The last few months have seen an explosion of new agentic-like AI tools that can automate workflows within productivity applications, but analysts are recommending that enterprises carefully assess deployment risks before adopting them. “Enterprise leaders need to come up with a strategy for thinking about agentic AI in its true sense, and also governance for these solutions,” said JP Gownder, vice president and principal analyst on Forrester’s Future of Work team. Gownder said these tools aren’t truly agentic AI, but “agent-ish” tools, which is how Forrester is terming these applications. “Agent-ish solutions can solve point problems, but they aren’t the ‘digital coworkers’ that vendors are talking about, at least not yet,” Gownder said. Google and Microsoft have been developing AI plugins for their productivity suites, and third parties are providing their own AI plugins for the applications. At last week’s Google Cloud Next, Google introduced new AI tools for its Workspace suite to automate work across productivity apps. Workspace Flows can automate a stream of jobs involved in document reviews, customer support requests, or product analysis. The “help me analyze” feature in Google Sheets can take information from tables and provide instant data analysis and insights. Microsoft celebrated its 50th anniversary earlier this month with a gaggle of unique Copilot automation features that can work with the 365 productivity suite or independently.  The company’s Microsoft 365 roadmap touts an exhaustive list of upcoming Copilot automation features, including automation tools to write Word documents, learning companions based on learning needs, and automated HR tools. This week, Anthropic introduced the integration of its Claude AI model with Gmail and Calendar, in addition to Google Docs, which helps “gain deeper insight into your work context” by searching emails, reviewing documents, and scanning the calendar, the company said via email. “Whether you’re managing projects, planning trips, or handling personal and family tasks, Claude now understands your full context to provide better assistance. Navigate to settings to connect to the Google Workspace beta,” Anthropic said. And Adobe is developing an Adobe Express AI agent for Microsoft 365  that can create multimedia and presentations via Copilot. Deploying these specific AI plugins in productivity applications safely could involve years of fine-tuning and implementation; first, enterprises need to understand what the tools do, understand potential security issues, and prepare guardrails, analysts said. A technical rollout of AI systems and agents doesn’t mean enterprises are doing the right thing — they could be moving too quickly. “We’re at the very beginning of the journey toward agentic AI, but vendors like to use the term ‘agent,’ even though they lack the autonomy and actionability they will eventually possess,” Gownder said. Microsoft Copilot Agents are good for tapping into specific data sources — such as a SharePoint site — to ensure that the right data sources are being queried, Gownder said. “It’s helpful for a sort of RAG (retrieval augmented generation) model of computing. But these aren’t complex agents,” he said. Agents can be stacked so that one agent calls another to do a special task, a process that could include several agents before a result is shown to the end user, said Jack Gold, principal analyst at J. Gold Associates. As with any new technology, enterprises need to learn how to best use the technology, including asking the right questions in the right sequence so that the agent itself is productive. “It’s also critical that the questions asked are succinct, because agents have a tendency to ‘wander’ if you don’t give them specific criteria,” Gold said. The “user training” process will likely include trial-and-error, and it can be quick or not, depending on the role and complexity. “The biggest challenge will be: can you trust the results that the agents just provided you,” Gold said. As with any enterprise-class system, the CISO needs to be involved to determine any danger points, and HR and legal teams might also need to be involved in rollouts that involve corporate data. “But with AI, it’s amplified by the ability for these systems to actually enhance the types of information that can be retrieved by surmising new data from the old. It’s a different level of requirements,” Gold said.