AI Enablement Radar week 28: give AI a real assignment and a fair test

AI was given bigger assignments this week. OpenAI launched ChatGPT Work for projects that can run for hours, UST is putting Claude into industrial processes, and Alberta is using around fifty agents to review government code. The interesting part is not that agents can do more. Good adoption now starts with a real assignment and a test that shows whether the result holds up.

Top signals this week

• ChatGPT Work takes on longer workflows. The agent can work across apps and files, create documents, spreadsheets, and presentations, and pause for questions and approvals. OpenAI says more than five million people use Codex each week and more than one million use the technology outside software development. (Source: ChatGPT is now a partner for your most ambitious work)
• UST combines the technology with broad training. The company plans to train 20,000 engineers, architects, and consultants on Claude. In its iDEC platform, the existing closed-loop pipeline has cut chip validation cycles by 50 to 70 percent, from four days to 48 hours, and Claude is now being added as the reasoning layer. (Source: UST is bringing Claude to physical AI)
• Alberta shows how a large assignment can be split up and reviewed. A team used around 50 Claude Code agents to scan 466 million lines of code in 20 hours. Each finding cited the exact file and line, and people approved patches before deployment. (Source: Government of Alberta uses Claude to find and fix cybersecurity vulnerabilities)
• GitHub now measures what happens after code is written. The Copilot usage API reports time to first review and the number of review cycles for pull requests that actually merged. Those measures are closer to business value than prompt counts or active users. (Source: Add review cycles and time to adoption phases in the usage API)
• The EU is linking AI adoption with cyber capability. The Commission's new action plan includes capacity to evaluate models before market entry and a secure testing platform for critical sectors. (Source: EU Action Plan on Cybersecurity and Artificial Intelligence)

What organizations are actually doing with AI

UST's example is unusually concrete. Claude Code will read schematics and pinouts, write regression tests, run them, and compare measurements with a digital twin. Human approval remains in place in healthcare and telecom workflows before a recommendation reaches a member or an action moves forward. This is an agentic workflow: AI handles several steps and tools toward a goal, while people own the decision points.

Source: UST is bringing Claude to physical AI

Alberta also chose a bounded but real assignment: find, explain, and propose fixes in legacy code. The agents first used rules to find known patterns, then performed a second review with exact citations. That made the result checkable. The province is also training employees and the public through the Alberta AI Academy; more than 10,000 members of the public have used the platform.

Source: Government of Alberta uses Claude to find and fix cybersecurity vulnerabilities

OpenAI describes the same move in office work. In early tests, RingCentral used ChatGPT Work to review launch plans, Jira tasks, and schedules and went from supporting one product manager to around 50. Virgin Atlantic cut a comparative analysis from weeks to hours. NVIDIA replaced an Excel workflow that took roughly 40 percent of its preparation time for GTC.

Source: ChatGPT is now a partner for your most ambitious work

The lesson for a Nordic team is fairly plain: do not set "use AI more" as the goal. Pick a job with clear inputs, a visible output, and a person who can decide whether the work is good enough.

The tooling layer: platforms, agents, and workflows

GPT-5.6 is becoming the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork. The model is entering tools many organizations already pay for, rather than another separate pilot environment. Model changes and output quality now need to be tested inside existing document and analysis routines.

Source: GPT-5.6 is now the preferred model in Microsoft 365 Copilot

Voice is taking on a different role too. GPT-Live can listen and speak at the same time while delegating search or heavier reasoning to another model in the background. API access is announced but not yet available. For schools, field teams, and people who prefer speaking to typing, the signal is worth watching: voice can become a work interface, not just dictation.

Source: Introducing GPT-Live

OpenAI's audit of SWE-Bench Pro is this week's useful reality check. An eval is a repeatable test of how well an AI system performs a defined task. After agent-assisted analysis and review by five experienced developers for every flagged task, OpenAI estimates that around 30 percent of the benchmark tasks are broken. A precise score from a poor test is still poor evidence.

Source: Separating signal from noise in coding evaluations

GitHub is also making operations easier to manage. Organizations can direct OpenTelemetry data from Copilot in VS Code and CLI to an approved collector, including whether prompt, response, and tool content may be captured. Settings can be deployed through Intune, Jamf, Group Policy, or a locked configuration file.

Sources: Enterprise-managed OpenTelemetry export for VS Code and CLI and Deploy managed Copilot settings via MDM in VS Code and CLI

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

The EU AI Act becomes fully applicable on August 2, 2026, with exceptions and later dates for parts of the high-risk rules. Transparency obligations also begin applying in August 2026. For most everyday workflows, the practical starting point is straightforward: list where AI is used, who owns it, which sources and systems are connected, and where a person reviews or approves the work.

Source: AI Act

Controls do not need to make AI toothless. Keep keys in environment variables or a secret manager, give connected systems scoped permissions, redact sensitive fields, require approval before external sends or writes, and keep logs that can be reviewed. The agent can then do useful work without putting passwords or broad access rights into a chat.

Code that adds AI features also needs ordinary security testing. CodeQL 2.26.0 includes a query that detects when untrusted user input can flow into a system prompt and influence model behavior. It also covers more prompt surfaces in SDKs from OpenAI, Anthropic, and Google.

Source: CodeQL 2.26.0 adds Kotlin 2.4.0 support and AI prompt injection detection

Anthropic's new reflection view points to a softer control: look at what AI is actually used for. It summarizes usage patterns across 1, 3, 6, or 12 months and connects them with the Delegation, Description, Discernment, and Diligence framework. A team version of the same question is worth asking: which tasks did we hand over, and did the work improve?

Source: Introducing a way to reflect on how you use Claude

This week's practical Hammer test

Set aside 40 minutes and give AI a real assignment with a fair test. Use a recurring job your team already knows how to judge, such as meeting follow-up, reviewing a proposal, a lesson overview, or a weekly customer list.

1. Write the assignment in five minutes. State the desired result, the sources AI may use, and the required delivery format.
2. Write three acceptance criteria. For example: every claim has a source, no required field is missing, and the next owner is named.
3. Run the workflow for 15 minutes. Connect a relevant tool if needed. Grant only the access the job requires and add approval before external publishing, sending, or system changes.
4. Review for ten minutes. Someone who knows the job marks each criterion pass or fail and records what needed correction.
5. Choose the next step. Keep the workflow if it saves time without lowering quality. Adjust the test if the criteria were unclear. Stop if the benefit is not visible.

Starter prompt to copy:

Help us with a recurring workflow. Start by asking for the goal, allowed sources, required delivery format, and the person who approves the result. Then propose three concrete acceptance criteria that a human can check. Do not run the workflow until the criteria are approved. When the job is complete, report which sources you used, which assumptions you made, and what the reviewer needs to inspect.

Companies and tools to watch

• OpenAI ChatGPT Work: longer assignments across files, apps, and scheduled events.
• UST: a clear example of training and workflow integration moving together.
• Government of Alberta: large-scale agent use with verifiable findings and human-approved patches.
• GitHub Copilot: measurement, central settings, and telemetry are becoming one operating surface.
• EU AI Office and ENISA: planned testing capacity may shape how European organizations trial advanced models.

If you want to turn one test into a connected and managed workflow, this fits naturally with Hammer's Tool Forge. Start with the job and its acceptance criteria. Choose the tooling afterward.