Microsoft enters decision models, and the category becomes a platform feature

Decision models as a category are only about four weeks old. Jev launched in mid-September 2026 and Clef on October 1. I found no product that tracks calibration specifically for decision-API outputs against labelled outcomes, or that flags drift when a customer switches between Jev, Clef and Perplexity's decider. I also found no vendor selling insurance-, lending- or trust-and-safety-tuned decision models with model-risk documentation. So both ideas are open today. They won't stay open for long, though. General ML monitoring vendors already serve insurers' fraud and claims models and could add Brier score, expected calibration error and reliability-curve tracking per vendor in a sprint. Those are Fiddler, Arize and Monitaur. ValidMind already packages SR 11-7 documentation for third-party models. Cloudflare's fine-tuning service lets buyers make their own vertical models. The base-model price war also squeezes margins for anyone selling generic decision models. The best place for a newcomer to win is a neutral calibration and swap-certification layer. It would treat the decision API as an untrusted vendor, sit in the request path or read logs, and produce evidence that a model risk examiner would accept. That evidence would include holdout calibration before and after any model or version change, a recalibration step such as isotonic or Platt scaling per vendor, and threshold re-tuning. Model-risk teams at mid-size banks and insurers can't get that from the decision-model vendors themselves, who can't credibly grade their own calibration. Public doubts about Jev's calibration claims help this pitch. The vertical-model idea is weaker on its own, because Cloudflare's fine-tuning and the open Clef weights make the models easy to reproduce. It holds up better if the defensible part is the validation packs plus proprietary labelled data partnerships, sold as documentation and monitoring wrapped around whichever base model is cheapest, not as the model itself. Speed matters, because the incumbent monitoring and model-risk vendors will likely announce decision-model support within a quarter or two.

• TypeSafe AI (Jev)

  The closed decision-model API that started this category. It launched mid-September 2026 and raised $870M at a $7.5B valuation. It's the main vendor both ideas would sit on top of or compete with. One independent blog test reported calibration failures on 7 of 8 datasets, which is a public data point that supports a monitoring product.

• Cloudflare Clef / Clef-flash and RL fine-tuning platform

  Open-weight (Apache 2.0) decision models that accept Jev's API format, so customers can swap one for the other. They're trained with a Brier-loss calibration term, and Cloudflare offers RL fine-tuning as a service. That service is the closest threat to the vertical-model idea, because it lets any insurer or lender fine-tune on their own labels. Cloudflare itself tells customers to calibrate thresholds on their own data.

• Perplexity pplx-decider and Amazon Strands Decider

  More decision models from the same launch wave. Together with OpenAI's Decisions API they add to the price war. Each new interchangeable vendor gives a swap-and-drift monitor more reason to exist, but also makes the base models cheaper and more commoditized.

• Fiddler AI

  An established ML monitoring platform that already covers fraud-detection models, including drift, performance on imbalanced classes and false-positive tracking. It could add calibration tracking for decision-API outputs without much effort.

• Arize AI

  An ML observability platform used by insurers, for example to monitor Clearcover's claims-fraud models. It already checks drift and runs threshold monitors, so it's a likely fast follower on calibration drift for decision APIs.

• Monitaur

  An AI governance and monitoring platform built for insurance. It monitors bias, drift and anomalies and offers independent validation. It overlaps with both the monitoring idea and the model-risk documentation angle for insurers.

• ValidMind

  A model risk management platform for banks and insurers. Its Advantage program helps third-party model vendors produce SR 11-7 and EU AI Act validation documentation, and its Experian partnership ships templates for credit and fraud models. It's the main incumbent for the validation-pack part of the vertical-model idea.

• jev-claim-processor (open source)

  A hobby GitHub project that uses Jev to triage insurance claims, with confidence thresholds and a rule-based fallback. It shows developers are already applying decision models to claims, but it isn't a commercial product.

Checked 2026-10-10