A UK operations lead who already spends Monday morning copying figures out of email, into a spreadsheet, and then into a slide deck will recognise the pitch. The new assistant is not another chat window that stops after one answer. Google is offering an agent that takes an objective and works through the steps.
That promise is attractive, and it is also where the risk sits. An objective such as "prepare the monthly sales note" hides a chain of lookups, drafts and files. If the first lookup uses the wrong spreadsheet, the finished note can still look complete. The useful question is not whether the agent can click through more apps. It is which apps it should be allowed to open, and who reads the result.
What Google announced
At its Gemini at Work 2026 event, Google introduced a Gemini agent meant to plan tasks, connect to workplace systems, use tools and return finished work from one interface. The company says a person can assign an objective rather than write an instruction for every step. That is the shift from a chatbot to an agent. Google's account of the event sets out that direction, and TechCrunch reported the business launch on 8 October 2026. MediaPost's report covers the same workflow claim.
For a UK buyer, the announcement is a product boundary, not a reason to connect every login on the first afternoon. The agent is designed to do more than answer a question. Doing more means reading, drafting and, where permissions allow, acting. Those are different levels of trust, and they should be switched on separately.
Where the agent is expected to work
Google describes the agent working across Workspace applications: Gmail, Drive, Docs, Sheets, Slides, Chat and Calendar. It also describes connections to other systems, including Microsoft 365, Slack, Jira, Salesforce, ServiceNow and several data platforms. A firm that already lives in that mix will see the point. Staff spend real time moving the same fact from one login to another. An agent that can read the source and write the destination could shorten that.
Google Workspace's own lesson on this kind of work shows the same shape. Gemini Enterprise is presented as one place that reaches business apps, including Drive and tools such as ServiceNow, and a person builds an agent for a multi-step task rather than a single question. The lesson is on YouTube. It is Google's training film, not a recording of the October 2026 announcement, so it explains the workplace agent rather than standing in for the launch itself.
The quality of the result still depends on the records the agent can see. A CRM that is months out of date, a shared drive full of conflicting versions, or a mailbox that mixes customer mail with supplier noise will produce a confident document built on the wrong material. Connecting more tools does not repair the underlying files. It gives the agent a wider set of places to be wrong. Current, approved company data is the part the business controls, and it matters more than the length of the connector list.
Work that continues after the laptop is closed
Google highlights a further feature: the agent can coordinate specialised sub-agents. Those agents can work in parallel or in sequence on tasks that may take hours or days. The Gemini agent can also keep running in the cloud after someone closes their device, and it keeps the task context so the person can return later. That is useful for a long piece of research, or for a report that has to wait on a data export. It is also a change in supervision. Work that continues while the office is empty needs a record of what ran, what it opened and what it wrote.
Google says the service includes controls for identity, permissions, auditing, sandboxing and spending. Those controls are the practical part of the announcement. Identity decides whose access the agent inherits. Permissions decide which mailboxes, folders and customer records it may open. An audit log is how a colleague reconstructs the work the next morning. Sandboxing and spend limits matter if the agent can call paid tools or external systems. A control that exists in the product and is left on the default is not a control the business has chosen.
The work it could take on
The opportunity Google describes is multi-step office work: researching a topic, preparing a report, analysing data and drafting a presentation. A UK firm might recognise a weekly routine in that list. Pull last week's enquiries, summarise them, and draft slides for the management meeting. That is a fair trial if the source documents are approved and a person reads the draft before it is sent.
It is a poor first trial if the agent can email customers, change a price, approve a refund or alter a live record without a second check. High-impact actions need a person. The agent can propose. The business still decides. That split is easier to keep when the first workflow is narrow: one objective, a named set of folders, and a reviewer who is actually on the rota.
- Write the objective the way you would brief a new starter, including what "finished" looks like.
- Name the files and systems the agent may use, and leave everything else off.
- Mark the actions that must stop for approval: sending, paying, deleting, publishing, or changing a customer record.
- Compare the draft with the log. If the two accounts do not match, do not add the next system.
UK GDPR still applies when the material identifies a customer, a member of staff or a supplier. An agent that reads a mailbox is processing personal data. The questions are the usual ones, asked of a new kind of user: what the purpose is, which systems are in scope, where the processing happens, and who can see the log. A vendor's enterprise label is not that record. The same caution applies to any agent that can act across tools, which is set out in how businesses can reduce the security risks of AI agents.
How a UK firm can start
Begin with one workflow you can describe in a paragraph, then measure it. Keep a note of time saved and of errors. A draft that a manager still has to rebuild has not saved the morning. A draft that is right, and that a person can check quickly, has. That note is what justifies a wider permission later. Firms that plan to roll agents out more widely will also need a governance model, not only a product switch. Deloitte's 2026 findings on agentic AI show how far adoption plans can run ahead of that model.
AI development for a UK business usually starts as one job inside software the firm already runs, with a boundary around the data. Putting that boundary into business software you control is often clearer than giving a general agent the keys to Workspace, the CRM and the finance tool on the same day.
Read the launch as a permission decision
Google has put an agent in front of businesses that can plan, use tools and keep working after the screen is shut. The firm's job is smaller and more specific. Choose the workflow, limit the access, keep a person on the high-impact steps, and keep a record of what happened.
Web Works Rise is a UK-led team in Birmingham, with delivery capacity from a development hub in Tunisia. If the useful version of this is an assistant limited to questions you already answer, or a first automation that stops for a colleague before it sends anything, talk to the team and describe the task. The task is the brief. The model name in the news is not.