Guide — AI literacy

Four different things are being sold to you as an AI agent.

Only one of them is worth money in a 20-person business.

You have been pitched agents by four vendors this year and every pitch described something different. That is not you missing something. The word genuinely covers four separate products, and the gap between them is the gap between a $200 tool and a $20,000 build.

Book the Architecture Assessment See what we build

$250 — credited 100% toward any build. If nothing needs building, we'll tell you that.

One of four is the one most vendors mean when they quote you.

What are the four things people call an AI agent?

Written in plain English, in the order of what they cost you. Read all four before you take another vendor call — most of the confusion in this category is two people using one word for different products.

1. A chatbot with a better memory

It answers questions from a set of documents you gave it. It does not touch any other system, and it cannot do anything except talk. This is genuinely useful on a website or an internal help desk, and it is the cheapest of the four by a wide margin. If a vendor is quoting agent money for this, you are paying for the word.

2. A scripted automation wearing the word

When this happens, do that. It is a fixed path with a language model somewhere in the middle writing the text. Most of what gets sold as an AI agent to small businesses is this — and that is not an insult, because this is also what most small businesses actually need. It is predictable, cheap to run and easy to fix. It just is not an agent.

3. A copilot sitting inside a tool you already own

The assistant already built into your CRM, your accounting software, your inbox. It helps the person doing the work go faster. It does not do the work while nobody is watching. You are frequently already paying for this one and not using it — which is a real finding we hand people more often than we would like.

4. An agent that actually decides

It gets a goal, picks its own steps, calls the systems it has access to, and hands back a record of what it did. This is the real thing. It is also the one that needs monitoring, guardrails and a human approval step wherever it touches a customer or money — because software that chooses its own steps will eventually choose a wrong one, and the design question is what happens next.

When is an agent the right answer, and when is it the expensive one?

Here is the test we use, and you can run it yourself in about a minute.

Describe the job out loud as a set of steps. If you can get to the end without saying "well, it depends," you want an automation. Fixed steps are cheaper to build, cheaper to run, and they fail in ways you can predict. An agent pointed at fixed steps is a more expensive way to get the same result with more things that can go wrong.

If you say "it depends" three or four times — it depends on what the customer wrote, it depends on which supplier, it depends whether the job is commercial — then the variation is real, and fixed steps will break on it. That is where an agent earns its cost, because judgment is the thing it is actually for.

The honest version of this is that most operations contain a bit of both. The estimate that gets built the same way 300 times is automation work. The email replying to a lead who wrote three sentences about a specific job is agent work, because the reply has to be about what they wrote. Building the second when you needed the first is the most common expensive mistake in this category.

Where are the hours actually going?

Before choosing between the four, it is worth knowing what you are buying back. The research on this is consistent and it is not about AI at all — it is about the cost of work that is spread across tools that do not talk.

58% of the working day

goes to coordination rather than the skilled work people were hired for — chasing status, hunting for information, moving data between tools. Measured across 9,615 knowledge workers in six countries.

Anatomy of Work Global Index 2023, GlobalWebIndex for Asana

9.5 minutes to refocus

is the average time it takes to get back into productive concentration after switching between applications. The same research found 59 minutes a day lost just searching for information spread across tools.

Qatalog with Cornell University's Ellis Idea Lab, Killing Time at Work, 2021 — three surveys, 1,000 participants each

Why that research matters more than the AI question

Read those two numbers together and the picture is not a business that needs smarter software. It is a business where the connective tissue between systems is being carried by people, a few minutes at a time, all day.

That is good news, because it is the cheapest category of problem to fix. It does not need an agent that reasons. It needs the gap closed so nobody is carrying data across it by hand. We have watched a business get roughly 45 hours a month back by connecting two tools it already paid for — no agent involved anywhere in the build.

So the useful order of questions is not "which AI agent should I buy." It is "where is my operation leaking, and what is the least clever thing that would stop it." Sometimes the answer to the second question is an agent. Frequently it is a setting you already own.

What should you ask a vendor before you sign anything?

Five questions. You can ask all of them in a first call, and the answers tell you more than any demo will.

Which of the four is this?
Ask them to place their product in the list above. A straight answer takes ten seconds. A long answer is itself the answer.
What does it touch?
Name every system it reads from and writes to. The number of integrations drives the cost and the fragility far more than the model does.
What happens when it is wrong?
Not whether — when. Which actions need a human approval, and where does the failure get reported? If nothing catches a bad output, you have bought a faster way to be wrong.
Who owns it if we part ways?
Does the thing live in accounts you own, or in theirs? This question is boring right up until the day it is the only question that matters.
What is the cheaper version of this?
A vendor who cannot name the cheaper alternative to their own product either does not know the category or is not going to tell you. Both are useful to find out on the first call.

Questions operators ask about agents

Do I need an AI agent or an automation?

Most small businesses need an automation. If the work follows the same steps every time, an automation does it more cheaply and more predictably. Agents earn their cost when the input varies enough that fixed steps break — a reply that has to be about what the customer actually wrote, for example.

Will an AI agent replace someone on my team?

In a business of 5 to 30 people, almost never. What it replaces is the fragment of several people's days spent re-typing, chasing and searching. Those fragments are where the recoverable hours are, and nobody in your building will miss them.

How much does this cost?

The model usage is usually the cheapest part. Cost is driven by how many systems it has to touch, how much judgment it has to encode, and how much monitoring it needs. A single-purpose build with two integrations is a different order of cost from something that touches your whole operation — which is why we map the operation before quoting anything.

What happens when it gets something wrong?

That is a design decision, not an accident. Anything a customer sees or that moves money passes a human approval step. Anything reversible and internal can run unattended. In one build of ours, roughly 300 estimates ran through the system over five months and a person approved every single one before a customer saw it.

Everyone says we're behind. Are we?

Probably not in the way the ads imply. Every service business will run on connected systems within a few years. The only question is whether you lead that shift or catch up to it later — and leading it starts with knowing which of your work is repetitive, not with buying an agent.

Where should I actually start?

With measurement, not purchase. Pick the job that annoys your team most, write down every step and every place it gets typed twice. If you would rather have someone do that with you, that is exactly what the Architecture Assessment is.