Agents are arriving inside the software you already pay for

What happened

In August 2025, Gartner predicted that about 40% of enterprise applications will include task-specific AI agents by the end of 2026. The figure at the time was below 5%.

It laid out five stages. First came assistants embedded in apps. Next, in 2026, come task-specific agents that can handle a job from start to finish. By 2027 it expects agents with different skills working together inside one application. By 2028, networks of agents working across applications, with roughly a third of user experiences led by an agent instead of by clicking through screens. By 2029, at least half of knowledge workers building or governing agents on demand.

It also warned about "agentwashing", where an assistant gets relabelled as an agent. Its advice to software leaders was blunt: settle your agentic strategy within three to six months, or risk falling behind.

So what for your business

There are two sides to this.

If you buy software, agents will appear in your CRM, your ticketing tool and your finance system as a feature update. Somebody has to decide what they are allowed to touch. The risk is not the technology. It's an agent with access to customer records or payment actions turning up in a release note that nobody in security ever read.

If you build software, your customers will soon expect the same from you. Gartner's stages are a handy roadmap. Most products sit at stage one. The jump to stage two, an agent that finishes a task instead of suggesting one, is where both the value and the risk go up.

What to do this month

1. List the tools shipping agents. Read the release notes for every major system you use. Mark which ones can write data or take actions, not just read.
2. Set a default. New agent features stay off until someone with authority approves the permissions they need.
3. Centralize the log. You want one place to see what agents did, whichever vendor ran them.
4. If you build software, finish one task end to end. One task completed reliably beats ten suggestions.
5. Test for agentwashing. Ask a vendor to show a task completing with no person in the loop.