Technical decision guide · 8 minute read

Choose the operating model before the controls.

For health-tech and healthcare teams, the first decision is not a checklist. It is who uses the system, where the workflow runs, what data crosses each boundary, and which OpenAI service your agreement actually covers.

Source review: Independent Ginylil guidance

Decision map

Start with the user and system boundary.

Managed healthcare workspace

ChatGPT for Healthcare

Choose this path when
Clinicians, administrators, or researchers need a governed ChatGPT workspace built for a healthcare organization.
Investigate before choosing
Which users, features, connectors, retention settings, and administrative controls your BAA and rollout cover.
Not a fit when
The AI capability must live inside your own application or product workflow.

Managed enterprise workspace

ChatGPT Enterprise with Regulated Workspace

Choose this path when
Employees need a governed general-purpose workspace with enterprise identity, roles, and administrative controls.
Investigate before choosing
Whether the required functions are enabled and covered, including any connector, search, sharing, or local-client use.
Not a fit when
The end user is a patient or customer interacting through your own product.

Custom application

API with Modified Retention

Choose this path when
Your organization must own the application experience, authorization, data flow, logging, human review, and operational controls.
Investigate before choosing
BAA execution, organization provisioning, endpoint eligibility, storage behavior, third-party tools, and the exact account and project configuration.
Not a fit when
You need a ready-made employee workspace rather than a product engineering program.

Local engineering client

Codex Local

Choose this path when
Authorized engineers need CLI, IDE, or desktop assistance and your organization can manage the workstation and repository environment.
Investigate before choosing
Authentication path, workspace eligibility, local retention, repositories, MCP servers, browser or computer use, secrets, and third-party services.
Not a fit when
You cannot enforce managed endpoint, identity, repository, and tool-use policy.

Five questions

The minimum useful discovery.

  1. Who is the user?Employee, clinician, researcher, engineer, patient, or customer.
  2. Where does the workflow live?OpenAI workspace, local developer tool, or your application.
  3. What crosses the boundary?Data classes, files, prompts, outputs, metadata, and connected systems.
  4. Which agreement applies?Product, BAA, eligible functions, retention controls, and account provisioning.
  5. Who operates the controls?Identity, endpoint, application, logging, review, incident, and vendor owners.

Official sources

Validate before implementation.

Apply the map

Turn one use case into an architecture decision.

Use sanitized facts. Do not send sensitive information.

Review the Architecture Sprint