At Gemini at Work 2026 on October 8, Google Cloud chief executive Thomas Kurian introduced the Gemini agent, which Google describes as "a single, universal agent for work." The agent runs persistently in the cloud and ships with a shared memory and personalization graph, built from four kinds of AI agent memory: session, semantic, procedural, and episodic. Workers can chat with it, assign it objectives, schedule tasks, or trigger it from events. Jobs that take hours or days keep running after the laptop closes. (Futurum Group)
The timing is pointed. Days earlier, research from Templafy found that AI without access to company knowledge is creating an "AI productivity tax." In its survey of 2,000 knowledge workers in the UK and US, 96% of UK respondents still edit AI-generated business documents, and 35% say editing them can take longer than creating them from scratch. (UK Tech News)
The two announcements describe opposite sides of the same problem: the industry is giving agents memory, but most companies have not given their agents anything worth remembering.
Agents are getting memory
The Gemini agent is Google's attempt to collapse enterprise AI into one interface. It works across Google Workspace, Microsoft 365, and Slack, connects to a tools registry spanning Salesforce, ServiceNow, Jira, Snowflake, and Databricks, and spins up temporary sub-agents for individual jobs.
The most consequential piece is the coworker agent: a persistent agent with its own Google Workspace identity, email address, calendar, Drive, and storage, acting under its own identity rather than the user's. Google paired it with cryptographically attested agent identity, OAuth-propagated permissions, and action-level audit trails attributed to the agent.
Google also introduced a Knowledge Catalog that maps business definitions such as net margin and reads metrics from Databricks, dbt, LookML, and SAP. Bloomberg Media reported a 63% lift in SQL query accuracy during initial development, evidence that grounding agents in company definitions matters more than bigger models.
The scale numbers come from existing deployments, not the new agent. Google reported that nearly 90% of the Fortune 100 use Gemini Enterprise and that nearly 500 Google Cloud customers have each processed more than one trillion tokens. Commerzbank cut manual document review from 20 hours to one; SOMPO has deployed more than 10,000 custom agents. Google disclosed no pricing or general availability dates for most of the new announcements.
The AI productivity tax
Templafy's report, The Business-Ready AI Gap 2026, finds AI embedded in daily work: 54% of UK respondents use it to create business documents at least once a day. But 51% say reviewing AI-generated work leaves little actual time saved, and 18% of knowledge workers spend six or more hours per week reviewing, editing, validating, or correcting AI output.
The root cause is a context gap. Forty-two percent of UK knowledge workers say their AI tools lack sufficient access to approved company content, while 38% say their tools struggle to incorporate it. The most common correction is accuracy of facts, numbers, and claims (40%), followed by generic phrasing (36%).
"AI may be fast to ask, but it is still slow to finish," said Christian Lund, Templafy's co-founder. His answer is not a more powerful model but an orchestration layer that gives AI company context, approved content, rules, and instructions at the point of creation.
Coworker agents are already moving in
Google is not alone. At WebexOne 2026 on October 7, Cisco introduced Claude Managed Agents inside Webex, built on Anthropic's technology. Agents can join spaces, meetings, and calls as standing participants: @mention an analytics agent in a Webex space to pull usage numbers, then ask a presentation agent to drop the figures into a PowerPoint deck while the call is still going. (Startup Fortune)
The pattern is clear. Agents are becoming persistent colleagues with identities, email addresses, and memories. What they lack is the one thing only the company can provide: an accurate, current, accessible record of how the business actually works.
That is the gap the Templafy data exposes. An agent with four memory types that cannot reach approved company content is a fast drafter and a slow finisher. The businesses that maintain a living record of their decisions, definitions, and context will extract far more from the agent era than those relying on each employee to prompt their way out of it.