AI assistants and more advanced agent-based tools are gaining visibility in the workplace, even as most organizations remain cautious about deploying them at scale. Analysts say that might change as the technology matures, but only if businesses address persistent challenges around security, governance, and trust.
A Gallup poll in November showed that just 18% of US workers use AI tools on a weekly basis, and just 8% use AI daily, highlighting its still limited use in the workplace. A separate PwC survey of 50,000 workers globally found similar results: 14% of respondents use generative AI (genAI) daily, while 6% interact with AI agents each day.
Even so, analysts envision some organizations moving beyond pilot projects in the near future. When it comes to AI in collaboration software applications, Irwin Lazar, principal analyst at Metrigy, sees signs that businesses intend to move more aggressively from experimentation to broader adoption this year.
Lazar said companies increasingly fear falling behind if they fail to adopt the technology, particularly given its potential to streamline collaboration and save time. “I expect you’ll see a large movement into real-world adoption, whereas last year it was more about pilots and trying to figure out how do we deploy successfully?”
“Adoption is picking up,” said Ethan Ray, senior analyst at 451 Research, part of S&P Global Market Intelligence. The research firm found that more than half of enterprises already have agents in production or testing, and organizational integration of genAI use is expected to jump from 27% to 40% within the next 12 months. That, he said, assumes businesses can overcome nagging deployment challenges.
“Progress will depend on building trust as leaders need strong governance, observability, and security controls, because top concerns are data privacy, accuracy, and reliability,” said Ray.
AI assistants still struggle to scale in the workplace
Even with a wide range of AI tools available to workers, deployments have been limited so far. Take Microsoft 365 (M365) Copilot, for example: two years after its full launch, businesses remain slow to adopt the AI assistant.
“Despite the hype, Microsoft has really struggled to make huge headway in terms of deploying it at scale,” Max Goss, senior director analyst at Gartner, said at the Gartner IT Symposium/Expo in Barcelona in November.
An audience poll during Goss’s presentation showed that most either remain in pilot deployments or have rolled out to a small group of less than 20% of employees. Few have deployed M365 Copilot widely across their workforce, mirroring the broader pattern of business adoption Gartner has noted, said Goss.
Several factors have slowed wider adoption, including security and governance worries, and the need to train staffers to use the AI assistant. An unclear ROI case has also put the brakes on any expansion plans and is having “real impact on Copilot adoption” when it comes to larger rollouts, said Goss.
Still, business interest in M365 Copilot remains high, he said, an indication that Microsoft’s marketing efforts are paying off in some ways. A Gartner survey showed that IT leaders’ priorities for AI assistants over the next 12 months largely center around M365 Copilot, both for the paid version (86%) and the free Copilot Chat (68%).
Businesses are interested in other AI assistants too: 56% of IT leaders plan to roll out OpenAI’s ChatGPT to staff, according to Gartner data, with Google’s Gemini, Anthropic’s Claude, and Amazon’s Q also piquing interest.
In fact, most organizations are looking at multiple AI assistants. Only 8% are focused on a single tool, with an average of at least three enterprise AI assistants in use at surveyed organizations. “The AI race is still very much on, and Microsoft has genuine competition,” said Goss.
AI tools start to mature
While customers are cautious, software vendors continue to add AI features into their products. Almost every vendor in the collaboration software market has an agent offering at this point, said Lazar.
“They’re going from standalone agents [where] you have to build capabilities, to agents that are already available within the applications,” said Lazar. This includes off-the-shelf agents that users can select for tasks such as project management, sales management, or IT service desk support. Customers only need to grant access to relevant data and set governance rules before putting the agent into use.
“Now you’re really starting to see this agentic era start to move forward, at least from the vendor standpoint,” he said.
“In 2026, vendors will move past just adding assistants and start building features that make agents reliable, explainable, and easy to govern,” said Ray. He expects more focus on “things like memory (so agents remember context), transparency in decision-making, and guardrails for safety.”
Agents get connected
One development that could enhance the usefulness of agents is the ability for AI assistants to interact with each other.
Employees can get frustrated with AI tools that are confined to a single app, which is at odds with the way employees work. “Work doesn’t live in one software tool,” said Will McKeon-White, senior analyst at Forrester. “I suspect most platforms have now realized the need for multi-vendor, multi-agent orchestration.”
To address this challenge, tech companies have been finding ways to simplify communication between agents — most notably turning to Anthropic’s model context protocol (MCP) and Google’s Agent2Agent (A2A) protocol.
MCP servers have been built into a wide variety of collaboration and productivity tools already. “Vendors have realized they can’t own everything, and so they’re building MCP servers to essentially federate the data they have with other AIs,” said Lazar.
The use of MCP servers could change how employees interact with collaboration and productivity tools, he said, by allowing businesses to choose a primary AI model and pull in data from multiple sources. “It saves the user from having to move back and forth between applications in order to do things like summarize chats or get a pulse on what’s happening in the company,” said Lazar.
Security and governance
Alongside the potential benefits, the use of MCP also introduces new security risks.
“The big concern I hear when I talk to folks is related to security of MCP servers,” said Lazar. “They will be the number one target for attack as they become more widely available, because that’s the gateway to enterprise data.”
For attackers, MCP servers present a “target-rich environment,” whether for data exfiltration or data poisoning. “If there’s any limitation on deployment, that’s going to be what people are concerned about,” he said.
In his presentation, Goss said security and governance will continue to be key considerations for IT decision-makers rolling out M365 Copilot, though the challenges continue to evolve.
Oversharing – where M365 Copilot surfaces sensitive corporate data to users not authorized to have the information — remains a priority, for instance. Other risks have emerged, with “agent sprawl” becoming a notable topic in 2025 as businesses deploy agents and workers build their own.
In 2026, he said “multimodel agent sprawl” could be an emerging issue for M365, as Microsoft offers the option to connect its AI assistant to a wider range of models, notably Anthropic’s, as it moves beyond OpenAI as its main partner.
“When Microsoft integrated Anthropic, they took the decision not to host an Anthropic model: that model is still in the AWS environment,” he said. “As Microsoft onboards more models, it’s going to be very difficult for them to host them all and do what they’ve done with with OpenAI. So, we’re now going to have to start to think about: how do we manage agents and models that are outside of the Microsoft trust boundary, as well as the ones that are within? What do you do about that? What strategy should you have?”
He recommended that organizations use “adaptive governance” to manage agents, setting the level of governance controls in relation to the level of risk. This approach enables the creation of a “self-governed, safe zone where users can create low-risk agents using Copilot Studio or other tools that will help them improve their productivity without exposing you to risk,” he said.
Goss said governance concerns shouldn’t be a reason to avoid deploying AI assistants or agents. “For me, governance is the ultimate enabler of AI — but we’ve got to get it right, and we’ve got to spend some time on it,” he said. “While the value around Copilot is still a little bit mixed, it’s a perfect time to think about how we get the foundations in place. Because I think there will come a tipping point…where most people are deploying Copilot at scale.
“…We’re not there yet, so it’s a great opportunity to fix the foundations.”
