blog

AI Models

AI Agents

The model is a supplier. The context is the system.

The model is a supplier. The context is the system.

Gal Dudi

Gal Dudi

|

|

Reading Time:

3

min

Table of Contents

I have a simple view of AI. The model is a supplier. The context is the system.
Everything else I believe about building with AI follows from that line.

The bottleneck moved

A modern model can reason through a hard problem better than most software we had five years ago. That argument is over.

Point it at a real company anyway and it fails. It has no idea who asked, what they own, or what was decided last time.

AI stopped being a capability problem. It is now a data problem. The gap between a good demo and a system someone depends on is almost entirely a data engineering gap.

Buy the intelligence. Build everything else.

Intelligence arrives on a price list, for everyone, at the same time. Reasoning, language, token throughput, model weights. You rent them, and so does everyone else.

What you build around them is different. Clean data. A map of who and what connects to what. The rules your business actually runs on. The record of what was decided and why.

Everything you buy has a shelf life measured in months. Everything you build accrues.

Models commoditize. Context does not travel.

Every advantage that lives inside a model disappears the day someone ships a better one.

Context is different. The history of who asked, what they own, what was decided and why is built per company and owned per company. You cannot license it, resell it or hand it to whoever comes next.

It is not a feature. It is an accumulation. I would rather own the accumulation than rent the feature.

Small and sharp beats big and clever

Once the context is structured, the model does not need to be large. It needs to be narrow, and it needs to be evaluated.

One model that can do anything is impressive. A set of models that each do one job is operable. You can name each job, measure it, and replace it without touching the rest.

Narrow models are also cheap and fast. That changes which work is worth automating at all.

Bounded context makes a decision defensible

When AI makes a call, the first question is always the same. Why?

If the answer is "the model felt like it," you do not have a system. You have a gamble. A decision is defensible when the facts behind it are small, named and the same every time. Reproducible means defensible.

I do not trust any AI whose reasoning I cannot reread.

The loop that compounds

This is the part I care about most, because it is the only part that compounds.

A human approves a decision. That approval becomes precedent. The next similar situation arrives with the earlier decision and its reasoning already attached.

So cost per unit of work falls with every unit of work. Not because the model got cheaper. Because the system stopped having to think as hard.

Autonomy is earned, not requested

The best model and the best data are useless if nobody lets them touch anything. So permission has to be engineered.

Watch first. Let the AI work live cases and execute nothing. Compare its calls against what people actually did. Promote one job at a time, with a scope you choose. Keep every action reversible.

Nothing acts until you say so. Every step is reversible. Every decision is documented.

I think this is the only honest way to bring AI into work that matters. Trust is not a setting. You earn it, case by case.

AI does the work. Humans own the decision.

I am not trying to build an AI that replaces people. I am trying to build AI that finishes the job, and leaves a record that proves it.

The person stops chasing context and starts making calls. The machine does the work. The human owns the decision. That split is the whole point.

We are not trying to own a model. We are trying to own the context and the record.

Questions? Knock twice.