Bring Your Own AI Key: Pay Providers Directly
Jonas Budelmann11 min read
On this page
Bring your own AI key means your support tool runs on an AI provider account that you own, and the provider bills you directly for what the models use. The software vendor charges you for the platform around that: the agent, the inbox, the workflows, the reporting. Nobody sits between you and the token bill, so you pay published rates instead of a resale price. That gives a support team two things it didn't have before. When model prices drop the savings are yours rather than somebody else's, and you can finally see where the money goes.
The pitch for bundled AI support is that you never think about any of this. That's genuinely worth something when you're small. It stops being worth much once your volume is high enough that the markup outweighs the convenience, and what AI support actually costs you runs the numbers on where that crossover lands.
What does bring your own AI key mean in customer support?
Bring your own AI key means the agent that answers your customers runs on an account with your name on it, not the vendor's. You create an account with OpenAI, Anthropic, or Google, generate an API key, and connect it to the support platform. From then on the platform sends requests to the model using your credentials, and the provider invoices you. The three consoles work differently enough that we wrote the walkthroughs separately: how to get an OpenAI API key, how to get a Claude API key, and the Gemini one, where a free tier complicates the choice.
The term is borrowed from a much older idea in enterprise software, where "bring your own key" describes customer-managed encryption keys for cloud storage. That version is about who can decrypt your data. The AI version is about who owns the account, and therefore who gets the bill. Both meanings are alive and well, and vendors will use them in the same breath, so pin down which one somebody means before you compare anything.
Software outside customer support adopted the AI meaning first, and developer tools got there earliest because their users already had provider accounts sitting around. Coding assistants, editor plugins, and API gateways have shipped BYO support for a couple of years now. Customer support software has been slower, which is why most of what turns up in a search is written for developers picking an IDE plugin rather than for a support lead choosing a helpdesk.
How does the billing actually work?
Your support conversation produces two charges on two separate invoices, and the split is clean: the model provider charges for the thinking, and the platform vendor charges for everything around it.
When a customer types a question into the chat widget, the platform assembles a request. That request carries the question, the relevant chunks of your help content, and whatever conversation history matters. It goes to your provider under your API key. The provider runs the model, returns an answer, counts the tokens on the way in and out, and adds them to your monthly bill at the rate published on its pricing page. The platform takes that answer, renders it in the widget, logs the conversation in your inbox, and applies whatever routing rules you configured.
So one invoice is model usage and nothing else. The other is seats and volume and nothing else. Your total won't always come out lower, and anyone promising otherwise hasn't seen your volume. What you do get is two numbers you can actually do something about.
This also helps the first time somebody asks you where the customer data goes. Because the request runs through your own provider account, the terms that apply are the ones you agreed to when you signed up. Whatever you set for data retention and training covers your support conversations too, and you can check those settings yourself instead of taking a vendor's word for it. That beats forwarding the question to a supplier and waiting. None of that lets you off the hook for what you send, so think about which help articles and customer details end up in a prompt.
Contrast that with bundled pricing, where a single per-conversation price covers the model and the platform together. Intercom prices Fin at $0.99 per outcome, on top of seats that run from $29 to $132 per person per month depending on plan. That's a perfectly sensible product, and an easy number to forecast, which is the point of it. What it won't tell you is how much of that $0.99 was the model and how much was margin. Without that split, you can't tell whether a cheaper model would have saved you a penny.
The two arrangements differ on five things that matter once you're running real volume:
| What changes | Bring your own key | Bundled per outcome |
|---|---|---|
| Who bills you for the model | Your AI provider, at its published rate | Your support vendor, folded into one price |
| Who picks the model | You, and you can vary it per workflow step | The vendor |
| When provider prices fall | Your next invoice is smaller | The vendor's margin improves |
| Cost visibility | Token usage itemized in your provider dashboard | One blended number per conversation |
| Who manages the key | You, including rotation and spend limits | Handled for you |
What holding the key buys you
Model choice, price transparency, and the freedom to leave are what you're really buying. Everything else people claim for BYO tends to reduce to one of those three.
Model choice is the one that surprises people. The published gap between the cheapest and most expensive models is enormous, and you can only use it if you hold the account.
That spread is the argument for routing rather than standardizing. Sending "what are your opening hours" and a disputed refund through the same expensive model wastes money on the first without helping the second. When you hold the account you can put a cheap model on the routine traffic and save a reasoning model for the steps that earn it. Optlo lets you mix and match models from those three providers across different parts of a workflow, and that per-step routing is where the savings show up.
Output tokens are where the real spending happens, and the ratio is consistent enough to plan around.
Because output is the expensive half, the length of your agent's replies is a budget decision rather than a style preference. A tool that answers in three tidy sentences costs less to run than one that produces a wall of text for every question. You can only tune that if you can see the split, and you can only see the split if the invoice is yours.
Price cuts are the second thing that reaches you directly, and they arrive without anyone consulting you. Anthropic launched Claude Sonnet 5 at an introductory $2 and $10 per million tokens, told everyone it would rise to $3 and $15 in September 2026, then canceled the rise and made the launch rate permanent. Under a bundled contract that kind of decision improves the vendor's margin, and your invoice stays exactly where it was. Under BYO, your next invoice is simply smaller.
The unglamorous gain is that you can leave. Your provider account, your usage history, and your spending controls belong to you rather than to the support tool. Switching platforms means reconnecting a key, not renegotiating the AI relationship from scratch.
50x
The gap between the cheapest and most expensive published input rates across the three major providers. Once the account is yours, that whole range is yours to pick from.
The two invoices, and what sits on each
The bit that surprises most teams is the ratio. Model usage on a support conversation is tiny, and the platform fee is usually the bigger line. The expensive part of bundled AI support was rarely the intelligence.
We've done the full comparison elsewhere rather than repeat it here, including how a bundled per-resolution price stacks up against tokens at cost across different monthly volumes. If you want the worked example, the true cost of AI support is the piece to read.
If you're on a bundled plan today, the more useful exercise is running your own numbers instead of reading ours. Ours takes your monthly conversations, seat count, resolution rate, and choice of model. It puts your total next to what Intercom, Zendesk, or Tidio would charge for the same work. Most of the gap turns out to be the AI layer, which is the half you stop paying a markup on.
The thing to notice is that a per-resolution price sits still while the model cost underneath it keeps falling. Every provider price cut widens the gap between what you get charged and what the work costs. None of that widening shows up on a contract you already signed, and holding the account is what puts you on the right side of it.
Two bills also give you two levers. If your model spend looks high, you switch model or tighten up the replies. If your platform spend looks high, you change plan or trim seats. With one blended number you get neither, just a total that climbs with volume.
Where does bring your own AI key go wrong?
Bringing your own key moves real work onto your side of the line, and teams who skip it get caught out.
Key hygiene is the part people underestimate. An API key is a credential that can spend your money, so treat it like any other production secret. Store it somewhere sensible, rotate it when people leave, and keep it out of shared documents. If nobody on the team owns that job, it stays nobody's job right up until the day it matters.
Spend control is the one that actually bites. Every major provider gives you usage limits and billing alerts, and every one of them ships with defaults looser than you'd choose. A workflow that retries too eagerly can burn through money fast, and unless you've set a cap the invoice is how you'll find out.
Somebody also has to own the account. Setting one up, generating a key, and reading a usage dashboard isn't hard, but it does assume there's a vaguely technical person somewhere near the support team. Plenty of small businesses simply don't have that person, and for them a bundled tool is a fair buy even with the markup on top. The convenience is real, and that's what they're paying for.
None of this makes the models themselves cheaper. You pay the same published rate anyone else pays, and what changes is that nobody marks it up before passing it along.
How do you tell a real BYO offer from a relabeled one?
Check who receives the money for model usage, because that single question separates the two. If the answer is anyone other than your provider, you're looking at resale with better marketing.
- The provider bills you directly, on your own invoice, and you can see that invoice without asking the vendor for a report.
- You generated the key inside an account you control, rather than receiving a key from the vendor.
- You can name the specific models running, and change them, rather than being told "our AI" handles it.
- Usage and cost data live in your provider dashboard, not only in a vendor summary screen.
- Nothing on the platform bill is priced per resolution, per conversation resolved, or per credit consumed.
Two setups fail this test while sounding like they pass. One is the vendor who takes your key and still charges per resolution on top, so you get the setup work of BYO with the pricing of a bundle. The other is a credit system, where you buy vendor credits and spend them against models. Credits are a currency the vendor sets the exchange rate for, and that rate is a markup wearing a different hat.
There's a quick way to catch both. Ask what happens to your bill the day your provider announces a price cut. With real BYO the answer is immediate and obvious. With the other two you'll get a paragraph.
Who should hold the key
If your support volume is high enough that the AI line on your bill is noticeable, holding your own provider account is close to a straightforward win. You pay published rates, you send the simple jobs to cheap models, and price cuts land without a negotiation. If your volume is low and nobody wants to own a credential, a bundled tool is a fair choice, and the markup is what you pay for not thinking about it.
Most people land somewhere in the middle, and they tip toward BYO sooner than they expect, because the crossover isn't really about volume on its own. It comes down to whether you ever want to answer the question "what is our AI actually costing us, and why". A blended per-conversation price can't tell you. Two separate invoices can, and once you can see the split you usually spot something worth changing.
That split is why we built Optlo the way we did. You connect your own OpenAI, Anthropic, or Gemini account, and we charge for the software rather than the tokens. The seven day trial includes AI access, so you can watch an agent handle real tickets before you open a provider account at all. If you want to see that running on your own help content, the 7-day trial includes AI usage so you can test before wiring anything up. Bringing your own key is what keeps the agent live once the trial ends.
Common questions
What does bring your own AI key mean?
Bring your own AI key means the software you're using runs on an AI provider account that you own rather than one the vendor owns. You generate an API key inside your own OpenAI, Anthropic, or Google account and connect it to the tool. The provider then bills you directly for model usage at published rates, and the vendor bills you separately for the software itself.
Is BYOK the same as bring your own encryption key?
No, though the acronym is shared and the confusion is common. In cloud storage and security, bring your own key describes customer-managed encryption keys, which is about who can decrypt data. In AI tooling it describes who holds the model provider account, which is about who gets billed and who picks the model. When a vendor uses the term, it's worth asking which one they mean.
Does bring your own AI key actually save money?
It removes the markup on model usage, which is the biggest savings for teams running real volume. You still pay the provider's published rate, so the models cost the same as they would for anyone else. What you save is the multiple a vendor would have added on top, plus whatever you claw back by sending routine work to cheaper models.
Which AI providers can I connect?
Optlo supports OpenAI, Anthropic Claude, and Google Gemini today. You can connect more than one and use different models for different parts of a workflow, which is how most teams end up balancing cost against answer quality. Support for further providers is planned.
What happens if my API key stops working?
The agent can't reach the model, so replies stop until the key is fixed. Usually the key itself is fine, and what's happened is that you've hit a spending cap you set months ago and forgot about. That's the argument for setting those caps generously and adding a billing alert that warns you before you reach one. Treat the key like any other production credential, with a named owner and a rotation plan.
Do I need a technical person to set this up?
You need somebody comfortable creating an account and copying an API key, which is closer to setting up a payment processor than to writing code. The provider dashboards are built for developers, so the vocabulary can be unfamiliar at first. Past that first setup there's very little to do, since the key stays where you put it and the billing alerts tell you if anything needs attention. Whoever already administers your other billing tools can usually take it on.
