What an AI agent is, and what it is not

An AI agent is software that uses a language model to decide its own next step toward a goal, using tools you gave it: read this inbox, draft in this CRM, search this folder. The deciding is the point. A chatbot answers when spoken to. An automation follows a fixed script. An agent looks at the situation, picks an action, acts, looks again.

The three words vendors blur together

• Chatbot: converses. Waits for a human, responds, forgets. Useful, but it does no work on its own.
• Automation: executes. When X happens, do Y, every time, no judgment. Reliable exactly because it cannot improvise.
• Agent: pursues. Give it a goal and tools; it chooses the steps. Powerful exactly because it can improvise, which is also the risk.

Most of what gets sold as an agent is a chatbot with a nicer dashboard. The test is simple: if you stop typing, does it keep working? If not, it is not an agent.

What agents are genuinely good at today

• Lead intake: reading an inquiry, drafting the reply, logging it where your team already looks, within minutes instead of a business day.
• Research: gathering and summarizing across sources that would eat an afternoon.
• Drafting: follow-ups, reports, content first passes, at a quality worth editing rather than discarding.
• Reporting: pulling the same numbers every Monday without being asked twice.

Where the human stays in the loop

Anything that spends money, signs, publishes, or promises. A well-built agent system routes those moments to a person by design: the agent prepares the action, a human approves it. These approval gates are not a limitation, they are the difference between a system a business can trust and an incident waiting for a timestamp. If a vendor's demo never shows an approval step, ask why.

The same logic applies to cost. An agent that can call models without limits can also run up a bill without limits. Caps, budgets, and logging go in on day one, not after the first surprise invoice.

How to start without betting the business

Pick one workflow with real volume and low stakes, usually intake or reporting. Wire the agent into the tools you already run rather than migrating anything. Document it so your team owns it. Expand only after the first workflow has earned trust. At Monolith, automation setups run $2.5k-$7.5k, ship documented, and always include the approval gates; the point is a system your team runs, not a dependency on us.