SignalsOperating intelligence
Open navigation

Operating question

Useful AI work becomes easier to trust when a person can inspect the source, choose the form of the output and join the work before a consequential step is taken.

Human-Centered Architecture

3 Things AI: The “Make the Evidence Easy to See” Edition

3 Things AI 5 min3 sources

For

Leaders and workflow owners

You will leave with

3 operating decisions

Reading mode

5 min · 3 verified sources

Reading guide3 decisions · 4 sections+

Decision points

  1. 01Test cell-level citations by measuring supported rows and review minutes before moving generated research into a working system.
  2. 02Compare transcript formats against one human-checked recording and score names, numbers, speakers, omissions and editing time.
  3. 03Replay a closed incident in a shared agent channel with a named owner, allowed sources, a stop condition and no deployment authority.

Companion tool

Human-in-the-Loop Review Map

Preview

Three practical tests for cited spreadsheet research, fit-for-purpose transcripts and shared agent work before they become everyday habits.

The most useful AI feature this week may be the one that makes a result easier to question. New tools for spreadsheets, transcripts and shared agent work all put more context beside the output. The opportunity is to test whether that context helps a real person make a better decision.

1. Make spreadsheet research easier to challenge

A sales coordinator is preparing a list of 300 potential distributors. The spreadsheet can summarize each company quickly, but a polished row is not useful if nobody can tell which website supplied the market, product or ownership detail. The practical bottleneck is verification, not filling cells.

On October 28, Google said Fill with Gemini in Google Sheets now places web-source citations in individual cells. Google also expanded a request to as many as 3,000 cells and reported an internal 95-person study in which a 500-cell task ran up to 55 times faster than manual entry. The rollout is gradual for eligible Workspace and consumer plans.

The benefit is a shorter path from a generated claim to the page that supports it. The tradeoff is that a citation can make weak research look finished. A source may be old, promotional or about a different company with a similar name. Google's speed result is an internal comparison, not a forecast for your workbook.

Take 20 rows from a completed research sheet. Ask the tool to rebuild only the factual columns, then have the person who uses the list open every citation and mark it supported, unclear or wrong. Compare accepted rows and review minutes with the old method. Do not import the test into the CRM. Availability depends on the account and rollout, and cell-level links do not establish that every generated interpretation is correct.

2. Choose the transcript your team actually needs

A bilingual project meeting includes names, part numbers, interruptions and a decision that changes halfway through the call. One colleague needs a verbatim record; another needs readable sentences with speakers and timestamps. Treating both needs as one generic transcript creates cleanup and disagreement later.

NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weights speech model that uses a natural-language instruction to select verbatim or written-form transcription, speaker attribution, translation and word timestamps. NVIDIA says it supports 40 languages and recordings up to one hour, and published model weights, a licence and a fine-tuning toolkit.

The benefit is the ability to ask for a transcript shaped for the job without training a separate model for each format. The tradeoff is new operational work: an open model still needs suitable hardware or hosting, privacy decisions, quality testing and someone responsible for corrections. Vendor benchmark results do not tell you how it handles your accents, jargon or noisy rooms.

Use one consented, non-sensitive 15-minute recording that already has a human-checked transcript. Test a verbatim instruction and a readable meeting-record instruction. Score names, numbers, speaker changes, omissions and editing minutes; keep the original audio and human record as the reference. The release does not prove the model is accurate enough for legal, personnel or customer commitments.

3. Give shared agent work a shared room

A developer asks an internal agent to investigate a failed deployment. The agent gathers logs and proposes a configuration change, but the service owner sees the work only when a long private session is forwarded. By then, the wrong assumption has shaped the whole recommendation.

Slack made its Code API generally available on October 28. A custom agent can open a dedicated code channel, add people and files, post progress and artifacts, accept steering from teammates and archive the channel when work ends. Slack says the agent inherits the channel's permissions and administrative controls; some related pull-request and overview features are described as coming soon.

The benefit is earlier participation by the people who know the system and will approve the change. The tradeoff is more channels, notifications and retained technical context. Visibility is useful only if the right owner joins before the decision, and a familiar chat interface does not make an agent's proposal safe.

Replay one closed, low-risk incident. Have the agent create a channel with the incident owner, its allowed data sources, a stop condition and a rule that it cannot merge or deploy. Measure how quickly someone catches a bad assumption and whether another teammate can understand the final decision without reading every message. The API provides a collaboration surface; it does not prove that your agent, permissions or review practice are reliable.

The bigger pattern

AI output is easier to use when its supporting context travels with it: a source beside a cell, a transcript format chosen for the reader or a shared place to inspect an agent's work. More visible context still needs judgment, but it can make review faster and disagreement more specific.

Where does your team lose the most time today: finding the source, reshaping the output or reconstructing how a decision was made?

Verified sources

Continue your decision path

Move from understanding to action.

01 · Apply

Human-in-the-Loop Review Map

Turn this edition's decision points into a concrete working plan.

02 · Go deeper

3 Things AI: The “Keep the Next Step Reviewable” Edition

Three practical ways to keep AI-assisted work reviewable: name the next step after a meeting, write the rule before choosing a guardrail and test one company-data question.

Read next
03 · Assess

Apply this signal to your architecture.

Identify the workflow, context, and controls to structure first.

Open Architecture Assessment