While the industry continues to grapple with defining AI agents, we researched where AI agents are today, how they’ll evolve over the next five years, and what technical and nontechnical limitations are slowing AI agents’ ability to reach the promised land of autonomy.

The key takeaway? AI agents are eyeing autonomy, but your poor documentation means they may not reach this threshold.

The Race For AI Agents Is A Marathon, Not A Sprint

AI agents are apparently everywhere, if the market is to be believed. In usual analyst fashion, it’s our job to break down the complicated reality of “well, no, but also yes.”

AI agents in their glorious self-driving, self-deciding, business-transformative manifestation aren’t in production yet. While they haven’t reached this apex of evolution, it doesn’t mean that they aren’t already delivering value, even in their current nascent stage. That’s why clients should be investigating AI agents now and exploring how they could benefit from AI agents while being realistic about what’s feasible today and will likely be feasible tomorrow.

While AI agents are expanding their capabilities quickly, there are several critical technical and nontechnical components that the industry and organizations still need to resolve, including:

  • Inconsistent or inaccurate permissions. Our colleague Heidi Shey summarizes AI agents as such: “They’re like toddlers: They will pull off the shelf whatever they can access.” Like toddlers, the AI agent won’t consider factors like what it’s grabbing, if it’s appropriate for what the AI agent is trying to accomplish, or if it’s sensitive (fragile). Toddler see; toddler grab; toddler (probably) throw. Chaos ensues!
  • Poor knowledge quality and process documentation. AI agents need step-by-step instructions on how to execute tasks. For most businesses today, this know-how lives in fragmented workflows, undocumented data, and unofficial processes. Ask yourself, “Do I know exactly where to find formal documentation on how a certain task is done?” and “Does the documentation reflect how tasks are actually done?” for a gut check on your organization’s readiness.
  • Limited agent orchestration within disparate ecosystems. As AI agents proliferate, they need to work together to accomplish complex tasks while hiding this complexity from end users. For example, users shouldn’t need to know which IT-department-owned AI agent manages PC troubleshooting. While bridging agents across ecosystems is technically possible today, it requires using intermediary orchestration engines, which are themselves configured through constant manual tuning. Protocols to standardize agent-to-agent communication are developing but nascent.

Prepare For AI Agents By Going Back To The Basics

Ironically, the right approach to AI agents is to tune out the hype and start small. Define your use case first: What exactly do you want the AI agent to do? We cannot stress this enough! AI agents are worth experimenting with, but the most valuable task you can do right now is map out your data and tech requirements, identify collaborators, and get involved with your company’s AI governance team.

Expect meetings to document and agree on official process flows and definitions.* Involve the AI agent deployment team and (if possible) representatives from the technology provider, too.

Read the report, AI Agents: Ready For Enterprises, And Moving Toward Autonomy, and set up a guidance session with Craig (on automation and AI agents’ evolution), Will (on conversational AI), or Steph (on marketing use cases and privacy concerns) for a deeper dive.

*Yes, meetings. We’re serious, and we’re sorry.

Thanks to our coauthors, Craig Le Clair and Christina McAllister!