Use Fill with Gemini cell sources to review a team’s capability table

The October 28 update adds cell-level web sources; check access first

Google’s October 28 announcement describes web sources attached to AI-function cells in Sheets on the web. It separately dates the gradual rollout from October 20, with up to 15 days for visibility. The announcement date and rollout start differ; a calendar date alone does not prove that every account has the feature on November 3.

The announcement lists Business Standard/Plus, Enterprise Standard/Plus, Google AI Pro/Ultra, Google AI Pro for Education and AI Expanded Access. Actual administrator, language and service-access conditions also matter. Check an existing eligible account and entry point before this five-row exercise, without changing permissions or buying a subscription merely to force a trial.

The official AI-function help describes text outputs with a prompt and optional context range. Here, draft review text inside an existing Sheet. If you first need an editable file generated from chat, use the Gemini task-tracker handoff case, then define the question each cell needs to answer.

Complete input: five deliberately imperfect capability claims

This is an editorial exercise. Candidate claims are neither official quotations nor observed model responses. A GitHub announcement also dated October 28 supplies a public reference, showing how one product name can contain different stages, platforms and client conditions.

The input has only claim_id, draft_claim and official_source. Preserve the specified first-party entry for each record so later human review can locate a particular section. One plausible product overview cannot accept every row in the table.

claim_id,draft_claim,official_source
C01,"Intelligent local model routing is generally available.",https://github.blog/changelog/2026-10-28-github-copilot-fusion-now-supports-selection-of-local-models/
C02,"The managed Windows ML model download also works on macOS.",https://github.blog/changelog/2026-10-28-github-copilot-fusion-now-supports-selection-of-local-models/
C03,"The Copilot app intelligently routes to any compatible BYOK local model.",https://github.blog/changelog/2026-10-28-github-copilot-fusion-now-supports-selection-of-local-models/
C04,"Running inference locally makes the Copilot session offline.",https://github.blog/changelog/2026-10-28-github-copilot-fusion-now-supports-selection-of-local-models/
C05,"Copilot CLI can route supported manually configured local models on Windows, macOS and Linux.",https://github.blog/changelog/2026-10-28-github-copilot-fusion-now-supports-selection-of-local-models/

Give AI the current row and request a short textual draft

In Sheets on the web with this capability available, put the CSV in A:C and name D draft_review. Enter the candidate below in D2, select the cell and use Generate and Insert. Labels can be localized. Fill other rows with their own ranges rather than combining every claim into an untraceable overall verdict.

A2:C2 gives the function this row’s candidate and source. The help says it does not automatically read the entire sheet or other Drive files. Do not ask it to review the material above without providing the required range. The output remains draft text; keep the human decision in another column.

If the function is unavailable, open the same announcement manually and complete the ledger below, recording that the AI step was not executed. Do not fill D with the authored reference answers and present them as actual generated results.

=AI("Review the candidate claim in this row against its official source. Return a short finding, corrected wording and stage/client/platform conditions. If the source cannot be checked, keep the claim unresolved. Do not infer offline operation from local inference.",A2:C2)

Open Sources and make three matches

The official help says to select the output cell, hover over its content and open Sources. Inspect what actually appears. Do not copy the requested URL into a verified-source field merely because it was in your prompt. Missing sources, inaccessible full text or another generation of the feature leave a claim unresolved.

First match the entity: is the page about Copilot Fusion generally or its local model routing? Then match the field: direct model selection, intelligent routing or managed download? Finally match conditions: have the client, platform, experimental settings or device requirements disappeared?

GitHub’s October 28 local-routing announcement supplies this exercise’s reference. The earlier Copilot CLI discovery and session guide addresses the actual discovery entry. This announcement additionally says that adding or discovering a BYOK model does not automatically make it eligible for Fusion routing. Discovery, eligibility and a claim in this table require separate checks.

Complete reference: four corrections and one conditional acceptance

This ledger was derived by reading the original announcement. It does not measure Gemini accuracy. All five candidates concern local models, but they do not receive the same acceptable answer. Treating a Sources icon as a universal pass would miss the distinctions.

Four and one describe only these five deliberately constructed candidates. They are not a real-world error rate or evidence that the product outperforms a human reviewer.

claim_id,review_result,reference_answer
C01,overstated_stage,"The local-routing capability is an opt-in experimental preview."
C02,wrong_platform,"The managed download experience is Windows-only on supported devices."
C03,wrong_client_scope,"The app's intelligent routing currently supports managed Windows ML models only."
C04,wrong_offline_inference,"Local inference does not make the session offline; Fusion still makes routing decisions in the cloud."
C05,supported_with_conditions,"The announcement supports this CLI platform scope for appropriately configured eligible local models."

Repair a false conclusion even when its link is genuine

A faulty teaching draft could attach the genuine GitHub link to every row and mark all five supported. That establishes the existence of a link rather than support for each inference. C04 makes the shortcut visible: removing the cloud-routing qualification from a page about local inference reverses its operating boundary.

Retain the candidate and actual AI draft, then add reviewed_answer, scope and review_state. Do not overwrite the location of the defect. Repair C01–C04 to the narrow answers above, keep C05’s conditions, and ask another reader to locate the basis from the source materials.

When inputs or sources change, decide when to regenerate and review again. Google’s help specifically warns that concatenating cell values into a prompt does not automatically prompt a refresh after those values change. Generated text is not a continuously maintained fact database. Include the review date and source version in the handoff.

Deliver a scope-reviewed table rather than a collection of source icons

Retain all five IDs, candidate claims, actual AI drafts, opened sources, reviewed answers and conditions, review dates and unresolved items. The editorial references here were checked by reading; whether your Google account generates matching text and Sources requires its own observation.

The function returns text. It does not establish account permissions, configure local models or complete software availability acceptance. This exercise also does not make the GitHub capability a KitNelo tool. A real team’s purchase or deployment decision still needs the actual account, client and task at that time.

This five-row case needs no bulk-capacity or performance promise. Start with sources you can open and inspect individually, then decide whether expansion is useful.