Included AI vs Bring-Your-Own Provider Key

The decision, in one paragraph

AI Provider Choice (BYOK) gives you two routes. Included platform AI means your plan’s token allowance covers the usage and there is nothing to set up. Bring-your-own-key means you supply a provider key, usage runs through your own provider account, and that provider bills you directly. Both use the same grounding behaviour from the Grounded AI Agent & Knowledge Base, the difference is who owns the AI account, not how the answers work.

Side by side

	Included platform AI	Bring your own key
Setup	None	Select provider and model, supply and test a key
Billing for AI usage	Included in your plan allowance	Your provider bills you separately
Token allowance check	Applies	Avoided
Other plan limits	Apply	Still apply
Model choice	Platform catalogue	Your provider’s catalogue
Who to contact when the model misbehaves	Us	Your provider, for their side

Start on the included AI

For almost everyone, the right first move is to use the included AI. Build the knowledge base, pick a grounding mode, run a flow, read some transcripts. That work is identical either way, and doing it first means you are choosing a provider with real usage data rather than a guess.

The allowance differs by plan, 250,000 platform AI tokens a month on Free, rising through the tiers. Note that tokens are not messages or conversations, so an allowance is hard to translate into “how many chats” without watching your own usage for a while. Usage reporting is there for exactly that.

When BYOK makes sense

You already pay a provider. If your business has an account with committed spend, routing chat usage through it consolidates the bill.

Procurement requires it. Some organisations need AI usage to sit on an account they control, with their own terms and their own data handling agreement.

You want a specific model. Provider and model selection lets you pick from your provider’s catalogue rather than the platform’s.

Your volume is high and predictable. Once usage is large and steady, running it on your own account can be the more economical arrangement.

What BYOK does not change

This is the part people misread most often, so it is worth being blunt.

Bringing your own key avoids the platform AI-token allowance check. It does not remove any other limit on your plan: agents per site, saved bot flows, knowledge sources, knowledge pages, storage, configured retention and concurrent active sessions are all unchanged. A Free plan with BYOK is still a Free plan in every other respect.

It also does not change how answers behave. Grounding modes, retrieval strictness, refusal messages and handoff thresholds are properties of your agent configuration, not of the provider.

Setting it up safely

Three controls exist for this, and all three are worth using.

Key testing validates the connection before you rely on it.

The grounding probe checks that retrieval is returning relevant knowledge, which catches the case where the key works and the knowledge base is the actual problem.

AI error and status information tells you when the provider is failing. Provider errors also escalate the conversation to a person rather than stalling, which limits the damage while you sort it out.

If you use OpenRouter, custom model IDs require validation, test before you switch a live site over.

You can change your mind

Provider configuration is a setting. Start included, move to BYOK when the reasons above apply to you, move back if they stop applying. Nothing about your knowledge base, flows or inbox changes when you do.

A sensible sequence

Rather than deciding up front, most teams do better following the same order.

Start on the included AI. Build the knowledge base, choose a grounding mode, publish a flow, read a fortnight of conversations. This work is identical under either arrangement.

Watch usage reporting. After a few weeks you will know roughly what your volume looks like, which is the only reliable basis for a cost comparison.

Switch if the reasons apply. An existing provider account, a procurement requirement, a specific model, or volume that justifies running it yourself.

Test before you cut over. Key testing validates the connection and the grounding probe confirms retrieval is behaving. Do both on a quiet day rather than a busy one.

Operational differences after you switch

BYOK changes who you talk to when something goes wrong, and it is worth knowing that in advance.

Answer quality problems are still yours and ours. They come from the knowledge base and the grounding configuration, not the provider. Availability problems belong to your provider, and AI error and status information will show them. Billing questions about token usage go to the provider directly.

Two things worth setting up on day one either way. Keep an eye on usage reporting so a limit does not arrive as a surprise mid-week. And make sure your escalation path is configured, because provider errors route conversations to a person, and that is the behaviour you want when a model is unavailable.

Read next: what grounded AI means for website chat, or the pricing page for what each plan includes.