AI Enablement Radar week 36: when AI gets permission to act

The week's most important AI change is practical: tools are gaining more ways to act inside real work. Google Workspace Studio can move files and reply in Gmail or Chat. GitHub Copilot can approve a pull request when an administrator enables that option. As AI moves from writing suggestions to performing actions, teams need to decide which steps may run directly, which require a person, and what receipt the system should keep.

Top signals this week

• Google Workspace Studio added steps to copy or move Drive files, reply in Gmail, and send replies in Google Chat. Administrators can disable steps or require approval when data is shared externally. Source: Google Workspace Updates on the new automation steps.
• GitHub Copilot code review can now submit an approval that counts toward a repository's required approvals. The feature is off by default, can be limited to selected file paths, and its approval is dismissed when new code is pushed. Source: GitHub Changelog on Copilot approvals.
• DXC Technology has rolled Amazon Quick out to more than 115,000 employees in 70 countries. DXC's lesson is that logins measure use, while AI fluency shows whether the work has changed. Source: DXC's conversation with AWS about the global rollout.
• Google found four recurring patterns among strong entries in its AI Agents Challenge: bidirectional MCP, event-driven concurrency, identical validation for primary and fallback models, and tiered routing. One team let a deterministic first pass handle more than 40 percent of messages before a larger model call. Source: Google Developers on four agent engineering patterns.
• The European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act after the service reported at least 45 million monthly EU users. Its additional obligations are due by January 2027. Source: The European Commission's DSA decision.

For a smaller company, a school, or any team without a dedicated AI department, the useful question is concrete: what permission should AI have inside this particular process? It leads to a workable design faster than another debate about which model to buy.

What organizations are actually doing with AI

DXC measures AI fluency in the work

DXC is not treating its Amazon Quick rollout like another SaaS launch. The platform reaches more than 115,000 employees, but the company puts changes in work ahead of active-user counts. One example is an agent for resource requests. Another handles lead routing for sales. Both take on manual steps that already existed in the business.

That measurement approach is worth borrowing. Track an operating result such as time to response, correctly handled cases, or how often a person must redo the AI's work. A login count says almost nothing about value.

Source: DXC, "Co-innovating enterprise AI."

The University of Nevada grounds its student agent in official sources

The University of Nevada, Reno has launched Ask Wolfie for enrolled students. The agent answers questions around the clock about housing, events, student services, and other campus topics. It is trained on university information and points students to official web sources.

The assignment is tightly defined: answer recurring questions quickly and leave staff more time for individual conversations. A school testing the same model should begin with a defined information set, clear citations, and a simple route to the right person.

Source: The University of Nevada, Reno on Ask Wolfie.

The NFL connects incidents to one operating workflow

Microsoft describes how the NFL built a Game Ops Dashboard with Copilot Studio and Power Platform. The system standardizes incident reporting on game days and makes prior experience reusable. The process design matters. An agent gathers context. Rules and permissions govern the flow, while people handle exceptions.

For a smaller team, the equivalent could be complaints, absence cases, quotes, or fault reports. Start where several people already need the same facts to make a recurring decision.

Source: Microsoft 365 Blog on AI and business processes.

The tooling layer: platforms, agents, and workflows

An agentic workflow is a process where AI can choose and perform steps through tools rather than only writing an answer. This week's product updates bring those steps closer to everyday work.

Workspace Studio can now take concrete actions in Drive, Gmail, and Chat. GitHub lets administrators give Copilot a bounded right to approve code. These are two versions of the same design: the action lives in the tool, while permissions and conditions sit around it.

MCP, or Model Context Protocol, is an open standard for connecting AI systems to tools and data sources. Google's review shows why the connection alone is not enough. Strong systems filtered data before it reached the model, ran independent tasks concurrently, and used the same quality check when a fallback model took over. This is familiar system discipline applied to agents.

Sources: Google Workspace Updates on new Studio steps. GitHub on governed Copilot approvals. Google Developers on MCP, fallback models, and routing.

Google also added audit logs for Gemini Notebook in the Workspace Admin console. The records can include user identity, IP address, visibility, and resource context, and administrators can export them to BigQuery. This gives teams a way to follow AI use without building a separate control surface from scratch.

Source: Google Workspace Updates on Gemini Notebook audit logs.

Governance and risk: what needs to be in place before scaling

AI governance is the set of decisions, roles, and controls that determine how AI may be used and reviewed. It does not have to begin as a thick policy document. For one focused workflow, access, approvals, and logging cover much of the practical ground.

Anthropic introduced Enterprise Frontier Safeguards after working with more than 100 customers. The design lets customers keep monitoring data in their own cloud accounts under their own encryption keys and access policies. Automated alerts go to the customer's reviewers. This makes it easier to use the model in real work without moving data custody and review away from the organization responsible for them.

Source: Anthropic on Enterprise Frontier Safeguards.

Microsoft's guidance for edge AI states a simple principle: model output may recommend an action, but a deterministic component outside the model should authorize what actually runs. That component can allow selected actions, constrain parameters, limit call frequency, and release credentials only when its conditions are met. Irreversible or sensitive actions get separate human approval.

In practice, use environment variables or a secret manager for keys, scoped permissions, redaction for fields the model does not need, approval points, and a run log. Passwords do not belong in the prompt, and an agent should not receive broader access than its task requires.

Source: Microsoft Security on securing edge AI in customer-owned environments.

The European Commission's DSA decision applies at the platform level, but the deadline also gives buyers a useful check. A provider's risk work, transparency, and incident handling become inputs to your own vendor assessment. Ask how data is retained, how tool calls are logged, and who can stop an automated flow.

Source: The European Commission's decision on ChatGPT, Reddit, and Roblox.

This week's practical Hammer test

Spend 40 minutes mapping the action boundary in an existing workflow. Choose something that already recurs, such as sorting incoming customer questions, moving approved documents, or preparing follow-up after a meeting.

0–10 minutes: Write down the actions that actually happen. Put them into four levels:

• read and gather context
• write a recommendation
• perform a reversible action
• perform an action with an external or financial consequence

10–20 minutes: Give AI direct permission for no more than one reversible action. Mark the steps that always need a person, such as sending externally, deleting, publishing, paying, or approving a binding decision.

20–30 minutes: Set the technical boundary. Use one scoped integration, keep keys in environment variables or a secret manager, and redact fields the task does not need. Decide what source evidence the agent must show before the action.

30–40 minutes: Run one real example and save a short receipt: time, source, proposed action, performed action, reviewer, and outcome. If the receipt does not explain what happened, the workflow is not ready for reuse.

Paste this into the AI tool your team already uses:

Help us map the action boundary for the workflow below. Begin by asking for any missing systems, data sources, and rules. Then sort the steps into: read, recommend, reversible action, and action requiring human approval. Propose the minimum permissions, the fields to redact, and what a run receipt must contain. Do not change or execute anything without our explicit approval.

Companies and tools to watch

• Google Workspace Studio, because routine document and message actions can now sit inside Workspace with central administrative controls.
• GitHub Copilot code review, because AI approvals make permission design as important as the review itself.
• DXC and Amazon Quick, because their rollout distinguishes AI fluency from basic usage statistics.
• The University of Nevada and CollegeVine, because Ask Wolfie is a concrete model for source-grounded student service.
• Anthropic Enterprise Frontier Safeguards, because customer-owned logs and keys may make more regulated workflows possible.

If the test shows that the workflow is worth developing, Hammer Automation's Tool Forge can help connect the data, permissions, approvals, and logs into something the team can use day to day.