boost.ai Presis: self-hosted AI model classifies agent conversations 2.5x faster
boost.ai announced Presis about an hour ago — a fine-tuned, self-hosted language model built for one job: the classification decisions inside every AI agent conversation. Which part of the agent should answer, whether the question is in scope, whether the answer is safe to show, and how the finished conversation gets reviewed — Presis now makes those calls across boost.ai’s platform, and the company says it does so at least 2.5 times faster than the general-purpose models it replaces.
What we checked: boost.ai press release “boost.ai Launches Presis, A Self-Hosted AI Model Built for Speed & Accuracy in Regulated Industries,” published November 10, 2026 at 11:05 UTC via PR Newswire, cross-checked against Unite.AI’s November 10 coverage of the same launch.
- What: Presis, a fine-tuned, self-hosted language model from boost.ai that makes the classification decisions in and around every AI agent conversation.
- When: announced Tuesday, November 10, 2026, at 11:05 UTC (06:05 EST) via PR Newswire.
- Speed: at least 2.5 times faster than the models it replaces, per internal testing — most noticeable during peak-demand hours and in voice AI.
- Cost and accuracy: token costs down 33% on average; more accurate than the models it replaces on all five classification features, by 0.5 to 7.9 percentage points.
- Rollout: already powers those decisions across the boost.ai platform; the broader rollout continues over the coming weeks.
- Scale: boost.ai reports 650+ deployments, 600 live AI agents and more than 200 million automated conversations; customers include Nordea, DNB, Sage and Trading 212.
Presis replaces the general-purpose models boost.ai used to rent for the same job. Here is the head-to-head, on boost.ai’s own numbers:
| Presis | The general-purpose models it replaces | |
|---|---|---|
| Speed under load | At least 2.5× faster, per internal testing — steadiest at peak demand and in voice AI | Slows when third-party capacity tightens, worst at peak hours |
| Token cost | 33% lower on average | Baseline |
| Classification accuracy | 0.5 to 7.9 percentage points better on all five classification features | Baseline |
| Infrastructure | Runs on boost.ai-controlled infrastructure, on AWS in the same region as each customer instance | Exposed to third-party outages, timeouts and capacity swings |
Presis: a specialist, not a generalist#
The job is narrow, and it sits on the critical path. Every automated conversation passes through five classification decisions: which part of the agent should respond, whether the question is in scope, whether the answer is safe to show to an end user, and how the finished conversation gets reviewed — plus the system guardrails around them. In voice deployments those checks have to finish before a response can even begin, so a slow classifier is a slow conversation.
Presis is the renamed Boost Native Model — boost.ai says the Norwegian name conveys accuracy and timeliness, and the rename changed nothing about how the model works. “A model tuned specifically for classification can do things that a general-purpose model can’t, including staying fast under peak load and improving steadily because we control every part of the stack it runs on,” CTO Rasmus Hauch said in the announcement.
It already powers those decisions across the platform, with the broader rollout continuing over the coming weeks. And the scale behind it is not small: boost.ai reports more than 650 deployments, 600 live AI agents and over 200 million automated conversations, with customers including Nordea, DNB, Sage and Trading 212. The company was named a Leader in Gartner’s 2025 Magic Quadrant for Conversational AI Platforms.
The models it replaces: general-purpose, rented by the call#
The incumbents are general-purpose, third-party models — built to do everything for everyone, and rented by the call. That breadth is the weakness boost.ai is attacking: when a third party’s capacity tightens, routing and safeguard decisions slow down with it, exactly when call volumes peak. Because Presis runs on infrastructure boost.ai controls end-to-end, the company says, routing, scope and safeguard decisions are no longer exposed to the outages, timeouts and capacity swings that can affect third-party providers.
The timing is not accidental. boost.ai’s Version 14 splits the platform’s language-model work in two: GPT-5.1 becomes mandatory for built-in text generation, while Presis takes the classifier side. The company ties the shift to Microsoft’s scheduled April 2027 retirement of GPT-4o, GPT-4.1, GPT-4.1-mini and GPT-4.1-nano on Azure. The same pattern is spreading elsewhere: Meta’s Muse Ember is a small model built specifically for the agent grunt work of tool calls and routing, and Anthropic’s Claude Fable 5.2 arrived 30% faster and cheaper than its predecessor.
The trade-off: control in, provider freedom out#
The control cuts both ways. Under Version 14, external bring-your-own-model support is being removed for built-in classification — customers lose the freedom to plug a different provider’s model into those decisions, though bring-your-own-model remains for generative Agentic Actions, LLM Actions and Guardrails. Customers must also configure GPT-5.1 before upgrading; without it, the system migration will fail.
On the compliance side, the news is cleaner. Presis runs on AWS in the same region that already serves each customer instance, so the documented processing location does not change, and AWS is already an approved sub-processor — no new sub-processors, no data-processing-agreement change.
“For regulated organizations, every automated conversation matters – speed and reliability are core to how they build trust with their customers,” CEO Jerry Haywood said. “Presis reflects what we’ve learned from a decade of working inside those industries.”
The catch#
Every performance number in this story comes from boost.ai’s own internal testing. No independent benchmark has been published, and the 0.5-to-7.9-point accuracy gains vary by feature — the low end is a rounding error, not a revolution. The rollout is also still in progress, so most customers have not felt it yet. And the bigger prize — a dedicated model for generative features, the answers customers actually read — is only being evaluated, not launched.
The verdict#
Presis is a sensible, unglamorous bet: take the decisions a general-purpose model does adequately, give them to a specialist you control, and bank the speed and the cost savings. For banks and insurers, where a delayed or wrong answer is a compliance problem rather than an inconvenience, that trade is easy to justify — provided the internal numbers hold up outside boost.ai’s own tests. It lands in a week when agent infrastructure is moving fast everywhere: AWS just gave Bedrock agents web search. Watch whether the generative version of Presis follows; that is where the real test of the strategy lies.