Google’s Gemini agent: identity, audit trails and spend caps for AI agents

Global reference

• Vendor: Google Cloud
• Topic: Gemini agent, a general-purpose AI agent for work
• Key date: Announced 8 October 2026 at Gemini at Work 2026
• Status: Announced. Financial Services and Legal versions are in preview. No pricing or general release date given

What Google announced

• One agent that answers questions, carries out tasks and writes and runs code. Users give it objectives, not step-by-step instructions.
• It runs in the cloud, so work that takes hours or days continues after the user logs off. Work can be scheduled or triggered by events.
• It works inside Google Workspace, Microsoft 365 and Slack, and connects to systems such as Salesforce and ServiceNow.
• It uses Google’s Gemini models and Claude models from Anthropic, with other models to follow.
• Google says it can read data from SAP, Workday and Salesforce without copying it.

The controls

• Identity: every agent has its own identity, governed like an employee, with least-privilege permissions.
• Audit trail: every action is logged and attributed to the agent, not to a person.
• Spend caps: a hard limit on AI spend per project. When the cap is hit the agent pauses until someone resumes it. Costs are tracked per project, so they can be charged back to departments.
• Network: agents run in a sandbox, and their traffic passes through a gateway that enforces company policy.

The financial services version

• In preview. Google first announced it in August 2026.
• It draws on FactSet, LSEG, S&P Global and SEC filings, and shows source citations, data lineage and confidence scores.
• It is aimed at investment research, credit analysis and risk modelling.
• Google names CME Group and Deutsche Bank as users.
• Customer results quoted by Google: Commerzbank cut manual document review from 20 hours to one, and Bradesco cut document review from one hour to five minutes. These figures are reported by the vendor.

What to do now

• Add AI agents to user access reviews and the segregation of duties matrix.
• Give every agent a named human owner and an approval limit.
• Set a budget cap for each AI project before the pilot starts.
• Ask every vendor whether each agent action is logged, and whether the log can be exported for audit.

Our take

The models matter less than the controls. An agent with its own login, limited access and a full log can be supervised like a junior member of staff. An agent that borrows a person’s login cannot. Make identity and logging a condition of any agent pilot.

Important. This summary is for general information and is not legal, audit or financial advice. It is based on public Google Cloud publications, together with press reports, and may be incomplete or out of date. The official publications prevail; check the official source before acting.