AI Governance in Posit Team
TL;DR
- Posit's Prime Directive holds that software for data analysis must be trustworthy and shown to be trustworthy. Code is the only way to satisfy that standard, so Posit's AI governance rests on four principles: code first, human in the loop, session aware yet data private, and governed environments that enforce compliance automatically.
- Workbench controls who can use AI and which models they can use. A single configuration file turns Posit Assistant off by default, grants access only to approved groups, and enforces which model providers and models are permitted. All traffic can route through an AI gateway, and no individual ever holds a provider API key.
- Posit's tools already report on AI usage through Workbench auditing, Posit Chronicle, console auditing, and token level metrics in the Assistant, all logged on premise. Centralized utilization metrics, cost controls, and agent versus human attribution are in development.
- Package Manager extends its existing repository controls, curated lists, blocklists, and cool down rules to agent installs without extra setup, and now supports agent plugins that admins can pre-approve and lock. Posit Connect gives every piece of AI content a named owner, an access control list, and a full audit trail of deployments and changes.
Why govern AI with Posit?
Trustworthy AI requires trustworthy code, enforced across your entire toolchain.
Posit’s responsible AI principles
Four principles govern how Posit AI works: code-first, human-in-the-loop, session-aware yet data-private, and enforced by the system, not the individual.
Restricting AI access
One configuration file, driven by your identity provider's existing groups, controls exactly who can use AI. It’s off by default. You turn it on only for approved users.
Control model access
Providers.json enforces which model vendors, which models, and which network paths your team can use: for example, only Bedrock, only Sonnet 4.5, and all traffic routed through your AI gateway. No individual ever holds a provider API key. Workbench issues short-lived credentials automatically when a user logs in.
Credential Security and Usage Reporting
Workbench manages provider credentials for you, so no one copies, pastes, or stores an API key. Existing tools already report on session activity, console commands, and token usage, all logged on premise rather than sent to Posit. New features in development will centralize these metrics further and add cost controls.
AI plugin governance
Package Manager's existing repository controls, curated lists, blocklists, and cool-down rules apply automatically to agent installs, no extra configuration needed. Package Manager now also supports agent plugins, letting admins pre-approve and lock a set of skills, commands, and MCP server definitions that appear automatically when a user starts Positron.
Governing content
Every piece of AI content deployed to Connect, chatbots, Quarto documents, MCP servers, gets a named owner, an access control list, and a full audit trail of deployments and changes. Admins can see exactly who has access and review a detailed log of every action taken on the platform.
Live Q&A: Meet the AI Governance team!
The Posit team answers your questions on governing AI in enterprise environments: model access, token limits, package rules, logging, and deployment credentials.
- 00:00 - Meet the team! Introducing Sam, Jeremy, and Tom.
- 01:06 - Can OpenAI Compatible Endpoints be used?
- 01:27 - Does everything presented here also apply to Posit Team Workbench in Snowflake?
- 02:12 - Is there a way to configure a token cap? Is there a way to dynamically select which model is used based on task difficulty?
- 04:31 - Can you turn off console access for Posit Assistant?
- 06:53 - Is it possible to have different rules in Package Manager for agents versus humans?
- 10:39 - Allowed packages for AI use?
- 12:04 - When deploying AI chatbots on POSIT Connect, which user's credentials or API tokens are used, and who's billed for those requests?
- 13:03 - Can the console access relays be restricted at the user group level?
- 13:21 - Can you say more about what any logging or tracking model usage would look like once they are deployed?
- 16:27 - Are there endpoints for internally hosted LLMs?
- 17:58 - Are you able to enforce available LLM models per individual or group?
- 19:18 - In Posit Assistant, can you configure multiple models simultaneously for GAN (generative adversarial network) models and setup?
- 21:15 - Do we have a solution to control token allocation per group on each AI provider? Is there a way for each customer to perform audit AI usage?
- 23:13 - Tell me something that’s really exciting about AI for you right now. Personal or data science related.