Shadow AI Mitigation & Executive Steering Framework

Case study 01 · Business Operations Manager · Keshet Media Group · 2026

A single light glowing in a dark room

Context

Rapid adoption of generative AI across business units created workflows that sat outside R&D oversight.

Challenge

40+ employees were building AI workflows with no standard security protocols, budget lines, or data-protection rules.

My role

I owned
Choosing the approved AI tools, together with the CIO. Then the route for tools already in use without approval: a new approved solution for each one, and the governance around it, from the approval workflow and budget ownership to the security boundaries and the rollout.
Scope
60 people in the rollout, 40+ workflows

Actions

  • Helped establish an executive AI steering committee with the CIO and senior leadership.
  • Designed the approval workflow defining budget ownership and security boundaries.
  • Built a phased rollout to move unmonitored workflows into official, compliant tools without hurting productivity.
  • Reviewed ~750 license seats against actual usage ahead of renewal.
  • Ran a platform cleanup that archived or consolidated 900 project boards, and removed tens of excess admin-level permissions across document sites.

Results

  • 100% transition: All 40+ workflows brought under IT/R&D security compliance.
  • Zero security incidents in the first four weeks under the new data-privacy boundaries.
  • Approved tools took over: use of the approved AI tools rose after the rollout.
  • 7 redundant tools cut through unified AI tool licensing.
  • 900 project boards archived or consolidated.

The hard call

Whether every AI tool and workflow should need central approval. Approving everything centrally was the tighter control, but it would have made the approved route the slow one and pushed people back to personal accounts. I drew the line at money and security instead: anyone can build with the approved tools and share what works, and anything company-wide, expensive or touching sensitive data goes to the steering committee. The rollout also started small, with a first wave of builders, so the decision to widen it would rest on usage data rather than opinion.

What I’d do differently

I’d write the intake requirements and the rollout plan together, from the first week. A requirements document is a starting point, not a shipped system; the rollout is where governance actually gets tested.

Artifacts

The request flow before and after the intake path, the scale of the AI use-case mapping alongside the maturity model, and the license picture behind the unified AI tool licensing. A generalized version of the intake system, built on synthetic data, is the AI intake agent demo. Where the idea went next: Citizen AI Governance, three coded projects on governing the tools employees build and use.

Ad hoc requests vs. the structured intake path
BEFOREChat messageinformal askManager / platformteamno defined processDuplicate buildsUnclear ownershipRisk unweighedAFTERPlain-languagerequestsubmitsStructuredintake formchecksDuplication +risk scoreroutes by capacityRight team,right owner
Use-case mapping & AI maturity

80+

use cases identified across departments

~90%

of departments assessed

AI maturity by department

EarlyDevelopingAdvanced
Most departments
Engineering (slightly ahead)

Real figures from the engagement, not illustrative.

Microsoft license tier breakdown
E5
~685
E3
~65

Out of ~750 total seats

Real figures from the engagement, not illustrative.

Copilot pilot-to-target
Pilot (current)
50
Target
250

Pilot is 20% of the expansion target

Real figures from the engagement, not illustrative.

More work

All case studies