Blog
What a business actually means by 'AI agent'
The word covers three different things that cost and fail differently. Here's how to tell which one you need before anyone builds anything.
TL;DR — “AI agent” gets used for three different things: a scripted bot, an LLM with tools, and an autonomous multi-step system. Most service businesses that come to us need the middle one. Agreeing which one you’re buying — in a single sentence — is the cheapest scope decision you’ll make on the whole project.
The three things people call an agent
1. A scripted bot
Follows a decision tree. Cheap, predictable, brittle. Good for a narrow FAQ or a booking flow where every path is known ahead of time — “what are your hours”, “reschedule my appointment”, “which plan includes X”. Not an “AI agent” in any interesting sense, and it breaks the moment someone phrases things in a way the tree didn’t anticipate. Often mislabelled as AI. Often all you actually need.
2. An LLM with tools
A language model that can call your systems — look up an order, check a calendar, write to the CRM, send a templated email — and decides when to call them based on the actual conversation. Someone writes “hey did my thing ship yet, the Tuesday order” and it knows to look up the order, read the status, and answer in plain language.
This is the sweet spot for most service businesses. It handles the messy phrasing of real requests, but it only acts inside rules you set, and it hands anything outside those rules to a person. Our AI Caller demo is this category: it works a qualification script, books or declines, and escalates anything off-script.
3. An autonomous multi-step system
Plans, acts, checks its own work, and loops — often several models coordinating, each with a job. Powerful, meaningfully harder to make reliable, and rarely the right first project. It earns its cost when the task genuinely has many branching steps whose path changes based on what earlier steps turned up — a research task, a multi-stage migration, a workflow with real “it depends” in the middle.
How to tell which one you need
Two questions settle it most of the time:
- Does the task follow a fixed path, or does it branch on what it finds?
- If it gets something wrong, is that annoying — or expensive?
- Fixed path, low cost of error → scripted bot. Don’t overbuild it.
- Branches on the input, needs a bit of judgement, a human reviews the calls that matter → LLM with tools. This is where we spend most of our time.
- Long chains of dependent decisions, and you’re prepared to invest in making it reliable → autonomous system, scoped and budgeted as its own thing.
Why the label matters before you build
The three cost very differently to build, cost very differently to run, and fail in different ways — a scripted bot fails loudly and predictably, an autonomous system fails in ways that take a while to notice. If the person paying and the person building don’t agree on which one is being made, you find out three weeks in, and the fix is a re-scope.
So we ask it on the first call, in one sentence: which of these three are we building, and why that one? If you can’t answer it yet, that’s fine — it’s usually the first thing an AI audit makes clear.
Got a workflow and not sure which category it falls into? Book a diagnostic call.