40%
of enterprise applications were projected to include task-specific agents by the end of 2026, up from under 5% in 2025.
GartnerFree AI agents white paper
AI Agents at Work
Decide when an AI agent is worth trying, what it should be allowed to do, and how people will check its work. Includes a plan for a 60-day trial.
- 11-page guide
- Includes research and links to sources
- Download the PDF after filling out the form
Inside the guide
What you’ll learn
- The differences between fixed automation, AI assistance, and AI agents
- Questions to decide whether a task needs an agent
- Limits to set on agent access, tools, actions, and resource use
- System design, testing for misuse, a four-stage rollout, and a 60-day trial plan
From the research
Research cited in the guide
Here are a few findings discussed in the guide. Follow the source links for the full studies and their context.
Single digits
describe agent deployment across nearly all business functions in the 2025 survey data reported by Stanford.
Stanford AI Index — Economy~1 in 3
OSWorld computer-use tasks failed for the leading agents assessed in Stanford's 2026 report.
Stanford 2026 AI IndexCommon questions
Before you download
What makes an AI system an agent?
An agent uses an AI model to choose steps toward a goal, use permitted tools, check the results, and decide what to do next. A chatbot or a fixed process that makes one AI call doesn’t necessarily do that.
When should an organization use an agent instead of automation?
Consider an agent when the task has a clear goal but varies too much for a fixed set of rules. You also need to be able to limit its tools and permissions, check the results, and review or undo its early actions.
How much autonomy should the first agent receive?
Very little. Start by having it observe work or prepare drafts for review. Let it take selected actions only after testing shows it can respect access limits, produce reliable results, handle errors, and stay within your rules and budget.
Want help applying the guide?
Tell us about the task or system you’re working on. We can help you assess the options, plan a trial, or get the project started.