HistAI AI Autopilot

The models are the product. AI Autopilot exists to train and run custom pathology classifiers on whole slide images, built on HistAI's foundation models, and what the user buys is time to train and run models. The viewer and annotation tools on CellDX serve it, and do not stand as the offering. Without the models, nothing is left of this product.

Ask which foundation model sits under the custom models, and how its updates affect models already trained.

HistAI requires validation before clinical use, but publishes no threshold. Each custom model shows its performance on a dashboard as it trains. HistAI's terms allow a CLIA certified laboratory to use custom model outputs in a laboratory developed test only under the validation requirements they set out. Otherwise the platform is not cleared for clinical use.

What is missing is guidance on when a custom model is good enough: no minimum performance, no recommended test on held back data, and no statement of how a pathologist reviews outputs.

Ask what validation HistAI requires before a laboratory relies on a custom model, and what performance measures the dashboard reports.

HistAI describes its approach and its open models, without naming what runs under AI Autopilot. Custom classifiers are trained on features from HistAI's proprietary foundation models. Its open source Hibou models, released in 2024 under the Apache 2.0 license with a preprint describing their training on a large slide collection, show its approach.

AI Autopilot's page does not say whether Hibou or another model sits underneath, and no version or update practice is published.

Ask which foundation model and version AI Autopilot uses, and whether existing custom models are retrained when it changes.

HistAI builds its own models. Custom models are built on its foundation models and run on its platform, with no outside model provider named or implied.

The privacy policy names Stripe for payments and refers to cloud computing and analytics providers without naming them. No subprocessor list is published.

Ask for the cloud provider and the subprocessors that can access uploaded slides.

HistAI publishes research on the underlying models, and nothing on the product's own results. Its Hibou foundation models are described in a 2024 preprint on arXiv that reports results on public benchmarks. HistAI names academic institutions including Stanford, MIT, Harvard and Mayo Clinic among users of its platform.

AI Autopilot shows each user a performance dashboard for the model they train. No study of how custom models built this way perform against pathologists or established methods is published, and the preprint is not peer reviewed.

Ask for evaluations of custom models built with AI Autopilot on held back data, and for any results customers have reported.

HistAI's terms reserve broad rights, and do not say whether uploaded slides train its models. The terms give HistAI a perpetual, unrestricted license to content users contribute, and state that every custom model a user trains stays HistAI's exclusive property.

Neither the terms nor the privacy policy says whether uploaded slides or annotations train HistAI's own models or enter its Data Hub. Data is kept as long as necessary, with no period stated. No safety engineering for the models is published.

Ask in writing whether your slides, annotations or custom models can train HistAI's models or be licensed to others, and how long uploaded slides are kept.

HistAI asserts HIPAA compliance, without a status or agreement. Its privacy policy, effective 22 September 2025, covers the website, the CellDX app and related services, and names HistAI as controller of account data. It says HistAI processes personal and health related data in compliance with HIPAA and the GDPR, and that its slide datasets are de identified. Its terms make the user responsible for complying with HIPAA when uploading slides.

No business associate status is stated and no business associate agreement is published.

Ask whether HistAI signs a business associate agreement for laboratories uploading identifiable slides, and what de identification it requires before upload.

HistAI publishes no security audit or certification. Its privacy policy describes technical and organizational measures, including encryption of datasets. No SOC 2 report, ISO/IEC 27001 certificate, HITRUST certification, penetration testing statement or trust center is published.

Ask for a current security attestation with its scope and date before uploading patient slides.

HistAI makes no device claim, and scopes the product to research and laboratory developed tests. Its terms state that the platform and custom models are not FDA cleared or approved medical devices.

The terms allow CLIA certified laboratories to use custom model outputs in laboratory developed tests under validation requirements. That puts regulatory responsibility on the laboratory, under laboratory rules. The FDA's databases list no HistAI device.

Ask what validation documentation HistAI gives a laboratory building a test on a custom model.

Each model's builder sees its performance, but nothing is published on behavior across groups. AI Autopilot shows a model's performance on a dashboard during training, and the Hibou preprint reports benchmark results for the foundation models.

No results compared across scanners, stains, organs or patient groups are published for the foundation models under AI Autopilot. No governance framework or monitoring after deployment is described. Custom models trained on small or narrow slide sets are where performance is most likely to fail to carry over to new slides.

Ask how the dashboard tests a model against slides from other scanners and sites, and whether HistAI checks custom models before they deploy.

Users can check a model on its dashboard, and HistAI's liability is capped. Each custom model shows its performance as it trains, so its builder can test it before use.

HistAI's terms cap its liability at 1,000 dollars for any cause and keep custom models as HistAI's property. No warranty, error commitment or remedy attaches to model outputs.

Ask what HistAI does if a deployed custom model or its foundation model is found to give wrong results, and whether enterprise terms change the liability cap.

HistAI publishes no integration. Custom models deploy to HistAI's own CellDX platform, where slides are uploaded and viewed. No laboratory information system, image management system, scanner, standard or interface is named for bringing slides in or sending results out.

Ask how slides reach CellDX from your scanners or image management system, and how model outputs can be exported to your laboratory systems.

AI Autopilot runs only as a hosted service, and its location is implied, not stated. It runs with CellDX in HistAI's cloud, charged by GPU time, and the privacy policy refers to cloud computing services without naming the provider or region.

No on premises option, data location or separation between customers is described.

Ask which cloud provider and region host uploaded slides and models, and whether data can be kept in a chosen region.

Prices, units and a self service route are all published. HistAI lists subscription plans from 15 to 39 dollars a month, and model training from 15 to 50 dollars a model, charged for the GPU time actually used. A library of more than 66,000 slides is free for training, and sign up and payment happen online. Data Hub slides are priced separately, per slide.

Ask what a typical model costs to train and run at your slide volumes, and whether enterprise terms change who owns custom models.

HistAI sets no limits on where AI Autopilot applies, and nothing validates its use. It can be trained on any whole slide images a user supplies or selects, across the 20 organ systems and the H&E and immunohistochemistry stains in HistAI's library. It is offered to pathologists, researchers and laboratories.

HistAI does not publish where custom models built this way perform well or poorly, and its terms leave validation for any clinical use to the laboratory.

Ask which tissue types, stains and scanners the foundation models were evaluated on, and what minimum training set HistAI recommends for a reliable classifier.