Not tested by us yet. Presis is not sold as a model you can call or download. It runs inside boost.ai's own platform, so there is no endpoint for us to benchmark. Everything below comes from boost.ai's launch announcement. We'll update this post if boost.ai opens access or publishes a model card.
Presis is a fine-tuned language model from boost.ai, a Norwegian conversational AI vendor, that makes the small classification decisions around each AI agent conversation: which part of the agent should answer, whether a question is in scope, and whether an answer is safe to show to a customer. boost.ai says it is at least 2.5 times faster than the models it replaces and cuts token costs by about a third. It is aimed at regulated industries, and it is not a general-purpose chat model.
| Presis | |
|---|---|
| Maker | boost.ai |
| Announced | 2026-11-10 (Sandnes, Norway) |
| Type | Fine-tuned language model for classification |
| Parameters | Not disclosed |
| Hosting | Self-hosted on boost.ai's own infrastructure |
| Availability | Rolling out across the boost.ai platform "in the coming weeks" |
| Price | Not published; part of the platform |
| Open weights | No |
What it does
boost.ai describes Presis as covering five classification features inside its agent platform. Three are named in the announcement: choosing which part of the agent responds, checking whether a question is within the agent's scope, and checking whether a candidate answer is safe to show an end user. A fourth is the review of finished conversations. These are the "decision" steps that sit around a generative answer, and until now they were handled by general-purpose models.
That is the same job that dedicated decision models target. If you have not met the category, our write-ups of Kakao's Kanana-2-v-decide and Ingenix's Thalamus Decision show what a model looks like when it returns a choice instead of prose. The difference with Presis is that it is a private component, not a product you can buy separately.
What is new
The news is the move from third-party models to an in-house one for these steps. boost.ai's reasoning is about control. Because Presis runs on infrastructure the company operates end to end, routing, scope and safeguard decisions are no longer exposed to a third party's outages, timeouts or capacity swings. For any team running a customer-facing agent, a classifier that stalls takes the whole conversation down with it, so that argument is a practical one.
boost.ai also says it is evaluating a similar approach for generative features later. Nothing is announced there.
The claims, and who made them
Every figure below is boost.ai's, from internal testing. None is ours, and none comes with a public benchmark, dataset or model card.
- Speed: at least 2.5 times faster than the models it replaces, with the largest gains at peak demand and in voice AI deployments.
- Cost: an average 33 percent reduction in token costs.
- Accuracy: better than the replaced models on all five classification features, by margins of 0.5 to 7.9 percentage points depending on the feature.
Read those with the usual care. The comparison models are not named. The accuracy gains are small on some features (half a point is within the noise of most evaluation sets), and the largest is under eight points. The speed and cost numbers are averages of unspecified workloads. The claim that a specialised small model beats a general one at a narrow job is plausible and common, but this announcement gives you no way to check it.
Who should care
- Existing boost.ai customers. The rollout reaches you automatically. Ask what changes in your scope and safety rules, and whether you can compare before and after on your own conversations.
- Anyone buying a conversational AI platform in a regulated sector. The pitch is predictable latency and no third-party dependency in the decision path. That is a real procurement question, and worth asking every vendor.
- Teams building their own agents. You cannot use Presis, but the pattern transfers. Routing and safety checks are the cheapest place to stop paying for a frontier model. Our post on cheap-first routing with escalation shows the generative version of the same idea, and Salesforce's Koa Decisions is another vendor-side decision model still waiting for public access.
Skip it if you want to evaluate it independently. With no weights, no API and an undisclosed size, there is nothing to measure yet.
What we will test if access opens
If boost.ai exposes Presis, or a customer shares a way to call it, we would run it next to a cheap and a mid-priced chat model acting as classifiers, on one shared set of 30 to 60 routing and scope items. We would include control items with obvious answers and ask yes/no questions in both directions, since a classifier that flips when the wording is mirrored is not usable for safety checks. Median latency and cost per 1,000 decisions would go in the same table. Until then, treat the 2.5x, 33 percent and 7.9 point figures as vendor claims.