Six words that explain all of AI
Six words cover everything you will hear about AI for the next two years: model, API, tool, agent, skill, harness. Learn them in order and the sales pitches, the headlines, and the product demos all sort themselves into a drawing you can hold in your head.
I spend my days building AI systems for small businesses in Savannah. The owners I work with are not behind. The Federal Reserve's 2025 Small Business Credit Survey found that 46% of small employer firms already use AI in some way, another 15% plan to start within a year, and one third have no plans at all (Federal Reserve Banks, 2026 Report on Employer Firms). What the owners lack is a shared vocabulary. Half of them cannot tell me whether the thing a vendor pitched them last week is an agent or a chatbot with a new label. This post fixes that.
1. Model
A model is the engine. You type something, the prompt. The prompt goes to the model, which is the system under ChatGPT, Claude, Gemini, or Copilot. The model writes back. That exchange is a chatbot.
Nothing happened in the world. The model did not send an email, book an appointment, or move a file. It answered a question. Most people who say they use AI mean this, and only this. The Fed survey found that among firms using AI, about half describe themselves as experimenting, 44% have partially integrated it into a business process, and 7% have fully integrated it (Federal Reserve Banks). The chat window is where the experimenting happens.
Two facts about the model carry over to everything else. It is good with language and reasoning, which is why it can draft a quote or summarize a contract. And it will sometimes answer with full confidence and be wrong. The industry word for that is hallucination. Treat it as a known property of the machine, the way you treat a new hire's first week. Assume it will happen and check the work. The same Fed survey found accuracy was the top challenge named by AI users, at 46%.
2. API
An API is the pipe between your business and the vendor's computers. When you use a cloud AI tool, the model is not on your laptop. Your request travels to the vendor's servers, the model runs there, and the answer travels back. The pipe it travels through is the API.
The pipe matters for two reasons that come up again and again. Your data travels through it, which is a question of what you are willing to send out of the building. And you pay by what travels through it, which is how most AI beyond the chat window gets billed. Hold the picture: your business on one side, their servers on the other, a pipe between.
The alternative to the pipe is running the model on hardware you own, so nothing leaves the building. That is a separate decision with its own trade-offs, and I cover it in Where should your AI run?. For now, know that the pipe is a choice, not a given.
3. Tool
A tool is a connection from the model to something real: your calendar, your email, a spreadsheet, a database, the web. With tools, the model can look things up before it answers and take an action after. "Check my calendar and find a time next week" only works if the model can see the calendar. "Draft a reply and save it" only works if it can reach the inbox.
This is the step most businesses have not taken, and it is where AI stops being a smart search box and starts doing work. The Fed survey found that for firms planning to adopt AI, the top challenge was finding tools that fit the business, at 54%. Fit is a tools question. A model with no tools can only talk about your business. A model with the right tools can act in it.
4. Agent
A chatbot takes a prompt. An agent takes a goal. You say "put together Monday's report," and the agent runs a loop: plan the steps, act using its tools, check its own work, and if the job is not done, go around again.
The most important part of the loop is the gate. At the points you choose, the agent stops and asks a person before doing anything that matters. Send the email? Issue the refund? Change the price? A well-built agent pauses there. A badly built one does not, and that is where the horror stories come from.
The word agent is being attached to a lot of software that does not earn it. Gartner, in a June 2025 press release, estimated that only about 130 of the thousands of vendors claiming agentic capability actually have it, and coined the term "agent washing" for the rest (Gartner, June 25, 2025). The same release projected that over 40% of agentic AI projects will be canceled by the end of 2027 for escalating costs, unclear value, or inadequate risk controls. That is a projection, not a result, but the direction matches what I see.
The test is simple. If a salesperson says agent and cannot show you the loop, the tools, and the gate, you are looking at a chatbot with a new label.
5. Skill
A skill is the written playbook for one job. The Monday report. The first draft of a quote. How we answer a complaint. It is the steps, in plain English, that the agent follows when that job comes up.
The value of a skill is that it is portable. You can hand the same skill to a different agent, or to a new employee, and the job gets done the same way. Most small businesses already have this knowledge. It lives in one person's head, or in a half-finished Google Doc from 2019. Writing it down is the work. Once it is written, the agent is the easy part.
6. Harness
The harness is everything you would hand a new employee on their first day, wrapped around the model. It has five parts.
- Instructions: who we are, how we talk to customers, what never to do.
- Skills: the playbooks for each recurring job.
- Tools: which systems the agent can reach.
- Memory: what to remember from last time.
- Permissions: what it can do alone and what needs a sign-off.
The model is the talent. The harness is the job. Same model, different harness, different employee. When people in my field say they are building an agent, most of the time they mean they are writing the harness, and the harness is mostly plain English. If you can write a job description for a new hire, you can write a harness.
The drawing
Put the six words together and you have one picture. A model sits in the middle. A pipe (the API) connects it to your business if it runs in the cloud. Tools branch out from it to your systems. A loop around it, with a gate, turns it into an agent. Skills attach to the loop, one card per job. A frame around the whole thing, the harness, holds the instructions, memory, and permissions.
Every AI product you will be sold in the next two years is some subset of that drawing. A chatbot is the model alone. A copilot inside your accounting software is a model with a few tools and no loop. A true agent has the loop and the gate. When you can place a product on the drawing, you can price it, judge it, and decide whether it fits.
What to do this week
You do not have to build any of this to use it. The major vendors now sell personal AI assistants with a harness already written: an assistant that lives on your phone, reads what you give it, and does small jobs for you. They cost somewhere between a streaming subscription and a car payment per month, and they are the easiest place to start.
Give one a task you do every week and see how far it gets. That is the first experiment. The second is to write down, in plain English, the steps for one job you repeat. You have just written your first skill.
If you want to work through this on your own business, book a free fit call. Thirty minutes, one process from your week, and you leave with a plan for it. If you would rather read first, the newsletter sends the six-words card, the three-tier data card, and the decomposition worksheet to new subscribers, and one short note a week after that.
Sources: Federal Reserve Banks, 2026 Report on Employer Firms: Findings from the 2025 Small Business Credit Survey (fielded Sept to Nov 2025, 6,525 firms with 1 to 499 employees); Gartner press release, June 25, 2025. Figures checked October 7, 2026.