AI Enablement Radar week 32: how AI becomes shared team capacity

One person with a good AI chat can save an hour. A team gains real capacity when the work has shared sources, the right access, a reviewer, and a metric everyone can see. Week 32 offered several concrete examples of that move. AI is entering documents, support queues, and developer tools while organizations start operating it as a shared workflow.

Top signals this week

• Karlsruhe Institute of Technology moved from uncontrolled AI use to a governed toolbox for 25,000 students and 10,000 employees. Access requires a qualification module, and faculty built 31 teaching chatbots during the first seven days. Source: SAP on KIT's AI program.

• Mobileye built an internal support agent for routine questions that previously required 15 clicks across several systems. The company reports that the agent automates 66 percent of ticket volume, has a 98 percent success rate, and responds in roughly one minute. Source: AWS and Mobileye.

• Adobe now brings more than 70 document, design, and video tools into one plugin for ChatGPT Work and Codex. A team can move from a spreadsheet or raw asset to a reviewable document without a long chain of exports. Source: Adobe.

• Agent Plugins 1.0 describes how skills and MCP servers can be packaged for several agent clients. MCP, the Model Context Protocol, is a standard for how AI clients discover and use external tools and data sources. Source: Google Developers Blog.

The common thread this week is quite ordinary: companies get more done when AI work becomes visible, shared, and measurable. The license is the starting point. How the team works with it determines whether the result can be repeated.

What organizations are actually doing with AI

German insurer DOMCURA chose a bounded document workflow. Its KIM platform reads claim documents, assists with filing, and can pay qualified claims in about ten minutes. DOMCURA reports a 50 percent reduction in operating costs. The word "qualified" matters because people still handle exceptions and harder judgments.

Source: Microsoft's DOMCURA customer story.

Target did a different kind of preparation. It brought transaction, graph, vector search, and full-text search workloads into Spanner Graph before extending its conversational Gift Finder. Target says the simpler data platform cut developer time spent on database administration and upkeep by 50 percent. A smaller retailer does not need the same technology. It should clean up product data, permissions, and sources before letting AI face customers.

Source: Target Engineering via Google Cloud.

An agentic workflow lets AI perform several steps and use tools toward a goal while the organization sets access and decision points. DOMCURA and Mobileye show why the definition is useful: automation works independently inside a known case, while people handle exceptions and final judgments.

The tooling layer: platforms, agents, and workflows

AI work is increasingly landing where colleagues already work. Google Drive now attaches comments to an exact moment in a video. It sounds simple, but it fixes a common handoff: a reviewer can point to the right frame instead of writing "somewhere near the middle" in an email.

Source: Google Workspace Updates on timestamped video comments.

LangChain goes further with Managed Deep Agents in public beta. The service packages long-running agents with memory, isolated sandboxes, traces, schedules, and channels such as Slack. It can remove a lot of infrastructure work for a technical team, though the beta status and current US region mean a pilot needs explicit acceptance criteria.

Source: LangChain on Managed Deep Agents.

Two infrastructure signals are worth reading together. Agent Plugins 1.0 makes skills and tools more portable across clients. The stateless version of MCP, now a release candidate, removes session dependencies and places routing and audit information in HTTP headers. That could make remote tools easier to operate in load-balanced and serverless environments. The specification is not final, so test it behind an explicit version boundary.

Sources: Agent Plugins 1.0. MCP stateless update.

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

AI governance means the practical rules that determine who may use what, with which data, under what review, and how the run is followed up. This week's control signals make useful integration easier to deliver professionally.

GitHub has made centrally managed MCP lists generally available for the Copilot app, Copilot CLI, and VS Code. Administrators can allow or deny remote servers by URL and local servers by exact command and arguments. Malformed or unverifiable policies fail closed. That gives a team a real control point for tools that could otherwise spread user by user.

Source: GitHub Changelog on MCP allowlists.

The EU AI Office is also treating real-world usage and behavior data after launch as a complement to pre-deployment testing for general-purpose AI models. For an ordinary buyer, the translation is simple: keep a short run log, track failures and human corrections, and decide when the workflow should be paused or changed. This is taskforce guidance, not a new binding rule.

Source: European Commission on the fourth GPAI taskforce meeting.

Microsoft's updated Zero Trust assessment shows how a large program can be split into manageable stages. Its DevSecOps section contains 15 control groups and 91 tasks, with four covering AI-assisted development: code governance, tool allowlisting, data protection, and AI system supply chains. A smaller organization does not need to copy the whole catalog. Start with permissions, secret management, approval points, and a log people can read later.

Source: Microsoft Security Blog on Zero Trust for AI.

This week's practical Hammer test

Set aside 40 minutes and turn a private AI routine into a shared work card. Choose recurring work that can be judged, such as drafting a first customer reply, reviewing a video, or summarizing an internal support ticket.

1. Spend five minutes on the baseline: how long does the work take, which sources are used, and who approves the result?
2. Spend ten minutes on access: list the documents, systems, and actions AI needs. Use scoped permissions and keep keys in environment variables or a secret manager.
3. Spend ten minutes on the output: write two acceptance criteria and one clear stop condition.
4. Run one representative case for ten minutes. Save the sources, draft, corrections, and elapsed time.
5. Use the last five minutes for the decision: ready for three more runs, needs revision, or should stop?

Paste this as the work instruction:

Help us make this recurring AI routine shareable across the team.

Task: [describe the workflow]
Approved sources: [list documents and systems]
Allowed actions: [read, summarize, draft, or other]
The result is accepted when: [two measurable criteria]
Human approval is required before: [publishing, sending, changing, or buying]
Stop and flag when: [missing source, conflicting facts, or other]

First produce a short plan. Then run one representative test case and finish with a run log: sources, actions, uncertainties, human corrections, and elapsed time.

A good result is not an impressive answer. It is a work card a colleague can run again and assess in the same way.

Companies and tools to watch

• KIT, for pairing mandatory foundational knowledge with faculty-built assistants.
• DOMCURA, for the clear boundary between qualified claims and human exceptions.
• Adobe, because document and media production is moving into the same agent environment.
• GitHub, for centrally governed MCP servers with fail-closed enforcement.
• LangChain, for packaging long-running agent execution as a managed service.

If the test shows that the routine deserves a real integration, Hammer Automation's Tool Forge can help shape the access, workflow, review, and measurement without making the process heavier than it needs to be.