The trained system that reads input and generates a response; the engine of an AI tool.

Human Out of the Loop · Part 3 of 6

November 6, 2026

Getting in the saddle.

Model, Harness, Context, Prompt: The Parts of an AI Tool, Explained

Q: What is an AI harness?

Everything built around a model to make it useful and safe: the app, instructions, connections, permissions and output checks.

Q: What is a context window?

The amount of information an AI model can consider at one time.

Q: What is a prompt?

The instruction you give an AI tool.

Q: Is Microsoft Copilot a model?

Copilot is best thought of as a harness built into Microsoft 365 that runs on underlying AI models.

WRITTEN BY

Quinn Merio

Solutions Architect

Artificial Intelligence, Digital Transformation, Software Development

QUICK SUMMARY

The model is the engine. The harness is the car built around it. Context is the map, and the prompt is the steering wheel. Here’s how the parts of an AI tool fit together, and the questions to ask before you buy one.

Model, harness, context window, prompt, Copilot, ChatGPT, agent. People in tech use these words as if everyone knows what they mean, and often as if they mean the same thing. They don’t.

The good news is that you only need four of them to make sense of almost any AI tool. And they fit together like the parts of a car.

Every AI tool has the same four parts

The engine does the work. Everything else decides whether that work is useful and safe.

The car analogy we use across this series.

What is an AI model? The engine

What is an AI model?

The model is the trained “brain” that does the actual work: reading what you give it and predicting a useful response, word by word. GPT, Claude, Gemini and Llama are all families of models.

The model is the engine. It’s where the power comes from, and it matters. A bigger, newer engine handles harder jobs: longer documents, trickier reasoning, fewer mistakes. It usually costs more to run and can be slower.

But you rarely need the biggest engine available. Sorting emails or summarizing a meeting is a short drive to the store. Reviewing a 200-page contract is a long haul with a heavy load. Matching the engine to the trip saves money without losing quality.

Engines also get swapped out. New models arrive every few months, and good tools let you upgrade the engine without rebuilding the car.

What is an AI harness? The car

What is a harness?

The harness is everything built around the model to make it useful and keep it in check: the app you type into, the instructions it follows behind the scenes, the systems it can connect to, what it’s allowed to do and the safety checks on what comes out.

In part 2 we compared AI to a horse: strong, mostly obedient, with a mind of its own. The harness is what lets you steer it. In our car analogy, it’s the car itself: the body, seats, dashboard, brakes and doors.

Here’s the part most people miss. Two tools can run the very same engine and feel completely different. A general-purpose chatbot and an AI assistant built into your company’s systems might share a model under the hood. One knows nothing about your business. The other can find last month’s inspection reports, follow your approval rules and stay out of files you shouldn’t see. The difference is the harness.

Most of what people experience as “good AI” or “bad AI” at work comes down to the car, not the engine.

52%

of Canadian workers use AI for work in some form.

Punchcard / Angus Reid, 2026

32%

use standalone, general-purpose AI assistants or websites.

Punchcard / Angus Reid, 2026

use AI tools customized or configured for their organization.

Punchcard / Angus Reid, 2026

In other words, most people at work are driving a rental: a capable engine in a car that wasn’t built for their job.

Renting made sense when AI couldn’t do much. As it takes on real work, a car built around how your business actually runs matters more. We made that case in Surviving the ‘SaaS-Pocalypse’.

What is context? The map and the trunk

What is context?

Context is everything the AI can see while it works on your request: your question, the files you attach, the earlier conversation and any company information it’s connected to. The context window is how much it can hold at once, like the size of the trunk.

An engine with no map drives somewhere, just not necessarily where you need to go. Ask an AI to “draft a quote for a 40-hour equipment rental” and you’ll get something generic. Give it your price sheet, your standard terms and last month’s quote for the same customer, and you’ll get something you can actually use.

Context is also where risk rides along. Whatever you put in the trunk travels with the request. If a tool isn’t approved for customer or employee information, that information shouldn’t go in. We’ll cover that in part 5, on guardrails.

What is a prompt? The steering wheel

What is a prompt?

A prompt is the instruction you give the AI: what you want, for whom, and in what form. It’s how you steer.

You don’t need special tricks. Clear instructions work for the same reasons they work with a new employee: say what the job is, who it’s for, what good looks like, and give an example if you have one.

VagueWrite something about the new safety procedure.

ClearWrite a 100-word note to our site crews explaining the new lockout procedure attached. Plain language, no jargon, and end with who to call with questions.

Same engine, same car. The clear prompt tells it where to go.

Putting the parts together

Next time someone pitches you an AI tool, you can ask about each part of the car.

Part	In the car	Question to ask
Model	The engine	Which model does it use, and can it be upgraded or swapped later?
Harness	The car	Was it built for work like ours? What can it connect to, and what is it allowed to do?
Context	The map and trunk	What information can it see, and where does that information go?
Prompt	The steering wheel	Who writes the instructions, and can we save the ones that work?
You	The driver	Where does a person review or approve the work?

What to do Monday

List the AI tools your team already uses. Include the free ones people signed up for themselves. Most are general-purpose rentals.
Pick one job that keeps coming up. A weekly report, a type of quote, a recurring email. Write down the context it needs and a prompt that works.
Ask the five questions above about the main tool you’re using or considering. Gaps in the answers are where problems start.
Don’t chase the newest engine. If results are poor, check the map and the steering first: better context and clearer prompts fix most problems.

The five-second version

The model is the engine. The harness is the car. Context is the map, the prompt is the steering wheel, and you’re the driver. Most of the difference between frustrating AI and useful AI is the car, not the engine.

The series

Human Out of the Loop: AI explained for people who don’t work in tech

Frequently asked questions

What is an AI model?

An AI model is the trained system that does the work of reading input and generating a response. GPT, Claude, Gemini and Llama are examples of model families. Think of it as the engine of an AI tool.

What is an AI harness?

A harness is everything built around a model to make it useful and safe: the app, behind-the-scenes instructions, connections to other systems, permissions and output checks. The same model can feel very different in different harnesses.

What is a context window?

The context window is the amount of information an AI model can consider at one time, including your question, attached files and earlier messages. A bigger window means it can work with more material at once.

What is a prompt?

A prompt is the instruction you give an AI tool. Clear prompts say what you want, who it’s for, what good looks like and, ideally, include an example.

Is Microsoft Copilot a model?

Copilot is best thought of as a harness: an AI assistant built into Microsoft 365 that runs on underlying AI models and connects them to your organization’s files, email and permissions.

About the data. Survey conducted by Angus Reid Group for Punchcard, September 2–11, 2026, among 1,100 working Canadians. AI-use questions were asked of all workers; categories can overlap. Results are self-reported. Full results are in The State of AI at Work in Canada.

Prompts and quote examples are illustrative.