Boost.ai's Presis Model Runs Classification 2.5x Faster In-House

Boost.ai says its new self-hosted classification model, Presis, is at least 2.5 times faster than the models it replaces. The company launched the model on November 10, 2026, and reported that internal testing shows the speed gap is most noticeable during peak-demand hours and in voice AI deployments, where every routing decision sits directly on the response-time path.

That last detail is the crux of the story. In a voice conversation, an AI agent can't start speaking until it knows which part of the system should answer, whether the question is in scope, and whether the answer is safe to show the end user. Presis makes exactly those calls. It also handles how completed conversations get reviewed. When those checks are slow, the whole interaction stalls — which is why a classification model that runs 2.5 times faster matters more than the raw number suggests.

What does Presis actually decide?

Boost.ai positions Presis as the traffic controller for its agent platform. The model's job covers four kinds of judgment calls in every conversation:

• Routing: deciding which part of an AI agent should respond to a given user input.
• Scope: determining whether a question falls within what the agent is built to handle.
• Safety: checking whether an answer is safe to show an end user before it goes out.
• Review: shaping how completed conversations are examined after the fact.

The company says Presis already powers those platform decisions, with a broader rollout across the platform continuing over the coming weeks. In voice deployments specifically, the routing and safeguard checks must finish before a response can begin. Classification isn't a background task there. It's the gatekeeper on the critical path, and its speed sets a floor under how quickly the agent can reply.

What are the measured performance gains?

Internal testing at boost.ai reports two headline numbers. The first is speed: at least 2.5 times faster than the models Presis replaces, with the biggest difference showing up during peak-demand hours and in voice AI. The second is cost: an average 33% reduction in token costs.

Accuracy also improved across all five classification features the model handles, according to the company's own testing. The gains range from 0.5 to 7.9 percentage points depending on the feature. Boost.ai did not break out which features sit at which end of that range, but the company presents the improvements as consistent rather than concentrated in one area.

These are internal benchmarks, not third-party evaluations, and the company's announcement carries them as its own claims. Even so, the direction of the numbers tells a clear story: boost.ai believes it can run classification cheaper, faster and more accurately with a purpose-built model than with general-purpose ones.

Why self-hosting matters here

CTO Rasmus Hauch framed the move around control. Controlling "every part of the stack it runs on," he said, allows the classification-specific model to keep improving. That's the argument for owning the infrastructure rather than renting intelligence from a large model provider.

The technical setup backs the claim. Boost.ai's Version 14 documentation places the fine-tuned Presis models on internally managed, dedicated server clusters, with low response times and high capacity as the stated design goals.

The data-processing picture stays simple, and for enterprise buyers that simplicity is a feature. Presis runs on AWS in the same region that already serves each customer instance, so the documented processing location does not change. AWS is already an approved sub-processor for boost.ai's customers. That means Presis introduces no new sub-processors, requires no data processing agreement change and demands no additional agreement.

The company also argues the architecture removes a category of risk. Routing, scope and safeguard decisions are no longer exposed to third-party outages, timeouts and capacity swings. For companies running customer-facing agents at scale, a classification model that depends on an external API is a dependency that can fail at the worst moment — say, during peak-demand hours, precisely when Presis's speed advantage is reportedly largest.

What changes in Version 14?

Version 14 of the boost.ai platform splits its language-model work into two distinct tracks: generation and classification. The split is strict, and it redraws the boundaries of customer choice.

On the generation side, GPT-5.1 is mandatory for built-in text-generation features and is the default for new agentic actions. On the classification side, Presis takes over the classifier component of five system areas: Agent orchestrator, Few shot, AI review, the out-of-scope detector and system guardrails.

The migration path has a hard requirement attached. Customers must configure GPT-5.1 before upgrading to Version 14. Without it, boost.ai says the system migration will fail. There's no soft landing here — the platform enforces the configuration up front.

Do customers lose model choice?

Yes, for built-in classification, and that's the trade-off at the heart of this release. Version 14 removes external Bring Your Own Model support and provider configuration options for classification tasks. Customers who previously pointed those classifiers at a third-party model can no longer do so.

Bring Your Own Model remains available for the generative parts of the platform: Agentic Actions, LLM Actions and Guardrails. So the restriction is targeted, not blanket. But within the five classification features Presis now owns, the platform's model is the only option.

For boost.ai, the bet is that tighter integration and measured efficiency gains outweigh the loss of provider flexibility. For customers, the immediate arithmetic is straightforward: tighter platform control and reported cost and speed improvements on one side, less freedom over which model runs built-in classification on the other.

How does the GPT-4o retirement fit in?

The transition has an external deadline driving it. Microsoft has scheduled the retirement of GPT-4o, GPT-4.1, GPT-4.1-mini and GPT-4.1-nano on Azure for April 2027. Boost.ai ties its broader model transition to that date, giving customers a fixed horizon rather than an open-ended preference.

The groundwork started earlier than the Presis announcement. Early migration and testing capabilities arrived with Version 12.8.3 in August 2026, and the mandatory Version 14 changes were scheduled to start reaching production in mid-October 2026 — weeks before Presis was publicly named on November 10.

Presis itself is not a new model under the hood. It's the renamed Boost Native Model. Boost.ai says the Norwegian name conveys accuracy and timeliness, and that the rebrand did not alter how the model works. The launch is as much about consolidating the platform's story around a named, self-hosted asset as it is about new technology.

What comes next?

Boost.ai is evaluating a similar dedicated-model approach for its generative features. The company frames this as an evaluation, not a committed launch — there's no date and no product attached yet.

If it happens, the move would extend the same logic that produced Presis: own the model, own the infrastructure, control the stack end to end. For now, the concrete changes are Version 14's split architecture, Presis's rollout across the platform over the coming weeks, and the April 2027 Azure retirement looming behind the whole migration. Customers weighing the upgrade face the sharpest version of the industry's current question — efficiency and integration from one vendor, or flexibility across many.