The most useful sentence in an AI report may be “I couldn’t verify that.” Google’s October 13 announcement of Gemini 4 Argon says the system can work through long, mixed-format documents and report when it cannot find something or finish a task. For a business leader assigning a document review, that second behavior is as important as the summary. A colleague can act on an incomplete report if the gaps are visible. A polished answer that quietly fills them in is harder to use safely.

In brief: When you ask an AI agent to review contracts, accounts, policies, or customer records, require two outputs: findings tied to source documents and a separate list of claims it could not verify. Have the person who owns the decision review both before the result travels farther.

The missing document changes the decision

Imagine a team comparing a supplier’s renewal proposal with its existing agreement. The AI reads the proposal, last year’s invoices, and a folder of correspondence. It returns a tidy summary: price has risen, the service level looks unchanged, and renewal appears reasonable. The summary may be correct on every document it opened. But the folder lacks a signed amendment that changed the service commitment in July.

That example is illustrative, not a reported Google customer case. A summary can be accurate about the available files and still be wrong about the agreement. The reviewer needs to see the boundary of the search.

Google says Argon can read material combining text, tables, charts, and images and produce a structured answer grounded in those sources. It also says the model was trained to report negative findings honestly. Those are vendor claims, not proof that a team’s contract review will be complete. The question is whether the missing amendment appears as a gap, or disappears behind confident prose.

A good handoff sounds less like “renewal approved” and more like this: “The proposal increases the annual fee; here are the two pages that establish the old and new amounts. The documents provided do not establish whether the July service change was signed. I did not review the vendor’s current terms outside this folder.” Now the procurement lead knows what remains to be checked.

Treat absence as a finding

Most teams ask an assistant for conclusions. Fewer ask it to distinguish between a fact it confirmed, a fact it inferred, and a fact it could not check. That distinction can fit in an ordinary working document. Put a source link or file name next to each material claim. Then add a short section called “Not verified” for missing records, conflicting numbers, unreadable files, and questions that need a person.

This is not a demand that the AI prove a universal negative. “I did not find an amendment in the files I was given” is a bounded statement. “There is no amendment” is a claim about the whole world of company records. One is useful evidence about the assignment; the other may be a guess wearing a suit.

Google’s release cites ServiceNow’s early-access account: on its own enterprise task testing, Gemini 4 Argon achieved roughly 20% better task success than Gemini 3.8 Flash, with particular gains in HR and customer service. That comparison belongs to ServiceNow’s testing, as quoted by Google. It does not measure how often an arbitrary business report catches missing documents. It is a reason to run a bounded trial, not a reason to skip review.

For a smaller trial, use an old decision whose outcome is known. Give the assistant only the files a colleague would normally have at the start. Ask it to produce a recommendation, a source list, and an explicit account of what it could not establish. Have the process owner mark which gaps were correctly identified, which claims were unsupported, and whether any missed source would have changed the recommendation. Keep the human review step in place even if the first outputs look good.

Give the agent a colleague’s standard

When I hand a research task to a person, I don’t want a beautiful document that hides where they got stuck. I want to know which records they examined and which phone call or missing file stands between us and a decision. An AI collaborator should meet the same standard. It does not need to mimic a person’s voice or claim certainty it has not earned.

The assignment can be plain: “Compare the renewal proposal with the signed agreement and invoices. Quote the source for every price or service claim. If a document is missing or the files disagree, say exactly what you could not resolve. Draft a recommendation; do not send it to the vendor or treat it as approved.” The point is not a clever prompt. The point is agreeing on what a reviewable result looks like before anyone relies on it.

Then decide who owns the unresolved items. The AI can find a discrepancy. Procurement can request the signed amendment. Finance can confirm the invoice history. The person authorized to renew makes the call. That division of work is more useful than asking whether the agent can “do procurement” on its own.

Google says the new model is available in Gemini for business and through its developer platform. Availability makes this a current operating question, not a future one. But a launch announcement cannot tell you whether your own records are complete or your team will notice a bad conclusion. The first test is whether the agent can show its work and say where that work ends.

For the wider question of how people and agents share responsibility, see Human-AI Collaboration in Practice. A report that names its missing evidence gives the next person somewhere to start. A report that conceals the gap gives them a decision they may have to undo.

Research and draft: Mai. Editorial direction: John Lipe.