Presis handles several classification functions around agent conversations, including routing, scope detection, guardrails and conversation review. Boost.ai, which sells conversational AI to banks, insurers and other regulated companies, announced on November 10 that those functions across its platform will now be handled by a model of its own, called Presis.

The move also removes a dependency with an expiry date. Boost.ai's documentation notes that the features Presis takes over previously ran on a mix of GPT-4o and GPT-4.1 models, and boost.ai's documentation says Microsoft will retire GPT-4o, GPT-4.1, GPT-4.1-mini and GPT-4.1-nano on Azure in April 2027. Moving the classification layer in-house before that schedule is maintenance with a launch attached. The architecture explains the intended approach; the unpublished evaluation details prevent an independent assessment of the claimed gains.

Three numbers, one source

Boost.ai reports that its internal testing found Presis at least 2.5 times faster than the models it replaces, with the difference most noticeable at peak hours and in voice deployments. It reports token costs 33 percent lower on average. And it reports accuracy improvements on all five features it powers, by between 0.5 and 7.9 percentage points depending on the feature.

Every one of those numbers is labelled, in the release itself, as internal testing. What internal testing does not come with is a way to check it. No baseline is given: 2.5 times faster than what latency, measured over how many conversations, is not stated. The cost figure is a token-cost average with no workload described; it is not a total operating cost, a customer price or a guaranteed bill saving. The results are reported against the GPT-4o / GPT-4.1 replacement mix as a group, not demonstrated separately against each model. Without sample sizes, uncertainty estimates and per-feature results, readers cannot assess the reliability or practical importance of either end of the reported range. Publishing the range without the per-feature table lets the headline carry the 7.9 while the 0.5 stays technically disclosed.

None of this makes the claims false. It makes them unverifiable from the published material: so far only the company has done the measuring, in an internal evaluation whose methodology and detailed results are not published.

A new name within a mandatory Version 14 migration

Boost.ai's Version 14 documentation is more candid than its press release. A note in the model-update article states that the model previously called the Boost Native Model is now called Presis, a Norwegian word the company glosses as accurate and on time, and that nothing about how it works has changed with the name. The rename note concerns continuity between Boost Native Model and Presis. The performance claims compare Presis with the GPT-4o / GPT-4.1 models it replaces; they do not describe gains caused by renaming it.

The Version 14 changes around the rename are concrete, and they are required when upgrading to Version 14. Presis is required for the five listed built-in classification features; the documentation states it is not possible to configure a different model for those features, and that upgrading transfers them automatically. GPT-5.1 must be configured for built-in text generation, or, in the documentation's words, the system migration will fail. Provider configuration remains available for the specified generative features, including agentic actions, LLM actions and guardrails that customers configure themselves; Bring Your Own Model support and provider configuration are removed for built-in classification only. The mandatory updates began reaching production in mid-October 2026, before the November 10 announcement.

Read together, the two halves are specific rather than directional. For built-in classification, provider choice is removed. For generation, configuring one specified model becomes a condition of upgrading. A bank that chose this platform partly for model flexibility keeps that flexibility where it writes answers, and loses it for the built-in classification features.

Self-hosted, on someone else's cloud

The release headline calls Presis self-hosted, and its text says the model runs on infrastructure boost.ai controls end to end. The documentation adds the address: Presis runs on AWS, in the same region that already serves each customer instance, on dedicated server clusters hosted and managed by boost.ai. Per the documentation, AWS is already an approved sub-processor, so no new sub-processors are introduced and no data processing agreement change is required.

Self-hosted here means hosted by the vendor, not by the customer. Vendor-controlled model serving does not eliminate AWS dependency; the end-to-end control in the release means control of the model and its clusters, not of the cloud underneath them. The dependency has moved down a layer and narrowed. It has not disappeared.

What a customer can actually do with this

The sources reviewed describe Presis as a built-in platform component and provide no standalone download or public evaluation package. It arrives as part of the Version 14 upgrade, and the release says the wider rollout continues over the coming weeks. The company is also evaluating a similar dedicated-model approach for generative features, which is an evaluation, not a commitment.

So the practical question for a boost.ai customer is not whether Presis beats the old mix by 2.5 times on boost.ai's test. It is how the new classifier behaves on their own conversations. Per the documentation, migration and test tooling, including test-model features in its test studio since version 12.8.3 in August 2026, is available for comparing behavior before the transfer. Running already-reviewed conversations through the new classifier and comparing its scope and guardrail calls against the old ones would show a compliance team the classification behavior on their own material. The launch figures rest on an internal evaluation whose methodology and detailed results are not published. The upgrade requirements are concrete either way, which is precisely why the checking is worth doing.

Sources

  1. [1] boost.ai, "boost.ai Launches Presis, A Self-Hosted AI Model Built for Speed & Accuracy in Regulated Industries", PR Newswire, November 10, 2026, 11:05 UTC (read in full first-hand; all performance, cost and accuracy figures, rollout wording and company scale figures)Read source
  2. [2] boost.ai documentation, "Presis (boost native model)" (read in full first-hand; feature list, no alternative model configurable, automatic transfer on Version 14 update, the GPT-4o and GPT-4.1 mix it replaces)Read source
  3. [3] boost.ai documentation, "Model Updates coming in Version 14" (read in full first-hand; rename note, April 2027 Azure retirements as stated by boost.ai, mandatory GPT-5.1 configuration and migration failure, Bring Your Own Model removal for classification, feature list, Presis on AWS in the same region, no new sub-processors, mid-October 2026 production rollout)Read source
  4. [4] Unite.AI, "Presis Brings Self-Hosted Classification to Boost.ai's Agent Platform", November 10, 2026 (secondary summary used only to cross-check the release and documentation; no load-bearing fact rests on it alone)Read source