AI Enablement Radar week 43: make your AI routine ready to hand over

This week's most useful AI updates make a good working routine easier to share. Agent Skills gains an independent home, Google connects generated apps to Sites, and Asana brings more steps into one workflow. For a Nordic team, the next question is concrete: can a colleague use the routine without its creator sitting beside them? This Radar covers October 19–25, 2026. Undated customer examples and standing guidance are marked as background.

Top signals this week

• On October 22, Anthropic announced the transfer of Agent Skills to the Agentic AI Foundation. A skill is a package of reusable instructions and resources for an agent. Independent governance may support reuse, but does not guarantee that every skill behaves identically everywhere. Source: Anthropic: Agent Skills and AAIF.

• On October 19, Google announced that interactive apps from Gemini can be exported to a new, private Google Sites draft. Publishing remains a separate step. That gives a team somewhere to review and share a tool. Source: Google: Gemini apps to Sites.

• On October 22, GitHub began exposing names of custom skills, agents and MCP servers in usage reports. MCP, Model Context Protocol, connects AI to tools and data sources. The names make it easier to identify which internal routines people use. Source: GitHub: custom names in usage metrics.

• Asana's October 22 quarterly roundup describes a visual automation builder and native round-robin assignment of incoming tasks. The roundup does not establish that every feature launched that day. Source: Asana: Q3 2026 release.

What organizations are actually doing with AI

These customer stories have no clear publication date. They are useful examples, not confirmed launches this week. The accounts come from Anthropic and the customers themselves.

Carvana puts assistance in Slack. The company replaced two custom bots with Claude Tag. Its workflow moves from alerts to investigation and code proposals, with human review before production. Data access is scoped by team. The lesson for a smaller organization is to put assistance where questions already arise and keep the working instructions in a shared location. Source: Anthropic: Carvana.

Pictet combines access with practical training. Around 700 people have access to Claude Code and Cowork. Partner Artefact has trained more than 500 people through 25 hands-on workshops. Access is not the same as active use, but the approach shows how training can follow real tasks. In a school, two colleagues could try the same document routine and compare where the instructions need clarification. Source: Anthropic: Pictet.

Qonto separates preparation from decisions. Its AI Lab built agents to prepare transfers, draft invoices and analyze transactions. Actions affecting finances or customers require user confirmation. This is a useful pattern for administration: let AI prepare the material while a named person approves the payment or customer commitment. Source: Anthropic: Qonto.

The tooling layer: platforms, agents, and workflows

Reuse the instructions; check the connections. The Agent Skills transfer includes the name, web domain and GitHub organization. Versioned releases and security guidance are on the proposed roadmap, not completed capabilities. When moving a routine between tools, you still need to check its data sources and permissions. Source: Anthropic: Agent Skills and AAIF.

Give a generated app a reviewer. Export from Gemini creates a new Sites draft and requires permission to create and edit sites. Existing Sites and Drive policies apply. Try an internal quote calculator with known examples, check its calculations, then decide who should be able to use it. Source: Google: Gemini apps to Sites.

Build reports against the right data. Gemini in Sheets can now analyze a full connected BigQuery dataset, rather than just visible rows. BigQuery is Google's service for analyzing large datasets. Gemini presents a plan for review before creating connected formulas, pivot tables and charts. Rollout began October 19 and can take up to 15 days. Access to the underlying data and an eligible Workspace edition are required. For teams already using BigQuery, a recurring sales report is a reasonable test. Source: Google: BigQuery analysis in Sheets.

Asana brings assignment and follow-up together. Its quarterly update describes AI Teammate connections to services including Google Drive, Gmail, Slack and HubSpot. Start with a recurring request: who receives it, what evidence do they need, and where does the next step appear? Check plan eligibility and actual access before making the routine depend on a feature. Source: Asana: Q3 2026 release.

Development teams get more deployment choices. Google's Agents CLI has an extension system; version 1.8.0 supports a first-party LangChain extension and deployment to Cloud Run. The extension system remains experimental. Google recommends pinning the CLI version. For the buyer, the question is simple: can someone maintain the solution without depending on one developer's laptop? Source: Google: Agents CLI extensions.

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

AI governance here means assigning responsibility for use, access and follow-up. GitHub's new model policy API can read and change which models are allowed at different organizational levels. That makes policy management programmable; it does not determine which model performs a task best. Source: GitHub: model policy API.

NCSC's new AI chapter in its research problem book, published October 20, addresses questions including how human oversight can work when AI operates at speed. It is a research agenda, not a new rule. The European Commission's AI Act overview provides standing context: documentation, logging and human oversight are among the requirements for high-risk systems. Classify the use before translating requirements into your routine; not every internal AI tool is a high-risk system. Source: NCSC: AI research chapter.

Source: European Commission: AI Act.

In practice, give the routine an owner, scope its access and store keys in environment variables or a secret manager. Log runs without secret values. Keep an explicit approval step when the routine publishes, pays or makes a commitment to a customer.

This week's practical Hammer test

Set aside 40 minutes for a handover test. The aim is for a colleague to run the same routine without extra verbal instructions. This is our suggested exercise, not a measured vendor result.

1. First 10 minutes: choose a routine, such as preparing a quote brief from a customer inquiry. Write down the required inputs, where the result belongs and who approves it.
2. Next 10 minutes: run the routine on a known example. Save the instructions, sources and an approved example result in the team's shared workspace.
3. Next 10 minutes: have a colleague run the routine on another example without your help. Ask them to mark unclear steps and missing information. Nothing should be sent to the customer during the test.
4. Final 10 minutes: fix the instructions where your colleague got stuck. Assign an owner and decide how to label the next version. Assess whether someone can review the result without opening the original chat.

Use this prompt after selecting your material:

Read our instructions and example result. Help me make the routine usable by a colleague: specify the required inputs, steps, expected result and what a human must approve. Ask for missing information rather than assuming it. Then suggest a second test case. Do not change or send anything in connected systems.

GitHub's new names in usage reports may later help larger teams see which routines people use. Historical reports are not backfilled, and some clients do not report names. Usage also does not establish that the result was correct. Source: GitHub: custom names in usage metrics.

Companies and tools to watch

• Agentic AI Foundation: follow how the skills standard is maintained rather than assuming full compatibility.
• Google Workspace: try the Sites draft as a review location for an interactive tool.
• Asana: investigate whether incoming work can be assigned without extra manual steps.
• Pictet: include practical peer training in adoption, not just licenses.

If the handover test shows you need to connect more systems, Tool Forge can help build and maintain the routine. Start with the step where your colleague actually got stuck.