AI Enablement Radar week 35: distribute AI like an internal product

AI tools are now distributed as more than licenses. Slack packages skills, Dropbox opens document workflows through MCP, GitHub manages model access centrally, and Anthropic is testing a shared standard for physical equipment. For a smaller team, a prompt, a connector, and a set of permissions have become an internal product. That product needs an owner, a clear approval point, and a way back when something goes wrong.

Top signals this week

• Slackbot adds Skill Sets, allowing several reusable skills to roll out together. Deep Research searches organizational data and the web, while an admin view exposes usage and permissions. This makes knowledge work easier to share and possible to audit.

Source: Slackbot: Your Agentic AI Workspace for Deep Work.

• Salesforce says its Employee Agent reached 70,000 users, handled more than 300,000 conversations in its first year, and achieved a 97.7 percent self-service rate. These are vendor-reported figures, but the operating pattern is useful: Salesforce tested the agent internally, placed it where employees already worked, and kept access tied to existing permissions.

Source: 1 Year, 70,000 Users: What Salesforce Learned as Customer Zero for Employee Agent.

• Anthropic has opened a limited research preview of the Model Hardware Standard, or MHS. It lets agents operate programmable microscopes, liquid handlers, and robotic arms through a model-agnostic interface. The interface can use MCP, the Model Context Protocol, a standard for connecting AI systems to tools and data.

Source: Previewing the Model Hardware Standard.

• GitHub is enforcing a global model policy for Copilot Business and Enterprise. New models, and existing models with no explicit setting, can follow the organization's default policy while deliberate choices remain intact. Open-weight models and models that require data retention are not enabled automatically.

Source: Global model policy generally available.

• The EU AI Act has entered a more practical enforcement phase. The European Commission's updated overview explains how responsibility is shared by the AI Office, the European Data Protection Supervisor, and national authorities. Some enforcement powers began to apply on August 2, 2026.

Source: The enforcement framework of the AI Act.

What organizations are actually doing with AI

Bain has made Claude available across the firm and formed a global partnership with Anthropic for client deployments. This says something about what follows a pilot: access becomes broad, while working methods, training, and support need to fit the roles using the tool. A ten-person team does not need Bain's setup, but it can borrow the principle. Decide what work the agent should help with before issuing more accounts.

Source: Bain & Company announces partnership with Anthropic to accelerate clients' enterprise AI transformations.

Gallup built real-time coaching for thousands of leaders with Amazon Bedrock. The system retrieves answers from research published by Gallup through RAG, or retrieval-augmented generation, which gives the model relevant sources for each question. AWS describes a move from prototype to production in weeks rather than months without a separate MLOps team. The transferable lesson is simpler: connect the agent to a bounded knowledge bank and expose the sources before expanding usage.

Source: Gallup scales real-time coaching for thousands with Amazon Bedrock.

Salesforce's first year also shows why placement matters. Employee Agent sits in the workspaces where questions already arise and can answer questions or complete certain administrative steps. Salesforce keeps sensitive decisions, including hiring and bereavement support, with people. That is a useful boundary: let the agent handle recurring service and preparation, but name the decisions that always belong to a person.

The tooling layer: platforms, agents, and workflows

Dropbox DocSend has opened a remote MCP server in open beta. It connects to ChatGPT, Claude, and Codex CLI through OAuth. An agent can read documents, find Spaces, search visits, and inspect activity. Administrators can disable or enable the connection in company AI settings. For sales and proposal teams, the signal is clear: the document room can become an agent data source, but the connector needs the same owner as the customer workflow itself.

Source: How to connect the Dropbox DocSend remote MCP server.

GitHub's Copilot harness is now generally available in JetBrains. Built-in JetBrains support for MCP remains in public preview, but it can expose IDE capabilities without relying only on external servers. /review can also run inside JetBrains. For technical teams, this means fewer hops between tools. For everyone else, the pattern matters more than the product: place agent tools close to the work and make their boundaries visible.

Source: Copilot harness generally available in Copilot for JetBrains.

Microsoft has lowered the barrier to testing Azure SRE Agent. New customers can trial up to three agents for 30 days without the usual always-on charge; active work still consumes Azure Agent Units. VNet integration is generally available, allowing the agent to reach private resources through existing network controls. Live Reports is in public preview and refreshes operational data whenever a saved report opens.

Source: Try Azure SRE Agent with no always-on charges.

Google Workspace is testing Gemini-based data classification in Drive. An administrator describes the classification levels, and Gemini proposes labels without requiring the organization to build a manual training set first. This can give agent workflows a better signal about which files they may read or process. The feature is an open beta and should be checked against the organization's own information model before it governs access.

Source: Gemini-based data classification in Google Drive is now available in open beta.

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

When a skill can be shared with one click and an MCP connector can expose real customer documents, reviewing the prompt is not enough. Review the whole chain: who can install it, which sources the agent reads, which tools it can run, which outputs are retained, and who approves the action.

AWS describes a concrete technical pattern for agent tools: inspect inbound data, check parameters before a tool runs, and validate the tool response before it returns to the agent. Bedrock Guardrails can work alongside regular expressions, schema validation, and lists of allowed values. The principle applies outside AWS too. Store authentication in environment variables or a secret manager, use scoped OAuth or API permissions, redact sensitive fields, and keep audit logs for runs.

Source: Extend Amazon Bedrock Guardrails to Tool Interactions Using the Strands Agents SDK.

The EU enforcement framework makes documentation more useful than a general AI policy. Record the model and version, the sources made available to the agent, the permissions in force, and the person who approved the result. This creates a run receipt that both the business and a future reviewer can understand.

This week's practical Hammer test

Run a 40-minute distribution rehearsal

Choose an AI routine that already works for one person, such as summarizing a customer meeting, checking a proposal, or sorting support requests. The goal is to see whether two colleagues can use it without the working knowledge remaining with its creator.

1. 0–7 minutes: Write down the trigger, sources, and concrete output. Use a real but bounded work case.
2. 7–15 minutes: List tools and access. Start with read permissions. Put authentication in environment variables or a secret manager, not in the prompt.
3. 15–23 minutes: Set a human release gate. Customer messages, payments, deletion, access changes, and external publication require explicit approval.
4. 23–32 minutes: Run one normal case and another with a missing source. The agent should stop and identify what is missing rather than filling the gap itself.
5. 32–40 minutes: Save a short run receipt: user, model, sources, tools, approver, result, exception, and rollback path.

The rehearsal passes when both colleagues can run the routine, see the same sources, and understand exactly where human judgment begins. If they cannot, you have found what needs to be built before a broader rollout.

Companies and tools to watch

• Slackbot: Skill Sets turn internal AI workflow distribution into a product feature.
• Salesforce: Employee Agent provides unusually concrete, if self-reported, usage and self-service figures.
• Anthropic MHS: shows how familiar integration questions move from documents and APIs to physical equipment.
• Dropbox DocSend: makes customer and investor documents accessible to agents through OAuth and MCP.
• GitHub Copilot: the global model policy shows how model selection becomes a central administrative concern.

If you want to move from a working personal routine to a shared, reviewable workflow, that fits Hammer's Tool Forge. We start with one concrete distribution rehearsal, set access and approvals, and only build further when the run receipt holds up.