AI Enablement Radar week 40: review AI changes where the work happens

AI suggestions are becoming easier to review where the work already happens. In week 40, Google's document APIs gain comments and suggested edits, while GitHub adds workflows with review stages. For a team spending too much time correcting AI answers, that is more useful than another chat: have AI show exactly what it wants to change, with evidence, before someone approves the result.

Top signals this week

• On September 30, Google announced comment support in the Docs, Sheets and Slides APIs. Docs also gains suggested text changes. The rollout may take 15 days. Source: Google Workspace on comments and suggestions.
• On October 1, GitHub introduced dynamic workflows in Copilot CLI, the app and SDK as a public preview. They can combine sequential and parallel stages with review checkpoints. Source: GitHub on dynamic workflows.
• Google starts rolling out reusable skills in Workspace on October 5 and in the Gemini app on October 13. Instructions do not sync between these environments. Source: Google Workspace on skills and Gems.
• Shopify launched Canvas on October 1, with direct editing, Sidekick chat and working store previews. Translations and app blocks are among the unsupported features at launch, which matters for Nordic merchants. Source: Shopify on Canvas.

What organizations are actually doing with AI

Scribd classifies documents at volume

In a September 25 customer story, Scribd and Google Cloud describe classifying more than 400 million documents. Gemini 2.5 Flash Lite performed the classification, with Gemini 2.5 Pro judging a second check. This is older context for the week, not a new release.

For an organization with a document archive, the useful idea is to sort the material before building more search or chat features. Start with a representative sample and check the categories manually. A second AI judge does not replace ground truth, and the story does not publish complete classification accuracy metrics.

Source: Google Cloud and Scribd on document classification.

Qonto prepares financial work for approval

Qonto describes agents that prepare invoices and transfers, while users approve actions with financial consequences. The customer story has no visible publication date and is included here as an adoption example. It shows a clear division of work: AI prepares the transaction, the person makes the decision. Qonto's time savings are vendor-reported rather than independent measurements.

Source: Anthropic on Qonto's financial administration.

KDDI sets quality thresholds with the product owner

KDDI's knowledge app Buffmee uses more than 100 content sources. In a September 9 technical customer story, KDDI describes product owners calibrating quality thresholds against a sample of answers. That provides a useful method for anyone building an internal knowledge assistant: decide which errors are unacceptable and test for them before rollout. Evals are systematic tests of AI results, rather than simply asking whether an answer sounds good.

Source: Google Cloud and KDDI on Buffmee and evaluation.

The tooling layer: platforms, agents, and workflows

Put the suggestion in the document

Google's update gives integrations a clearer place to leave work for review. An API is a programming interface that lets systems work with each other. Here, a workflow could propose a text change in Docs instead of overwriting the original. This does not mean every ready-made connector in Make or other tools already supports the feature. Check the actual integration before planning the workflow.

Source: Google Workspace on the new API capabilities.

GitHub's dynamic workflows address a similar need in coding: gather evidence, perform the task and have a separate stage review the result. An agentic workflow lets AI use tools and work through multiple steps. The new preview makes the sequence and review points more explicit, but does not prove that results will automatically be correct.

Source: GitHub on workflows with structured results.

Check instructions and defaults

Google's skills use the Markdown format SKILL.md for reusable instructions. The route for Gems in Workspace Studio is also changing: new flows can no longer use "Ask a Gem". Existing steps continue for now, but for business users they will stop working no sooner than March 1, 2027. Inventory the routines that depend on Gems rather than moving instructions without checking their dependencies.

Source: Google Workspace on skills and the change to Gems.

On October 2, GitHub enabled requests for Copilot code reviews through REST and GraphQL APIs. Since September 28, "Default" means "Balanced"; an explicit choice of "Lite" is preserved. That affects how teams should compare reviews before and after the update.

Source: GitHub on the review API and default effort level.

On October 2, GitHub also removed selected models from Copilot, including Gemini 3.5 Flash, Gemini 3.6 Flash and Claude Opus 4.7. Check model choices and your organization's allowed models in recurring jobs. This announcement concerns Copilot, not a global shutdown by the model providers.

Source: GitHub on models removed from Copilot.

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

AI governance means responsibility, rules and follow-up for how an organization uses AI. The European Commission's AI Act overview describes a risk-based framework with specific requirements for high-risk systems and transparency requirements for uses including chatbots. This is existing context, not a new rule from this week. Distinguish between suggesting a wording change and influencing recruitment decisions, for example; the same routine will not suit both.

Source: The European Commission's AI Act overview.

For this week's test, assign a reviewer and limit access to the relevant material. If you build an integration, store keys in environment variables or a secret manager, grant only the permissions it needs and log proposed and approved changes. Redact information the log does not need. That makes the routine possible to operate and troubleshoot.

This week's practical Hammer test

Set aside 40 minutes to improve an existing document, such as customer instructions or a procedure for colleagues. The goal is a reviewable change proposal. You do not need to wait for the new APIs; the test works with a document copy and comments today.

1. Prepare the material, 5 minutes. Choose the document and a current factual source. Decide who can approve changes.
2. Request suggestions, 10 minutes. Use the AI tool your team already has. Ask for up to five changes, each with original text, proposed text, a reason and a source.
3. Review, 15 minutes. Have a colleague assess each proposal. Mark it accepted, rejected or needing more evidence. Pay particular attention to numbers and terms.
4. Save the result, 10 minutes. Apply approved changes to the copy. Note how many suggestions were usable and how much review time they needed. Save the instruction if it is worth reusing.

Use this instruction:

Read the document and factual source I attach. Propose up to five changes that make the instructions easier to follow. For each change, show the original text, proposed text, reason and supporting evidence from the source. Keep numbers and terms unchanged unless the material explicitly supports a change. Flag anything you cannot determine. Do not edit the original.

A good result is not the largest number of edits. It is a set of suggestions the reviewer can assess without starting the whole document again.

Companies and tools to watch

• Google Workspace: check when comments, suggested edits and skills become available in your environment.
• GitHub Copilot: follow how the preview's review stages work for your recurring tasks.
• Shopify Canvas: watch for translation and app-block support before planning a store migration.
• Qonto: use its division between preparation and approval as a comparison when assessing finance agents.

These watch points follow the linked product announcements and customer examples above. If the document test works but gets stuck in manual copying, Tool Forge can help connect sources, proposed changes and approval within your existing workflow.