AI Tool Governance Program

Overview

Maintain the AI tool-approval program for the organization — deciding which AI tools are sanctioned on managed devices, enabling the approved ones properly, and restricting the rest to limit company data egress.

Problem

AI tooling arrives from the bottom up. Engineers adopt a coding assistant because it makes them faster, and the adoption decision gets made long before anyone asks where the code in the prompt window ends up. Proprietary source, customer data, and internal documentation move to a third party through a browser tab, and there is no procurement record of it happening.

The failure mode on both sides is real. Block everything and you lose genuine productivity while pushing usage onto personal devices where there is no visibility at all — the worst of both outcomes. Allow everything and company data leaves the environment on terms nobody reviewed.

Approach

Run it as an approval program rather than a blocklist, so there is a sanctioned path that is genuinely good enough to use.

GitHub Copilot and Claude Code are approved for engineering use, deployed and managed through the normal endpoint tooling rather than installed ad hoc.

Unapproved AI tools are restricted on managed endpoints, so the default for anything not reviewed is unavailable rather than silently permitted.

Apple Intelligence is handled deliberately at the enrollment layer — its setup step is skipped during device provisioning rather than left to a user decision during onboarding, keeping OS-level AI features from being enabled before the organization has reviewed them.

New tools enter through review rather than through whoever installed it first, and the sanctioned list is maintained as tools mature rather than frozen at whatever was approved once.

Architecture

Restrictions and enrollment behavior ship as configuration profiles through Jamf, applied to managed macOS endpoints. Approved tooling deploys through the same managed pipeline as any other sanctioned application, so an approved AI tool is an ordinary managed app rather than an exception.

Outcome

An organization with a real, usable AI toolchain for the people who need it and a defensible answer for where company data does and does not go. Engineering keeps the productivity; the review step keeps the data question from being answered by default.

Lessons Learned

Governing AI adoption is a supply problem, not an enforcement problem. Restrictions alone push people to their phones and personal accounts, which is strictly worse than sanctioned usage because it is invisible. The approval list only works because the approved tools are ones people actually want to use — the moment the sanctioned path is worse than the unsanctioned one, the policy is decorative.