What AI agents can and cannot do for a business in 2026
AI agents are useful for scoped, repeatable work across several tools. They are weak where processes are undocumented, data is siloed, or the decision needs real judgment.
Published 2026-10-15 · 6 min read
Where agents are genuinely useful
Agents work best when the input and the output are clearly defined. Pulling a lead from a form, looking up the company, scoring it against a simple profile, and drafting a reply for a person to send is a good example. Each step is bounded, and each can be checked.
Scheduled checks are another good fit: an agent that reads yesterday's enquiries and flags the ones nobody has answered is cheap to run and easy to trust.
Where they still fall down
A Celonis survey of 1,649 leaders, as reported in 2026 coverage, found that 76 percent said their business processes were not ready for AI. An agent cannot compensate for a process nobody has written down.
Siloed data is the other common failure. When each department keeps its own version of a customer, an agent that reads one system will act on an incomplete picture.
Long-horizon judgment stays with people
Strategic decisions, pricing changes, and anything with reputational or legal weight need a person who owns the outcome. Gartner guidance cited in 2026 commentary argues for matching autonomy to risk: low-stakes work can run more freely, high-stakes work keeps a human in the loop.
Treat an agent as a capable junior with a narrow brief, not as a replacement for the person who knows the business.
How to decide what to hand over
Ask three questions. Is the task repeated? Are the steps written down? Would a mistake be cheap to catch? If all three answers are yes, it is a candidate. If any answer is no, fix that first.
