Agent automation 101 — agents that do real work

An AI agent is not a chatbot with opinions. It is a software worker that follows a defined goal — find leads, draft outreach, answer support, classify documents — using tools and data you give it, inside guardrails you set. Done well, it is the most boring, valuable hire you will make this year.

What an agent actually is

Strip away the marketing and an agent is three things glued together:

1. A goal and instructions. “Find 20 local businesses with weak websites, write a personalized first message to each, and queue them for my approval.”
2. Tools. A browser, a search API, your CRM, your inbox, a document store.
3. Guardrails. What it may never do, what it must escalate, and where a human signs off.

The magic is not the model. It is the discipline of the instructions and the quality of the tools.

Keep the human where the judgment is

We are suspicious of “fully autonomous” for anything that touches a customer or a regulated process. Our rule of thumb:

• Let the agent do the volume. Discovery, drafting, classification, summarization, routing.
• Let the human do the verdict. Approving an outreach message, confirming a document’s category, sending anything with your name on it.

This is human-in-the-loop, and it is how you move fast without doing something you regret. The agent compresses hours into minutes; the human keeps the accountability.

Start with one painful, measurable job

Do not “roll out AI.” Pick a job that is painful, repetitive, and measurable, and automate that one thing end to end. Good first projects:

• A lead-discovery agent that builds a clean list every morning for your review.
• A support agent that drafts answers from your docs and hands the hard ones to a human.
• A document agent that captures, classifies, and files inbound paperwork.

Measure the before and after. If the agent saves real time and produces work you are comfortable shipping, expand. If not, tighten the instructions — that is almost always the fix.

The governance you need from day one

Agents that touch data need the same hygiene as any other system: an audit log of what they did, clear access scopes, a way to pause them, and a retention policy for what they produce. We wire this in first, not last — because an agent without a paper trail is a liability with a personality.

Start small. Keep the human. Write it down. That is the whole recipe.