AI Enablement Radar week 30: build AI workflows around a clear handoff

AI tools moved closer to the work this week. They now sit inside issue trackers, meeting folders, coding workflows, and analytics views. That makes them more useful, but it also exposes weak handoffs. When AI passes on a draft, classification, or change, the next person needs to understand what happened, which evidence was used, and what still needs a decision.

Top signals this week

• Hippo says more than 90% of employees used Claude or its other AI tools during the first month. Sixty percent used them daily, and the same share reported saving at least two hours per week. The company links adoption to training and real departmental work, while underwriters and claims professionals retain decision authority. Source: Hippo Holdings via PR Newswire.
• Brown Health has more than 400 active Dragon Copilot users and has built over 24 Copilot Studio agents for emergency guidance, translation, scheduling, policy lookup, and operations. It did not begin with a tool free-for-all. A focused AI Center of Excellence spent its first six months on governance and a risk-based review model. Source: Microsoft Customer Stories on Brown Health.
• GitHub Copilot's cloud agent can now take a Linear issue, work in an ephemeral GitHub Actions environment, open a draft pull request, and request human review. The handoff is the useful part: progress stays visible where the task already lives, and the result returns as a reviewable proposal. Source: GitHub Changelog on Copilot and Linear.
• GitHub Issues is previewing automations that attach confidence and rationale to changes. Administrators can let routine changes through while holding less certain suggestions for review. GitHub also says the approval feature is a workflow convenience, not a server-side security boundary. Source: GitHub Changelog on agent controls in Issues.
• The European Commission's new guidance for Article 50 of the AI Act puts a near date on transparency work: the obligations apply from 2 August 2026. They cover areas including direct AI interaction and certain AI-generated or manipulated content. Source: European Commission guidelines on transparency.

Attach AI to a defined job and require a delivery that a colleague can check. An agentic workflow, in this article, means a process where AI chooses several steps or tools to reach a goal instead of merely answering one question.

What organizations are actually doing with AI

Brown Health built adoption around clinical workload

Brown Health chose documentation as its first clear problem. Dragon Copilot captures conversations and helps draft clinical notes while the clinician can stay focused on the patient. The organization then built agents for policies, emergency care, patient pathways, and operations. That is a better adoption pattern than giving everyone a blank chat window and hoping inspiration follows.

For a Nordic business or school, the translation is simple: choose a workload that people can already describe. It might be meeting follow-up, proposal preparation, lesson adaptation, or first-pass support triage. Connect the tool to the place where the work is already managed, then name the person who receives the result.

Source: Brown Health scales Microsoft Dragon Copilot and AI agents to ease care delivery.

Hippo paired broad access with clear human responsibility

Hippo's early numbers come from an internal survey and should be read as the company's own measurement, not an independent productivity study. The usage pattern is still worth examining. Claude and other tools support underwriting, claims, audit, marketing, sales, and product work. AI helps with research, analysis, and more frequent checks, while professionals approve decisions and changes.

This is where many license rollouts fail. Access is not enough. The employee needs a task, relevant sources, and a clear answer to one question: "What should I do with the AI result when it comes back?"

Source: Hippo Rolls Out Claude and Other AI Tools Company-Wide.

AXA puts AI inside tools people already use

AXA is progressively rolling out Microsoft 365 Copilot in Teams, Outlook, and Word, three years after launching its internal Secure GPT service. The company is pairing the rollout with practical and ethical upskilling and human oversight for important decisions.

It looks ordinary. When AI sits inside an established workflow, the team can spend its time on working practice and quality instead of getting people to remember one more separate app.

Source: AXA accelerates its AI workplace strategy.

The tooling layer: platforms, agents, and workflows

GitHub makes agent work visible where the task begins

The Linear integration shows a useful principle even for teams that do not write code: the task, status, result, and review request should stay connected. The agent works in a bounded environment and returns a draft pull request rather than making an invisible production change.

GitHub's new Issues controls add rationale and confidence. A confidence rating is the agent's own estimate of how certain it is about an action. It is not proof of quality, but it can help a team sort which suggestions deserve the most attention.

Sources: Copilot cloud agent for Linear. Agent automation controls in GitHub Issues.

MCP becomes easier to scale and test

Model Context Protocol, or MCP, is an open protocol that gives AI systems a common way to connect to tools and data sources. GitHub has adapted its MCP server for the next specification, whose core becomes stateless and removes the previous sessions and initialize step. Official conformance tests also make it easier to check whether a client or server follows the specification.

Smaller teams do not need to rebuild everything now. The useful signal is that integrations are gaining more mature testing and scaling patterns. Ask a supplier to show how the connection is tested, how permissions are scoped, and what gets logged when the agent uses a tool.

Source: GitHub MCP Server supports the next MCP specification.

Zapier mixes rules and AI in one workflow

AI by Zapier lets a team add AI steps to ordinary Zap workflows. Rule-based parts can remain deterministic, while AI handles tasks that require interpretation, such as classifying a support ticket or summarizing a customer request. Agentic steps can use selected app actions, web search, and sources in Google Drive, Notion, or Zapier Tables. Structured results can then feed the next step.

Use fixed rules where the answer should be the same each time, and use AI where language or context needs judgment. Give the agent scoped permissions and add approval before sending messages, changing customer data, or taking other hard-to-reverse actions.

Source: AI by Zapier: Add agentic AI steps to your workflows.

Notion, Google Meet, and Shopify add operational context

Notion now shows Workers usage in its credits dashboard. Workers run custom code on Notion's infrastructure for data synchronization, Custom Agent tools, and webhook automations. Visible consumption lets a team connect a workflow to actual cost before the beta ends.

Google Meet is starting to collect meeting notes, transcripts, and recordings in a dedicated Meet folder in the host's Drive. Recurring meetings share a subfolder. That sounds trivial until an old script searches for "Meet Recordings"; Google therefore recommends that administrators review automations that depend on folder names or IDs.

Shopify now lets compatible apps add event annotations directly to analytics charts. Campaigns, launches, and supplier changes can appear beside sales, conversion, and fulfillment data. Annotations do not prove causation, but they make the week's events visible when someone interprets the result.

Sources: Notion Workers in the credits dashboard. Google Meet organizes meeting artifacts in Drive. Shopify annotations.

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

AI governance is the set of decisions, roles, and controls that determines how an organization may use AI. Two sources above provide a useful minimum this week: GitHub distinguishes workflow approval from a real security boundary, while the European Commission explains which transparency requirements begin on 2 August.

Make labeling and handoff part of the workflow itself. An AI draft should carry its sources, status, and next human decision. An integration should use scoped API keys from environment variables or a secret manager, then log which tools were called. You can give an agent enough access to do useful work without giving it free rein.

Google Cloud's CodeMender uses the same construction in a security workflow. The agent can scan code, verify suspected vulnerabilities by running generated tests in a customer-managed sandbox, and produce tested patches. A developer reviews and approves before anything is committed. A sandbox is an isolated environment where code or tools can run with limited impact on other systems.

Sources: European Commission guidance on Article 50. CodeMender in Google Cloud.

This week's practical Hammer test

Run a 40-minute handoff test. Choose one recurring workflow where AI already produces a draft, or where someone manually moves information between two tools. A meeting that should become tasks works well, as does a support message that needs classification and a suggested response.

1. Spend five minutes naming the starting point and the person who receives the result.
2. Spend ten minutes defining which evidence AI may read and which tools it may use.
3. Spend ten minutes defining the delivery: draft, status, sources used, uncertainties, and next decision.
4. Run one real example for ten minutes. Ask a colleague to review it without verbally explaining the workflow.
5. Use the final five minutes to rewrite the handoff wherever the colleague had to guess.

Paste this as the working instruction:

Map this task as an AI-supported workflow. Describe the starting point, allowed sources, allowed actions, delivery format, and who approves the next step. Then create a handoff template containing: work completed, sources used, changes or proposals, uncertainties, pending decisions, and log information. Mark which steps may run automatically and which require human approval. Do not change any systems during this test.

The test succeeds when the recipient can make the next decision without hunting for context or asking what the agent did.

Companies and tools to watch

• Brown Health: shows how a focused center of excellence can begin with governance and then build agents around real workload.
• Hippo: worth following as early self-reported time savings meet longer use and measurable business quality.
• GitHub: connects the task, agent status, proposal, and human review inside the same workspace.
• Zapier: makes it easier for nontechnical teams to add AI where rules are insufficient without rebuilding the entire automation.
• European Commission: Article 50 makes transparency a current implementation task rather than a future policy document.

If you want to build this kind of handoff in a real customer, document, or operations workflow, Hammer's Tool Forge is the relevant next step. We connect AI to existing systems with clear access, review points, and run logs that the team can actually use.