AI Project Policy

Appearance

Production Handoff Triggers ​

Purpose ​

Handoff is the process for moving an AI idea from informal experimentation to accountable company ownership.

The handoff brings in the domain specialists responsible for the affected systems, data, risks, and operations. The original team should continue to provide product context and business goals.

Basic rule ​

If an AI use case would create real risk if it failed, leaked data, acted incorrectly, became unavailable, or produced biased or misleading output, it is no longer only an experiment.

Start handoff before the work crosses that line.

Mandatory handoff triggers ​

An AI experiment must be handed off to the appropriate accountable teams before any of the following occur.

Production system access ​

The AI workflow will read from, write to, call, monitor, configure, deploy to, or otherwise interact with a production system.

Examples:

  • querying production databases or logs;
  • calling production APIs;
  • opening, closing, or modifying production tickets based on AI output;
  • deploying generated code or configuration;
  • running an agent with access to infrastructure tooling;
  • integrating with production CI/CD, observability, IAM, or incident tooling.

Sensitive data use ​

The AI workflow will process sensitive data, including customer data, employee data, secrets, financial data, security findings, private source code, production logs, or confidential company material.

Real-world impact ​

The AI workflow will influence a customer-facing experience, employee-impacting workflow, operational decision, financial decision, security action, or business-critical process.

Durable or operational use ​

The AI workflow is no longer a disposable prototype and is becoming part of a process that people or systems rely on.

Use by multiple teams or recurring automation does not require handoff by itself. Handoff is required when the workflow also creates production, sensitive-data, operational, security, capacity, or real-world-impact concerns.

Autonomous or semi-autonomous action ​

The AI workflow can take action through tools, APIs, scripts, service accounts, browser sessions, agents, or other integrations.

This includes workflows where a human gives broad approval but the AI performs multiple downstream actions.

Specialized access need ​

The experiment requires specialized model access, high quotas, persistent storage, external vendor access, or other capabilities that increase security, vendor, capacity, or operational impact.

Handoff destinations ​

Depending on the trigger, handoff may involve one or more specialist teams:

  • Dev leadership or the relevant engineering owner;
  • Dev;
  • SRE / DevOps;
  • data platform or data governance owners;
  • product owner or business owner;
  • legal, compliance, privacy, HR, procurement, or finance when their review is required.

For this policy, Dev defines the engineering, security, infrastructure, SRE, and DevOps handoff requirements. Dev will route to other functions when their specialist review is required.

Handoff information ​

A handoff request should give specialists enough context to take responsibility for their domain without losing the product intent.

Useful handoff information includes:

  • summary of the use case;
  • prototype owner or sponsor;
  • users or systems affected;
  • data types involved;
  • model, vendor, or resource requested;
  • production systems accessed or affected;
  • expected AI actions and outputs;
  • known access needs;
  • known risks or failure modes;
  • other teams that may need to be involved.

The handoff information does not need to be perfect. It needs to be clear enough for the responsible teams to evaluate the next step.