, so the gap between what you pay and what the AI costs grows on its own over time.
What are you actually being billed for?
"Per resolution" isn't a standard unit, and the definition is where the money hides. Intercom charges $0.99 per outcome, which means a conversation Fin closed on its own. Freshdesk's Freddy AI Agent bills per session instead. Its pricing page defines a session as "a unique interaction between an end-user and an AI Agent". For email that means "a 72-hour window starting from the customer's first email", where "all AI responses within that window count as a single session, regardless of the number of replies". Sessions are billed whether the agent resolved anything or not, at $49 per additional 100 once you pass the 500 a plan includes.
Zendesk counts differently again, judging an automated resolution by a quiet period of typically 72 hours with no customer reopening the conversation. The same 72 hours that defines a billable session at Freshdesk is the thing that proves a resolution at Zendesk. One clock bills you regardless of outcome and the other bills you only when the customer stays away.
That difference matters more than the headline price. A vendor charging $0.49 per session that bills every engagement can cost you more than one charging $1.50 per resolution, if your agent starts far more conversations than it finishes. Before you compare two per-conversation prices, check what each one counts, and whether a conversation the AI handed to a human still shows up on the invoice.
The two layers you're actually paying for
It helps to split an AI support bill into two layers. There's the AI layer: the model reading the question and writing the reply. And there's the platform layer: the inbox, seats, routing, workflows, and everything else your team touches to run support. Bundled tools fold both into one number, which hides the fact that they mark up each one.
Take Intercom as a concrete example. On their Advanced plan with Fin AI, say you run 1,000 conversations a month with 3 people on the team, and Fin resolves 70% of those conversations on its own. Both Intercom columns come from its published pricing, and the token figure from OpenAI's:
| Layer | Intercom | Optlo |
|---|---|---|
| AI layer | $693 for 700 Fin resolutions × $0.99 | ~$17 for GPT-5.4 Mini tokens, billed by OpenAI at cost |
| Platform layer | $255 for 3 seats × $85 | $59 flat Growth plan (3 seats included) |
| Per month | $948 | ~$76 |
$948 → ~$76
Same 1,000 conversations, same three-person team, same 70% resolution rate, and a twelfth of the monthly bill.
Both of the layers move at once. The AI layer drops from $693 to about $17 because you pay OpenAI's published token rate instead of a per-resolution price with margin folded in. The platform layer drops from $255 to $59 because you're on a flat plan rather than paying $85 per seat. Annual plans widen the gap further: Optlo includes two months free, so it works out to about $790 a year against roughly $11,376 with Intercom, a saving of more than $10,500.
Your own inputs will shift the totals, since more seats, a pricier model, or a higher resolution rate all move it. The shape holds either way: two layers, each cheaper when there's no middleman on either.
The upside of bringing your own AI
When you own the AI account, a few things change beyond the monthly number.
- You pay the provider's published rate with nothing added on top, so what you spend is the actual model cost rather than a number a vendor set.
- You can pick the model per job. Routing "what are your hours" and a billing dispute through the same expensive model is a waste; cheap model up front, reasoning model only when the question earns it.
- When providers drop their prices, you get the cut the same day, no renegotiation.
For most teams the savings get attention first. The control is what they end up valuing.
None of that arrives free of effort, and a cost article that pretends otherwise isn't much use. Somebody has to own the provider account, store the key properly, and notice when spend moves, which is work a bundled plan quietly absorbs on your behalf. For a small team with nobody technical nearby, that overhead can genuinely outweigh the markup it saves. The calculation turns on volume more than principle: at a few hundred conversations a month the convenience is worth paying for, and at a few thousand the markup stops being worth absorbing.
See where you land
The savings depend on your volume, your resolution rate, and which models you run, so the only number that really matters is yours. Put your own volume and model against Intercom, Zendesk, and the rest and see where the gap lands for your setup.
Common questions
How much does AI customer support cost per resolution?
Two numbers answer that, and they are far apart. The tokens for a support conversation run roughly $0.02 to $0.10 depending on the model, while published bundled prices run from $0.75 per resolution at Help Scout to $1.50 per automated resolution at Zendesk. The gap between those two figures is margin on the AI layer, and it is most of what you pay.
What does an AI support conversation actually cost in tokens?
Around $0.02 to $0.10, assuming a conversation of roughly 10,000 tokens across eight messages with your help content attached. The bottom of that range is Claude Haiku 4.5 and the top is Claude Opus 5. Anthropic's own worked example for support tickets is lower still, putting a conversation near 3,700 tokens and 10,000 tickets at about $37 on Haiku, which is under half a cent each.
Is bring your own AI key cheaper than per-resolution pricing?
On the AI layer, almost always, because you pay the provider's published rate with no markup. On the total bill it depends on your volume and your seat count, and anyone promising a guaranteed saving has not seen your numbers. The worked example in this article puts 1,000 conversations a month at three seats at $948 with Intercom against roughly $76 running on your own key, but your inputs will move that.
Why do AI support tools charge per resolution?
Because it is a simple number to forecast and a comfortable one to sell. A single price per resolved conversation hides the split between what the model cost and what the vendor added. That's fine while your volume is low and the convenience is worth paying for. It stops being fine when the markup grows with every conversation and you have no way to tell how much of the price was the intelligence.
Does the gap between token cost and per-resolution pricing get bigger?
Yes, on its own, without anyone changing a contract. Model prices keep falling and per-resolution prices do not, so every provider price cut widens the distance between what you are charged and what the work costs. Under a bundled agreement that saving improves the vendor's margin. When you hold the provider account it lands on your next invoice instead.
How much of an AI support bill is the platform rather than the AI?
Usually most of it, which surprises teams who assume the intelligence is the expensive part. Splitting a bill into the AI layer and the platform layer tends to show model usage as the smaller line once you are paying the provider directly. Seats and plan fees carry the rest. That is the argument for seeing the two numbers separately: they have different levers, and a blended per-conversation price gives you neither.
