3 Things AI: The “Choose the Handoff Before the Tool” Edition

1. Choose what the agent may see before asking it to act

A sales manager wants a short project update before a customer call. The facts are scattered across a shared folder, a chat thread and several emails. Asking an assistant to assemble the update sounds efficient, but the first decision is not how to phrase the prompt. It is which sources belong in the task.

Google announced five new agentic capabilities across Workspace on September 9. Gemini can gather context from selected files, email and chat, then create items such as documents, spreadsheets, presentations or draft messages. Google says the relevant sources are chosen by the user or enabled by an administrator, and the resulting files remain available for review.

The benefit is less manual copying between tools. The tradeoff is that convenience can hide the size of the context window: a useful project folder may sit beside an old forecast, an unrelated personnel note or a draft that was never approved.

For one recurring update, create a small named source set before writing the prompt. Include the current plan, the approved numbers and the latest decision log; exclude private or superseded material. Ask the reviewer to identify one missing source and one the agent did not need.

Availability varies by Workspace plan and feature, and Google’s announcement does not prove that a cross-app draft will be complete or correct for your files. The reasonable test is a bounded internal task where a person already knows the answer.

2. Check the facts again when the format changes

A controller has a dense monthly report that few people read. An audio recap or a clean set of slides could make the numbers easier to absorb. It could also turn a footnote into a headline, separate a chart from its definition or make a preliminary estimate look settled.

Adobe announced new Acrobat Productivity Agent capabilities on September 9. The company says Acrobat can turn document collections into interactive reports, summary slides and audio, and can extract structured information from large sets of files. It also says answers include clickable citations and that customer documents are not used to train Adobe’s generative AI models.

The benefit is reach: the same source material can become a format suited to a meeting, a commute or a quick comparison. The tradeoff is verification work. A polished deck or confident voice can feel more final than the source, while citations help only if someone follows them and checks that the surrounding table, date and qualification survived the conversion.

Pick one report with five known facts, including a caveat, a date range and a number whose unit matters. Transform it into one new format, then have a reviewer trace those five facts back to the original pages. Record whether each stayed accurate, remained in context and was easy to locate. If the caveat disappears, the faster format is not ready for an executive audience.

These are vendor-described capabilities, not independent proof of accuracy across every file type, language or document collection. Sensitive material also needs a separate privacy and access review before it is placed in any AI-enabled workspace.

3. Record the reason behind an expert override

An inventory planner changes a recommended allocation because a supplier mentioned a delay on a call. The spreadsheet records the new quantity but not the reason. A month later, the team can see what happened without learning why the expert departed from the model.

NVIDIA described a supply-chain workflow developed with Palantir on September 10. NVIDIA says planners had information that its quantitative optimizer could not see, including supplier emails, weather, geopolitical events and debrief transcripts. The workflow captures the decision, rationale, expected result and actual outcome, then uses point-in-time backtesting: it replays a past decision with only the information available that day. A specialized model recommends; a planner makes the final call.

The benefit is that expert judgment becomes reviewable material instead of disappearing into memory. The tradeoff is the discipline required to record it. A vague note such as “planner knows best” teaches nothing, while capturing every conversation can create noise, privacy concerns and a maintenance burden.

For one decision your team repeats weekly, add three fields to the existing record: what recommendation changed, which new fact caused the change and what outcome would show the override was helpful. Review five completed decisions after a month. You do not need to train a model; first learn whether the rationale is specific enough for another person to use.

NVIDIA and Palantir are describing their own implementation and development benchmark, not a general result for smaller operations. The approach depends on reliable outcome data, representative history and a task narrow enough to replay fairly.

The bigger pattern

The useful handoff sits around the AI output. Source selection determines what an agent is allowed to know. Fact tracing determines whether a new format preserved meaning. Override records show where human judgment added information the system lacked. Each step adds effort, but it also makes a small experiment easier to improve without pretending the first result is the finished process.

Where does your team lose more time today: gathering the right context, checking a polished output, or explaining an expert override?