What does model-agnostic mean for a customer support agent?

Model-agnostic means the platform is not tied to one large language model. You can run your agent on different providers, switch when a better or cheaper model ships, and route different tasks to different models. For support teams this protects you from a single vendor's model regressing, raising prices, or falling behind, because you can change the model without rebuilding the agent.

Can you bring your own fine-tuned or open-source model?

On some platforms, yes. Cognigy supports connecting a privately hosted or fine-tuned model, and Voiceflow supports bringing your own model, including open-source and fine-tuned ones, at the Enterprise tier. This matters when a general-purpose model handles your domain poorly, or when a data policy requires the model to run on infrastructure you control. Confirm the exact tier and deployment before you sign.

Does letting you choose the LLM remove vendor lock-in?

Only partly. Model choice removes model lock-in, but three other kinds remain: data lock-in (your transcripts and knowledge trapped in the vendor's system), suite lock-in (the agent only runs inside a larger platform you also buy), and pricing lock-in (a per-resolution meter that ties your bill to your success). Evaluate all four, not just whether you can pick the model.

Which AI Customer Support Platforms Let You Choose Your Own LLM

"Choose your own LLM" has become a line on enterprise support RFPs. The question sounds technical, like it belongs to the engineering team. It is really a procurement question, and a strategic one. Whoever picks the model that answers your customers also shapes your cost, your data exposure, and how hard it will be to leave in two years.

So the honest version of the question is not "does this platform run GPT or Claude." A few enterprise platforms now do. The real question is how much of the stack you still control once the contract is signed: the model, the data, and the exit.

Here is which enterprise AI customer support platforms let you pick the model, how they differ on the things that matter more than model choice, and what to check before you commit.

Short answer. Among enterprise AI customer service software platforms, Voiceflow, Cognigy (now part of NiCE), and Kore.ai let you choose the large language model your support agent runs on, across providers like OpenAI, Anthropic, and Google. Fin (formerly Intercom, now part of Salesforce) does not; it runs on Intercom's own proprietary models. Voiceflow goes furthest on control: you can bring your own model, including open-source or fine-tuned ones, at the Enterprise tier, and the agent runs on your stack and routes resolutions into the helpdesk you already use, so the transcripts and the audit trail stay yours.

Choosing Your LLM Is Really Three Decisions

When a buyer asks "can I use my own LLM," they are usually trying to get at three separate things without naming them.

The model. Which large language model actually generates the answer your customer reads. Provider choice, meaning OpenAI versus Anthropic versus Google, matters for quality, latency, and cost. Bringing your own model, an open-source or fine-tuned one, matters when you have a domain the general models handle poorly, or a policy that says customer data cannot leave infrastructure you control.

The data. Where the conversation, the knowledge base, and the resolution log actually live. A support agent reads your customers' messages and your internal docs on every ticket. Whose systems hold that, and do you keep a clean audit trail of what the agent saw and did.

The exit. What it costs to switch. If the model is locked, the data is trapped, and the meter runs on every resolution, leaving is expensive by design. Model choice is the first sign of whether a vendor built for your control or for your lock-in.

Read that way, "let me choose the model" is the visible tip of a bigger question: how much of your support stack do you still own after you buy.

Which Platforms Let You Choose Your Own LLM

Here is how the four platforms most often shortlisted for enterprise support compare on those three decisions.

Platform	The model	The data	The exit
Voiceflow	OpenAI, Anthropic, Google, or AWS Bedrock; bring your own open-source or fine-tuned model at the Enterprise tier	Runs on your stack; you keep the transcripts and the audit trail	Routes into the helpdesk you already run; platform subscription, not per-resolution
Cognigy (now part of NiCE)	Model-agnostic across OpenAI, Anthropic, Google, and AWS; private or fine-tuned models supported	Lives inside the Cognigy environment	Part of the NiCE contact-center suite
Kore.ai	Model Hub across OpenAI, Azure OpenAI, and AWS Bedrock, alongside its own XO GPT	Lives inside the Kore.ai platform	Adopt the broader console and toolset
Fin (formerly Intercom, now part of Salesforce)	Intercom's own proprietary models; no model choice	Lives inside Intercom, now Salesforce	Priced per resolution; folding into Agentforce

Two things stand out. Model choice alone no longer separates the field: three of these four let you pick a provider, so if "bring your own LLM" is the only box on your scorecard, you will not learn much. And the platforms diverge sharply on the other two columns, data and exit, which is where the real decision sits.

How Each Platform Handles Model, Data, and Switching

Cognigy, now part of NiCE. Cognigy is genuinely model-agnostic. You can assign models from OpenAI, Anthropic, Google, and AWS to different agent jobs, fall back between them, and connect a privately hosted or fine-tuned model. On the model axis it is a real peer. The trade is the platform it sits in. Cognigy is a contact-center suite, and since the NiCE acquisition it is one part of a larger CCaaS stack, so you get the model flexibility inside their environment. Our Cognigy review has the fuller picture.

Kore.ai. Kore.ai moved from a single proprietary model, its XO GPT, to a Model Hub that runs models from OpenAI, Azure OpenAI, and AWS Bedrock alongside its own. So model choice is there. Kore.ai is also a broad, feature-dense enterprise platform, and that flexibility lives inside a console and toolset you adopt more or less wholesale. Our Kore.ai review covers where the breadth helps and where it slows teams down.

Fin, formerly Intercom, now part of Salesforce. Fin is the outlier on model choice. It runs on Intercom's own proprietary models, and you do not select or swap the LLM. It is a closed, managed product, priced per resolution, and it is folding into Salesforce's Agentforce after the 2026 acquisition. If model control is a requirement, Fin is the one platform on this list that does not offer it, and the acquisition adds roadmap and stack questions on top.

Voiceflow. Voiceflow is model-agnostic across OpenAI, Anthropic, Google, and AWS Bedrock, and at the Enterprise tier you can bring your own model, including open-source or fine-tuned ones. What separates it from the others is not the length of the model list. It is that the agent runs on your stack and routes resolutions into the helpdesk you already run, instead of asking you to move into a new suite. You keep the transcripts and the audit trail. More on that below.

Get started

See how leading teams design, test, and deploy AI agents at scale.

Book demo

The Vendor Lock-In Question

This is what the model-choice question is really pointing at. Lock-in in AI support is not one thing. It shows up in four places, and a platform can be open in one and closed in the others.

Model lock-in. You cannot change the LLM. If the vendor's model regresses, gets more expensive, or falls behind, you wait for them. Fin is the clearest case.
Data lock-in. Your conversation history, your knowledge base, and your resolution logs sit in the vendor's system in a shape that is hard to export. Even with model choice, if the data is trapped, leaving means rebuilding.
Suite lock-in. The agent only works well inside a larger platform you also have to buy, a contact-center or CRM suite. The agent is the hook and the suite is the contract. That is the shape of the NiCE, Salesforce, and Kore.ai stories, in different ways.
Pricing lock-in. A per-resolution or per-conversation meter ties your bill to your success and your busiest days. The better the agent gets, the more you pay, which quietly argues against ever moving the volume anywhere else.

Model choice fixes the first one. It does nothing about the other three. A platform that lets you pick GPT or Claude but keeps your data behind its walls, only runs inside its suite, and bills per resolution is still a platform that is expensive to leave. So when you evaluate "can I choose my own LLM," keep going. Ask where the data lives, whether the agent runs without the rest of the suite, and what the meter does to your bill at three times today's volume.

Where Voiceflow Lands

Voiceflow's answer to all four kinds of lock-in is the same idea: you keep the stack.

On the model, it is agnostic. Choose OpenAI, Anthropic, Google, or AWS Bedrock, and at the Enterprise tier bring your own, including open-source or fine-tuned models. If a model regresses, or a cheaper one clears your quality bar, you make the switch yourself instead of filing a request.

On the data, the agent runs on your stack and reads your knowledge base, and you keep the conversation transcripts plus a clear record of what the agent retrieved and did. That matters for regulated buyers, and it is the difference between borrowing a vendor's visibility and owning your own. Voiceflow is SOC 2 Type 2 with PII masking, and our security and compliance guide covers what to ask any vendor on this.

On the exit, Voiceflow routes resolutions into the helpdesk you already run. It hands off to a human agent through your existing live-agent tools, with an AI-written summary and a warm transfer, and it has a Call Forward step for voice. It is not a ticketing helpdesk of its own, and that is deliberate: it sits on your stack rather than replacing it, so it does not become the suite you are then locked into. Pricing is a platform subscription rather than a per-resolution meter, so a better agent improves your margin instead of your invoice. The specifics live on our pricing page.

One boundary is worth drawing, because it is easy to blur. This piece is about whose model runs your agent. Whether a platform can actually reason through a multi-step support problem, rather than follow a scripted tree, is a separate question about architecture. Model choice and reasoning are both real requirements. They are not the same requirement, and a platform can pass one and fail the other. If reasoning is where your evaluation is, our guide to choosing an AI agent builder is the better starting point.

How to Pressure-Test Model Choice Before You Sign

Five questions separate real control from a checkbox. Send the same five to every vendor and read the answers side by side.

Which exact providers and models can I use today, and can I bring my own open-source or fine-tuned model? At which tier?
When a new model ships, do I switch to it myself, or wait for you?
Where do the transcripts, the knowledge base, and the resolution logs live, and can I export all of it?
Does the agent run on my existing helpdesk, or only inside your suite?
How am I billed, and what does the bill do if my volume triples?

The platform that answers all five cleanly is the one built for your control rather than your renewal. For the wider evaluation around these, our pillar on AI customer service software and the read on moving a pilot into production go deeper.