Experts adopted an AI agent's coin-flip preference, and Gemini lets people fix its files in the chat

The agent was Claude Code, Anthropic's coding agent, and the people using it were professional data analysts advising an invented farming cooperative. The work held four judgment calls with two sound answers each. A hidden coin flip decided which answer each agent preferred, and years of experience gave the analysts no protection against its influence.

The influence was strongest where it mattered most. When an analyst had stated a view and the agent disagreed, the analyst dropped that view 60% of the time. Most later explained their choices with the agent's arguments and called the result their own.

An agent that does the whole analysis makes these choices along the way and gives a sensible reason for each, so they arrive as finished work, not as questions. A reviewer checks finished work for mistakes, and neither answer is a mistake.

For example, a finance director asks an agent whether to close a regional warehouse. The agent must decide whether the month of a strike counts as a normal month, and whether shared costs are split by volume or by revenue. Either choice can be defended, and either can decide the answer. The report she receives shows one choice, with a good reason beside it.

If five teams use the same agent, five reports that agree may be one preference counted five times. Before the agent starts, ask it to list the judgment calls the work needs and to argue the case for each answer, then make those calls yourself.

Until now, Gemini could create these files but only as downloads. Now the file opens in the conversation: you highlight a paragraph, tell Gemini what is wrong, and it changes that section. PDFs work on the phone too; Docs, Sheets and slides need the web.

Before, the quick way to fix one wrong price in a quote was to ask for the whole file again, and a new version can bring new errors into the parts that were right. A highlighted fix leaves those parts alone, so a draft that is almost right is now worth keeping.

FINRA, which supervises brokerage firms in the United States, already uses AI to summarize thousands of complaints and to check advertising videos for fraud.

Inspection has always relied on samples: an examiner reads a few dozen of a firm's letters, and a problem in the rest usually stays unseen. A reader that never gets tired removes that protection. A misleading line repeated in 40 videos now shows up as a habit of the firm. Auditors and tax authorities will get the same tools, and a company whose own checks still read samples will learn about its patterns from them.

Salesforce and Anthropic announced the plugin in August, and it has been in beta since September for approved companies. It gives Claude 37 ready-made sales skills and lets a seller read and update CRM records by asking in plain words.

When factories first replaced their steam engine with one big electric motor, they gained little, because the belts that carried power to each machine stayed in place. The gains came when each machine got its own motor and the floor was rebuilt around the work. Sales routines were built around screens in the same way: reps fill in fields so that managers can see deals, and the weekly pipeline meeting asks about what the fields miss. If an agent reviews every deal first, the meeting can spend its time on the three deals that need a decision.