Same Engine, Different Truck

Most of the AI products you are being sold run on the same few engines.

TL;DR

• Every AI product is two things: a model, and the packaging around it. A handful of frontier labs build the best models. Vendors build the packaging: what the AI can see, what it can touch, and when it runs. A vendor can give you a better fit. The engine almost always comes from the labs.
• Judge the two separately. Check which model is inside, then judge the packaging on fit, access, approvals, and whether you can take your context with you.
• The same model climbs three rungs, Ask, Equip, and Station, as it gets more to see, more to touch, and a schedule of its own.
• Before your next AI purchase, run the five-question checklist below, or start with an AI readiness assessment.

---

Why does the difference between model and packaging matter?

It decides what you compare. If you believe each vendor sells a different AI, you compare demos. If you know most of them rent engines from the same few labs, you compare what actually differs: fit, access, approvals, cost, and who keeps your context. A confusing market becomes a short list of questions.

The cost of missing this is quiet. Teams pay twice for the same capability in two tools. They lock years of context into a product they can't leave. Or they pick a vendor that routes their work to a cheaper model without saying so.

A colleague told me recently that his team runs on Notion. Calendars sync, summaries go out on a schedule, and the AI can search every page they have written. He was sure it was a different kind of AI from what I had been describing. He was right that it works differently. The difference just isn't the AI. In 2025, Notion added the ability to switch between OpenAI and Anthropic models. The engine comes from the same labs. What Notion built is the truck.

How does an AI model actually work?

A model does not remember you. Each time it answers, it sees only what is in front of it at that moment, and it can act only through the tools it has been given. When an app seems to remember you, the app is putting saved notes back in front of the model. Everything else you experience as "the AI," including memory, search across your files, and a link to your calendar, is packaging. Packaging decides what goes in front of the model and what it may touch.

That one fact explains every AI product on the market. Three dials do the work:

• What it sees. Saved memory, search over your documents (often called RAG), written procedures, and project files all choose what goes in front of the model.
• What it touches. Connections to your tools decide what it can actually do, such as reading a calendar or updating an order. They are built through an API, a command line, or the Model Context Protocol (MCP).
• When it runs. A schedule or a trigger, like a new email, decides whether it waits for you or starts on its own.

Turn those dials and the same model goes from a clever chat partner to a colleague who knows your work.

What does an AI vendor actually sell you?

A vendor sells you a truck. The model is the engine. Trucks and vans from different brands often share an engine from the same maker. You buy a work truck for the bed, the cab, and the tow package, not because its engine is secretly better. AI products work the same way. Engineers even use the same word: the setup around a model is called its harness.

The analogy holds the nuance, too:

• The chassis matters. A great engine in a bad truck still loses. A May 2026 position paper argues that on long, multi-step tasks with comparably capable frontier models, the harness around a model often matters more than which model you picked.
• The truck can add real value. Some vendors bring data the labs don't have, specialized parts like transcription or search, or a lab's engine tuned for a narrow job.
• The truck can also take value away. A vendor can drop in a smaller, cheaper engine, shorten what it can see, or remove tools, and still call it AI.

So the honest claim is narrow. A vendor can give you a better fit, and sometimes a lab's engine tuned for one job. The engine itself almost always comes from the frontier labs.

What are the three rungs of AI use?

There are three ways to put the same model to work: Ask, Equip, and Station. The engine can be identical on every rung. What changes is what it can see, what it can touch, and when it runs. Most people meet AI on the first rung and assume that is all there is. The lines blur, because chat apps now add memory and connectors. The rungs describe what a setup can do, not which app you open.

The first rung is a brilliant new hire on their first morning, every morning. The second gives that hire a binder and a set of keys. The third gives them a desk that stays open after you go home, with one rule: anything that sends, spends, deletes, or publishes waits for your yes.

The explainer video above was made on the second rung. A frontier model, the same one you can open in a chat window, worked inside our own animation studio with our ant's rig, our scene files, and our render tools. Nothing shipped until a person approved it.

My colleague's scheduled Notion summaries sit on the third rung, prebuilt by a vendor. You can also build your own: rent the engine from a lab, build the truck around your work, and run it on a machine that stays on. The idea is simple. The craft is in the safety. An agent that reads incoming email is reading text from strangers, and that text can carry instructions that try to redirect it. This is called prompt injection, and it tops OWASP's Top 10 for LLM Applications. Least access and human approval are not optional on that rung.

How do you pop the hood on an AI product?

Ask five questions before you buy. They work for a chat app, a coding agent, a vendor platform, or something you build yourself. Write the answers down, because they are the real comparison between products.

1. Which engine is inside, and can I choose?

A serious vendor will name the model and the lab that made it. Ask whether you can pick a stronger one for hard work, and whether they ever route your work to a cheaper one.

2. What can it see, and who decides?

Ask where your context lives and what the AI reads before it answers. More is not better. A few relevant documents usually beat your entire drive.

3. What can it touch, and what waits for my yes?

List every system it can read and every system it can change. Anything that sends, spends, deletes, or publishes should wait for a person.

4. When does it run, and on whose machine?

Know whether it acts only when you ask or on its own schedule. Then find out where your data goes when it does.

5. If I leave, does my context come with me?

Your glossary, procedures, and memory are the most valuable part of any setup. Keep them in a form you own, like plain files, so they move with you to the next truck. We wrote about how owned context turned one person's insight into a verified saving.

What do people get wrong about AI vendors?

Three mistakes come up again and again.

"Their AI is smarter." Usually it is better fitted, not smarter. Fit is worth paying for. Just know that fit is what you are paying for.

"Local means private." A local agent usually still sends its work to a model running in a lab's data center. The harness runs on your machine. The engine usually does not, unless you run a local model, which today usually means a smaller engine.

"An always-on agent is just plumbing." The plumbing is the easy part. Deciding what it may read, what it may change, and where it must stop is the work.

Frequently Asked Questions

Will a vendor give me better AI than ChatGPT or Claude?

Almost never a better model. Most vendor AI features run on models from the same few frontier labs, such as OpenAI, Anthropic, and Google. A vendor can give you a better fit: your data, your tools, your schedule, and approvals built in. Judge the model and the packaging separately, and pay for packaging that fits your work.

What is an AI harness?

A harness is everything around a model that turns it into a working agent: the files and memory it reads, the tools it can use, the rules it follows, and the checks on its work. The same model can perform very differently in different harnesses, so the harness deserves as much scrutiny as the model.

Is a local AI agent private?

Not automatically. A local agent runs its harness on your machine, but it usually sends its work to a model in a lab's data center. Privacy depends on what the agent can read, what it sends, and the provider's data terms. Running a local model keeps more on your machine, with a smaller engine.

What is the difference between a chatbot and an AI agent?

A chatbot answers what you ask, mostly from what you give it in the conversation. An agent works through a multi-step job: it reads your files, uses your tools, and stops at the approval gates you set. The model inside can be identical. The difference is what it can see and touch, and when it runs.

How should a company choose an AI vendor?

Start with the work, not the product. Name the job, the systems it needs, and the steps that must wait for a person. Then ask each vendor which model is inside, what it can see and touch, when it runs, and whether your context leaves with you. Choose the best fit, not the best demo.

If you want a second set of eyes before your next AI decision, we'll pop the hood with you.

---

About the author — William Rodriguez is the founder of Analytical Ants. He builds applications, equips teams to use AI, and provides technical leadership. More about Analytical Ants.