Field Notes
July 20, 20266 min read

What an AI Agent Actually Is

Justin Henriksen
Justin Henriksen

Founder & CEO, GetLatest AI

A plumber I know put a chatbot on his website last year.

It answers questions. Hours, service area, rough pricing. Customers love it.

He still checks his email at 6am to see if any leads came in overnight.

That's the whole problem with chatbots. They wait. Someone has to show up and ask, or nothing happens. Nobody is watching your pipeline at 2am. Nobody drafts a response to the lead that came in while you were under a sink at a job site.

An agent does that.

Not because it's smarter. Because it's built differently - it watches, then acts. A chatbot sits there like a suggestion box on the wall. An agent is the employee who checks the box, figures out what needs doing, and either handles it or drops a note on your desk that says "here's what I'd recommend."

That's the whole distinction. Everything else is detail.


What an agent actually does

Three things. In order.

Observe. It reads your actual systems - CRM, inbox, calendar. Not invented data. What's there right now.

Reason. A language model looks at what it found and makes a judgment. "This lead matches your five best accounts." "This deal hasn't moved in 21 days." "Your rep is on vacation and this follow-up is overdue." It can handle the messy, variable input that breaks any fixed rule.

Act. This is where most people get the design wrong.

A good agent doesn't fire off emails. It drafts. Proposes. Waits for you to confirm before anything consequential goes out. Or it executes inside a tight lane - approved templates, specific deal stages, clear rules about when to stop and ask.

The act step is all design work. That's why the demo always looks better than production.


Chatbots and automations are not agents

Vendors blur these on purpose. Here's what's actually different.

Chatbot: responds when you talk to it. Doesn't watch your pipeline. Doesn't act while you sleep. Useful for answering questions at your front door. Zero leverage at 2am when a lead just submitted your form.

Automation: when X happens, do Y. Perfectly reliable until something falls outside the script - then it fires wrong or fails silently. Great for simple, predictable handoffs. Cannot handle the variation real business generates.

Agent: triggered by events, interprets variable inputs, acts across multiple systems, has clear rules about when to proceed and when to surface something for you.

A chatbot is a vending machine. An automation is a train on a track. An agent is a new hire on their first week - smart enough to figure things out, but checking with you before they send the proposal.


Start at the bottom, earn your way up

The businesses that get hurt give agents too much authority too fast.

Think of it like handing your car keys to a teenager. On day one, you sit in the passenger seat. After six months without a dent, they drive alone to school. After a year, they take the highway.

Same logic with agents.

Read and summarize. "Here's what's in your pipeline. Here's what needs attention." Nothing changed. You decide. Low risk, immediate value.

Recommend. "This lead scored 84/100. Suggested next step: discovery call." Still no action without you.

Draft. A follow-up email, a meeting summary. You review before it goes anywhere.

Execute with approval. "I'm about to send this to Sarah at Ridgeline Roofing. Confirm?" One click.

Execute within guardrails. Agent acts without per-action approval, inside tightly defined rules. Specific templates, authorized contacts, specific deal stages. You get here after careful design and real monitoring - not day one.

For most SMBs, the biggest near-term value is the first two rungs. Synthesis you weren't getting before. Work that was falling through cracks becomes visible. That's already a different business.


Your tools stay put

The most common mistake: treating AI agents as a reason to rip out your CRM.

It's almost never right. Your existing tools hold your data, your history, your integrations. Your team knows how to use them. Replacing them means migrating everything, retraining everyone, starting your contact history over - and then discovering the new tool has different gaps than the old one.

The right model: your CRM is still the system of record. Your email platform is still delivery. The agent is the connective tissue - reads from the CRM, reasons about what it finds, drafts in your email tool, updates the deal record, flags anything outside its lane.

That's also why integrations matter more than the AI model itself. An agent that reliably connects to your actual systems and acts on what it finds is worth more than a more powerful model that can't touch your real data.


Why every demo looks good

Clean data. One workflow. No conflicting records. No messy import history. The agent follows the happy path and everyone in the room is impressed.

Then you go back to your actual business.

Same lead submitted your form twice with different email addresses. The webhook connecting your form to your CRM went down overnight. The agent drafted a response citing a pricing tier you retired four months ago because nobody updated the document it reads.

None of that is a reason not to build agents. It's a reason to build them like a production system - real data from day one, clear rules, human approval at every consequential step, error handling, someone responsible for monitoring.

The businesses that get lasting value understood that going in. The ones that got burned thought the demo was the product.


One thing to do this week

Pick one place in your growth process where someone has to open multiple systems, pull together information, and produce an output before they can take any action.

Write down exactly what they look at, what they're trying to figure out, and what they produce at the end.

That's the shape of an agent opportunity. Hold onto it.


Want me to look at yours?

Bring one workflow to a 30-minute call with me. I'll tell you exactly where an agent would fit in your business, what it would take to build, and whether it's even worth it for you. No pitch - just an honest read from someone who does this for a living.

Book a mapping call

Justin Henriksen

Justin Henriksen

Founder & CEO, GetLatest AI

Justin is the founder of GetLatest AI. 25 years building and leading technology, from Principal SWE to CEO. He writes about AI agent architecture, production systems, and what actually works.

Want your AI team built and run for you?

Book a free call30 min · no commitment