Production AI Institute · Public recordRecord
Deployment GuidesPSF-aligned · May 2026

So you want to deploy
AI agents in production.

Microsoft, Anthropic, Google, and OpenAI are rolling out managed agents, plugins, and connectors that can autonomously interact with your business systems. Most organisations have no idea where to start — or what can go wrong. These guides exist to change that.

Start with the explainer →Jump to the M365 playbook →Business workflow library →
Open source · MIT

WorkflowOS — fork it, self-host it, ship it for clients

PSF Workflow Studio is our free reference implementation: canvas, PSF scoring, AI heal, and exportable deployment packages. MSPs and integrators can run it on their own infrastructure — no licence fees, full source on GitHub.

Canvas + simulatePSF 8-domain analysisPortfolio exportsBYOK AI
Star on GitHubRead why we open-sourced it →Try hosted Studio →
Why this section exists: Managed agents, plugins, and connectors are moving from product announcements into real deployment work. Practical, safety-aligned guidance for the teams implementing them is still scarce. This section exists to close that gap.
Deploying is not the win

Choose the evidence path before the agent goes live.

The deployment guides are free to read. The next step is deciding whether this is a readiness check, client service, organisation rollout, or external assessment.

I deploy for clients

Map every deployment to the controls it needs.

Use the vendor control mapper and control templates to structure readiness checks, rollouts, and governance around the PSF.

Map a vendor →

I have a live agent

Check the deployment before it becomes a dependency.

Run the readiness check when an agent can touch customers, systems, private data, or operational decisions.

Run readiness check →

I am building now

Benchmark readiness while the design is still movable.

Run the AI agent readiness check before the agent becomes a production dependency that is harder to change.

Run readiness check →

I manage rollout

Give every team the same open standard.

Adopt the PSF as the shared vocabulary for roles, training, policy, and deployment review across teams.

Read the PSF →
The answer

What “production-ready” actually means for an AI agent

A demo is production-ready when it works on the happy path in front of you. An agent is production-ready when it keeps behaving correctly when you are not watching: under adversarial input, during a vendor outage, six months after the person who built it has moved teams. Most “AI agent deployment” guides skip straight to SDK code. The harder and more consequential part is the five things below — skip any one of them and the agent will eventually fail in a way that reaches a customer, a regulator, or a courtroom.

This is the same framework PAI uses to assess and certify deployments — the Production Safety Framework (PSF). Each step below links to the full control guide.

1PSF D1Govern what reaches the modelClassify and filter inputs before they hit the agent — adversarial prompts, prompt-injection attempts, and out-of-scope requests need to be caught here, not discovered in production.2PSF D5Red-team before launch, not afterTest the agent against jailbreak patterns, edge-case policy questions, and failure modes deliberately, before real users find them for you.3PSF D6Make human oversight active, not nominalA human “in the loop” who isn't actually reviewing anything is not a control. Define what triggers escalation, and who is accountable for the final call on high-stakes actions.4PSF D4Monitor predicted vs. actual outcomesUptime is not enough. Track whether the agent's decisions are still correct in the real world, and alert before drift becomes a financial or safety problem.5PSF D8Plan for the vendor going downA single model provider is a single point of failure. Know your degradation path — cached responses, a secondary route, or a clean human handoff — before you need it at 2am.

What skipping this actually costs

These aren't hypothetical risks. Each is a documented, source-linked incident in PAI's public incident registry.

PSF D5 / D6Uber disabled emergency braking and treated a safety driver's presence as oversight instead of active monitoring — a pedestrian died.PSF D4 / D8Zillow's pricing model kept buying homes at above-market prices because nothing monitored predicted-vs-actual value in real time. $528M write-down.PSF D1 / D5A customer jailbroke DPD's support chatbot with a one-line prompt-injection attack the industry had already documented for years. No pre-launch red-team caught it.

Deployment Guides

Start Here🧩

AI Agents Decoded

What managed agents, plugins, and connectors actually mean for your business — without assuming prior AI engineering depth.

FoundationsNon-technicalBusiness owners
12 min read
Deployment Guide🔮

Deploying Claude Agents Safely

From API setup to production — tool use, guardrails, human oversight, monitoring, and PSF compliance. A practitioner can follow this end-to-end.

AnthropicClaudePythonTool UsePSF
40 min read
Deployment Guide⚡

Deploying OpenAI Agents Safely

Agents SDK with guardrails, interrupts, tracing, and PSF alignment — from zero to production in one walkthrough.

OpenAIAgents SDKPythonGuardrailsPSF
40 min read
M365 Playbook🏗️

Microsoft 365 Agent Deployment

The complete step-by-step playbook for deploying AI agents in Microsoft 365 environments, from tenant readiness to live monitoring.

Microsoft 365CopilotCopilot StudioPSF
45 min read

Every guide is PSF-aligned

The Production Safety Framework (PSF) is PAI's open standard for production AI. All deployment guidance in this section maps directly to PSF domains — so you're not just deploying fast, you're deploying safely.

D1 · Input GovernanceD4 · ObservabilityD5 · Deployment SafetyD6 · Human OversightD8 · Vendor Resilience
Read the full PSF →

Evidence tools for deployers

AI Agent Readiness Check
Score a live agent deployment against the PSF controls and publish a citable readiness report
Evidence pack builder
Turn readiness results into artifacts that survive procurement, audit, and internal review
Vendor control mapper
See which PSF domains your stack covers and which remain your responsibility

More guides coming

We're building out the full library of production deployment guides. On the roadmap:

  • Google Workspace agent deployment playbook
  • Salesforce Einstein / CRM agent deployment
  • Open-source agent stacks (LangGraph, CrewAI) in enterprise environments
  • AI agent deployment for regulated industries (financial services, healthcare, legal)
  • Multi-agent orchestration safety patterns

Have a use case you need covered? Tell us what you need.

The Production AI Brief

Keep your AI map current

Tool policy changes, new incident records, disclosure signals, and practical next steps. Public evidence, plain English, no hype.