AI Enablement Radar week 44: measure time to approved AI work

AI can produce more cells, files and presentations for your team. This week's useful question is how quickly you can decide whether that material is fit to use. Google is bringing sources closer to spreadsheet content, GitHub is coordinating multiple models, and pharmaceutical companies are showing what happens when AI output meets real review. These are the signals from October 26–November 1, followed by a test that measures time to an approved work product.

Top signals this week

• Google Sheets is adding web sources for individual generated cells in Fill with Gemini and support for up to 3,000 cells per request. Rollout began October 20; the announcement arrived October 28. A citation makes an answer easier to check, not automatically correct.

Source: Google Workspace: Fill with Gemini citations.

• GitHub Copilot Fusion became generally available on October 28. Automatic model routing can combine multiple models on one task. Count the cost of the whole task, not just the first model.

Source: GitHub Copilot Fusion.

• Gemini can create, preview and edit files directly in chat, according to Google's October 29 announcement. That shortens the path to an editable work product.

Source: Google Workspace: create, preview and edit files in Gemini.

• On October 27, Enveda reported faster manufacturing-document review when reviewers received AI comments beforehand. The result concerns a bounded experiment, not pharmaceutical review in general.

Source: Anthropic: Early progress on accelerating drug development with Claude.

What organizations are actually doing with AI

Enveda: read the result alongside the test conditions

Anthropic describes an Enveda test on 1,061 pages of manufacturing documentation. Claude found the eleven issues human reviewers had identified and flagged seven more. According to the vendor's report, human review time fell from over 25 hours to under three. That is a result from the test, not a guarantee that every future run will find every error. Humans still own the quality decision.

For a Nordic business handling proposals, quality records or supplier documents, the transferable idea is to let AI flag what a reviewer should inspect. Then measure both missed errors and review time.

Source: Anthropic: Early progress on accelerating drug development with Claude.

Generate Biomedicines and Bristol Myers Squibb: test against operational outcomes

Anthropic's October 27 report describes six Generate Biomedicines scientists using Claude Code to propose CAR T-cell designs. One component became roughly 70 percent shorter, and laboratory tests showed around ten times greater receptor production. These are preclinical results, not evidence of a finished treatment. Bristol Myers Squibb uses Claude to analyze proprietary data and inform clinical-study plans; its example is qualitative, not a controlled productivity measure.

The lesson for other teams is methodological: a polished report is not enough. Decide which downstream task should improve and test that outcome.

Source: Anthropic: Early progress on accelerating drug development with Claude.

The tooling layer: platforms, agents, and workflows

Sources and editable files closer to the work

Fill with Gemini shows cell-level web citations in Sheets on the web. It supports larger requests, but start with an amount you can review. Access includes Business Standard/Plus and Enterprise Standard/Plus. Check that the feature has reached your account before booking a team test.

Source: Google Workspace: Fill with Gemini citations.

Gemini's new file workflow includes PDF preview and editing on the web and mobile. Documents, spreadsheets, presentations, Markdown, text, and code are handled on the web. One useful test is to create a supplier comparison that opens in Sheets, then have a colleague check its conclusions there.

Source: Google Workspace: create, preview and edit files in Gemini.

On October 30, Google announced editing across multiple slides or entire presentations with Gemini in Slides. Changes remain editable and come with a summary for review. The launch is English-only, and rollout may take more than 15 days from October 29. Do not plan a Swedish workflow as though support were already generally available.

Source: Google Workspace: edit entire presentations with Gemini.

Multiple models need a shared quality test

Copilot Fusion can run a single model, cascade between models or have models critique one another's answers. Administrator model policies constrain the selection. Multiple models also mean multiple billed model contributions. Compare cost per accepted change.

Source: GitHub Copilot Fusion.

GitHub Agentic Workflows also became generally available on October 28. An agentic workflow lets AI use tools and take multiple steps toward a goal. Here, Markdown instructions compile into GitHub Actions. Agents default to read-only permissions and use controlled outputs for changes. Recurring issue triage is one possible use, subject to the same review requirements as other automation.

Source: GitHub Agentic Workflows generally available.

Google's October 28 guide shows how Agents CLI can generate, grade and compare tests of agent behavior. These tests are often called evals. The tool has no built-in scheduler; recurring runs need something such as cron or Cloud Scheduler. Keep individual test cases so that an improved average does not hide an important failure.

Source: Google: Agents CLI Evaluation.

Mistral released open weights for Mistral Large 4 on October 29. That offers a self-hosting option, but a trillion-parameter model is not an obvious local solution for an ordinary office. Its license also requires a separate commercial agreement for businesses offering models as a service. Distinguish open weights from unrestricted use.

Source: Mistral Large 4 open weights.

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

AI governance here means the decisions defining who owns the work, what the system may do and how results are approved. Two older references remain useful; they are not new rules this week.

The EU General-Purpose AI Code of Practice is a voluntary route for model providers to demonstrate AI Act compliance. Its documentation component helps buyers ask concrete supplier questions. Signing it is not blanket certification of your own workflow.

Source: European Commission: General-Purpose AI Code of Practice.

The UK Government's AI Playbook emphasizes human control, lifecycle responsibility and appropriate assurance. Translate that into operations: scoped permissions, secrets in environment variables or a secret manager, redacted logs and approval before external commitments. Keep an audit log showing what was actually sent or changed.

Source: UK Government: AI Playbook.

This week's practical Hammer test

Set aside 40 minutes to test an AI-generated decision brief, such as a supplier comparison or a review of course options. Use a tool you already have access to. This is Hammer's suggested exercise, not a vendor benchmark.

1. First 5 minutes: choose five questions the brief must answer. Write down what counts as a correct answer and which documents may be used.
2. Next 10 minutes: ask AI to create an editable brief with a source for each factual claim. Mark gaps rather than filling them with assumptions.
3. Next 15 minutes: have a colleague check five claims against their original sources, including one that affects a purchase or another decision. Record corrections and review time.
4. Final 10 minutes: compare with your usual approach. Keep the routine if time to an approved brief improves without more missed errors. Otherwise, change the instruction or source selection and test again.

Use this instruction after selecting and connecting the right material:

Read the selected sources and create an editable decision brief for our five questions. Cite a source for each factual claim. Separate documented facts from your proposals and mark missing information. Finish with the facts a colleague must check before making a decision. Do not send or publish anything.

A person makes the decision and approves external use. AI can prepare the brief and revise it after review.

Companies and tools to watch

• Enveda: watch whether document review works as well across campaigns beyond the current evaluation.
• Google Workspace: check Sheets citations and Slides language support in your actual account.
• GitHub: track the cost of the whole Fusion task, not just an individual model.
• Mistral: review operating costs and licensing before choosing a self-hosted model environment.

Once the test works, Tool Forge can help connect it to documents, permissions and review in daily work. Start with the brief that currently takes too long to approve.