Enterprise AI agents: beyond the demo

“Agentic AI” is the phrase of the moment, and most of what's sold under it won't survive contact with a real enterprise. That doesn't mean agents are hype — it means the gap between the demo and the dependable system is even wider here than usual. Here's how I think about where they fit.

What an agent actually is

Strip away the marketing and an agent is a language model given tools and a loop: it can decide to take an action — call an API, query a system, run a step — observe the result, and decide what to do next, rather than producing a single answer. That autonomy is the source of both the power and the risk.

Where agents earn their place

Agents shine on multi-step tasks with genuine branching, where the path can't be fully scripted in advance: triaging an issue across several systems, gathering and reconciling information, orchestrating a workflow that legitimately varies each time. If a task is the same every time, you don't want an agent — you want ordinary automation, which is cheaper and more reliable.

Where they quietly fail

Autonomy compounds errors. A small mistake early in a chain becomes a confident wrong action three steps later. In production that means real consequences — a wrong record updated, a wrong message sent. The engineering that matters is the boring kind: tight tool permissions, human approval on consequential actions, observability into every step, and hard limits on what the agent can touch.

My rule of thumb

Use the simplest thing that works. Reach for a single well-prompted call before a chain, a scripted workflow before an agent, and an agent only when the task genuinely demands dynamic, multi-step reasoning. The teams getting value from agents are disciplined about scope; the ones getting burned handed an autonomous system the keys and hoped. Governance isn't a footnote here — it's the difference between an asset and an incident.