BPFlow AI / News

Healthcare AI agents, child safety and sites built from sources

Issue date:

This issue covers tools for healthcare and content safety teams, plus a new way to present information from user-provided sources.

This issue may include important events published earlier.

News archive →
01

OpenAI adds agents and plugins to ChatGPT for Healthcare

OpenAI introduced agents that operate healthcare software through its interface, improved the search for medical evidence and added more than 20 plugins to ChatGPT for Healthcare.

Key facts

  • Computer-use agents in ChatGPT Work and Codex and improved search are available to ChatGPT for Healthcare customers, subject to workspace settings; administrators enable the new plugins. Individual clinicians have access to agents, search and select plugins. Free ChatGPT for Clinicians accounts are for eligible, verified U.S. clinicians.
  • OpenAI says physicians rated the updated experience more highly on all eight criteria, including factual correctness and safety, in a blind evaluation involving more than 1,300 reviews of responses.
  • OpenAI reports that a CVS Health agent cleared a months-long backlog of documents awaiting dispatch over a weekend in a production workflow. CVS Health uses validation checks and human oversight.

An objective look at what matters

Agents could connect steps across systems where staff currently move information by hand.

Business processes

  • Referrals and prior authorization — Prepare documents and gather information from different systems for specialist review.

Automation opportunities

  • Preparing an MRI request: An agent could review a chart and draft an MRI order and patient reply.
    Conditions: Access to approved applications and clinician review are needed.; HIPAA-compliant work with protected health information requires eligible services covered by a BAA with OpenAI, the required security and retention settings, and protections for connected systems.

Impact on manual work

OpenAI describes a production workflow at CVS Health but does not provide a general measure of labor savings across customers.

Limitations and risks

  • Clinical conclusions and agent actions need review; errors and data exposure can harm patients.

Sources

  1. OpenAI News: More ways to move care forward with ChatGPT for Healthcare · Source date:
  2. OpenAI News: Advancing computer-use agents for healthcare teams · Source date:
02

OpenAI releases classifier to detect child sexual abuse material

OpenAI released an image classifier to detect potential child sexual abuse material, including AI-generated images. Its scores help reviewers prioritize images for inspection.

Key facts

  • Vetted organizations can obtain access at no cost through the API, an interface for connecting to the service.
  • In OpenAI's evaluation of user-uploaded images, the company estimated that the classifier would flag 48% fewer images while finding roughly the same number of abusive images as its existing tool.
  • OpenAI reports that, in partner fal's test, reviewers confirmed 72% of high-confidence flags as previously unidentified child sexual abuse material.

An objective look at what matters

The tool could help safety teams prioritize cases and reduce unnecessary exposure to material.

Business processes

  • Reviewing user-uploaded images — Sort the review queue using classifier scores.

Automation opportunities

  • Initial image screening: Send classifier scores to a queue for specialist review.
    Conditions: Access is subject to organizational vetting; decisions require specialist involvement.

Impact on manual work

Tests suggest review work could fall, but the effect depends on an organization's data and process.

Limitations and risks

  • False positives and missed material are possible. A model score does not replace human review or established reporting procedures.

Sources

  1. OpenAI News: Strengthening the fight against online child sexual abuse with new tools and support · Source date:
03

Gemini Notebook creates interactive sites from sources

Google began rolling out a Gemini Notebook feature that creates interactive sites from a user's sources on request. Finished sites can be opened in a new tab or shared.

Key facts

  • Google gives sales dashboards, learning aids and task trackers with deadlines as examples.
  • The feature is rolling out globally on the web for users aged 18 and over, including Google Workspace subscribers. It is coming to the mobile app later.

An objective look at what matters

Teams could turn collected material into interactive views more quickly without building a site themselves.

Business processes

  • Data analysis and project management — Create an interactive dashboard or tracker from available sources.

Automation opportunities

  • Creating a sales dashboard: Generate a site from a description and source material, then share it with colleagues.
    Conditions: The web feature is still rolling out; users must be at least 18.

Impact on manual work

Generation could reduce manual site-building work, but Google provides no measured time savings.

Limitations and risks

  • Check the data and how the generated site works before using or sharing it.

Sources

  1. Google Blog: Transform notes, research, and sources into interactive sites with Gemini Notebook. · Source date: