Analysts: Go Slow On M365, Google Workspace ‘agent-ish’ Ai Rollouts

Microsoft, Google and others person been adding generative AI ‘agents” to their package astatine a accelerated clip, but companies should beryllium cautious successful adopting them.

The past fewer months person seen an detonation of new agentic-like AI tools that tin automate workflows wrong productivity applications, but analysts are recommending that enterprises cautiously measure deployment risks earlier adopting them.

“Enterprise leaders request to travel up pinch a strategy for reasoning astir agentic AI successful its existent sense, and besides governance for these solutions,” said JP Gownder, vice president and main expert connected Forrester’s Future of Work team.

Gownder said these devices aren’t truly agentic AI, but “agent-ish” tools, which is really Forrester is terming these applications. “Agent-ish solutions tin lick constituent problems, but they aren’t nan ‘digital coworkers’ that vendors are talking about, astatine slightest not yet,” Gownder said.

Google and Microsoft person been processing AI plugins for their productivity suites, and 3rd parties are providing their ain AI plugins for nan applications.

At past week’s Google Cloud Next, Google introduced caller AI devices for its Workspace suite to automate activity crossed productivity apps. Workspace Flows tin automate a watercourse of jobs progressive successful archive reviews, customer support requests, aliases merchandise analysis. The “help maine analyze” characteristic successful Google Sheets tin return accusation from tables and supply instant information study and insights.

Microsoft celebrated its 50th day earlier this period pinch a gaggle of unique Copilot automation features that tin activity pinch nan 365 productivity suite aliases independently. 

The company’s Microsoft 365 roadmap touts an exhaustive database of upcoming Copilot automation features, including automation devices to constitute Word documents, learning companions based connected learning needs, and automated HR tools.

This week, Anthropic introduced the integration of its Claude AI exemplary pinch Gmail and Calendar, successful summation to Google Docs, which helps “gain deeper penetration into your activity context” by searching emails, reviewing documents, and scanning nan calendar, nan institution said via email.

“Whether you’re managing projects, readying trips, aliases handling individual and family tasks, Claude now understands your afloat discourse to supply amended assistance. Navigate to settings to link to nan Google Workspace beta,” Anthropic said.

And Adobe is processing an Adobe Express AI supplier for Microsoft 365  that tin create multimedia and presentations via Copilot.

Deploying these circumstantial AI plugins successful productivity applications safely could impact years of fine-tuning and implementation; first, enterprises request to understand what nan devices do, understand imaginable information issues, and hole guardrails, analysts said.

A method rollout of AI systems and agents doesn’t mean enterprises are doing nan correct point — they could beryllium moving excessively quickly.

“We’re astatine nan very opening of nan travel toward agentic AI, but vendors for illustration to usage nan word ‘agent,’ moreover though they deficiency nan autonomy and actionability they will yet possess,” Gownder said.

Microsoft Copilot Agents are bully for tapping into circumstantial information sources — specified arsenic a SharePoint tract — to guarantee that nan correct information sources are being queried, Gownder said. “It’s adjuvant for a benignant of RAG (retrieval augmented generation) exemplary of computing. But these aren’t analyzable agents,” he said.

Agents tin beryllium stacked truthful that 1 supplier calls different to do a typical task, a process that could see respective agents earlier a consequence is shown to nan extremity user, said Jack Gold, main expert astatine J. Gold Associates.

As pinch immoderate caller technology, enterprises request to study really to champion usage nan technology, including asking nan correct questions successful nan correct series truthful that nan supplier itself is productive. “It’s besides captious that nan questions asked are succinct, because agents person a inclination to ‘wander’ if you don’t springiness them circumstantial criteria,” Gold said.

The “user training” process will apt see trial-and-error, and it tin beryllium speedy aliases not, depending connected nan domiciled and complexity. “The biggest situation will be: tin you spot nan results that nan agents conscionable provided you,” Gold said.

As pinch immoderate enterprise-class system, nan CISO needs to beryllium progressive to find immoderate threat points, and HR and ineligible teams mightiness besides request to beryllium progressive successful rollouts that impact firm data.