The harder conversation is with customers who have studied AI. Training is central in everything they read, and they arrive expecting it — usually without the data to do it.
"We want to train a model on our data."
We hear a version of that early in most conversations. It is a perfectly reasonable thing to want, and it is usually the wrong place to begin.
The harder version comes from the customers who have prepared.
The reading is not wrong
Someone who has spent months studying AI has read about training. Pre-training, fine-tuning, evaluation, the loss curves, the papers. Training is not a footnote in that material — it is most of it.
So they arrive expecting every AI solution to have a training component, and they are not wrong about the field. They are describing a different discipline from the one they are about to buy.
What they studied is largely about building models. What they need is a system that deploys one. Those two things share a name, a vocabulary, and remarkably little else. Nothing in the reading marks the boundary either, because from inside the research the deployment is somebody else's problem.
That is an uncomfortable conversation to have with someone who did the work. It is also the most useful thing we can offer them in the first month.
What that sentence usually means
There are a few different things hiding inside "train it on our data", and they do not do the same job.
Fine-tuning changes how a model behaves. Its tone, its format, the way it handles your terminology. Retrieval changes what a model knows about, by putting the relevant material in front of it at the moment it answers.
Most of the time, what someone wants is for the model to know things. That is retrieval, and it needs no training set at all. It is also faster to build and much easier to correct when it is wrong, because you are editing documents rather than retraining.
Reaching for the wrong one is slow to discover. Fine-tune on your documents and the model learns to sound like them, quite convincingly, while still inventing facts that were never in them.
And the data is not there yet
Even when training genuinely is the right tool, the second problem arrives.
A training set is not a pile of documents. It is examples of the thing you want, paired with what a good response looked like, in enough volume and variety to teach a pattern. The documents almost always exist. The pairs, in the engagements we have run, usually do not.
This is the point where the conversation gets difficult, because it sounds like a refusal. It isn't. It is a sequencing problem.
Your first deployment is a data strategy
Here is the reframe worth carrying into that meeting.
The first AI deployment is a data-acquisition strategy disguised as a productivity project. It has to earn its keep on productivity terms, and it should be scoped that way. But while it does that, it produces something that did not exist before. A record of what was asked, what the system proposed, what a person corrected, and what was approved.
That record is what a training set is made of.
For the customer who came in wanting to train, this is the honest answer to their question rather than a deflection from it. They asked how to get to a trained model. This is how — the path runs through a system that is already doing useful work.
Train when it earns it
The gate for moving on is not having accumulated enough rows.
It is that a trained model measurably beats what you already have, on the cases you actually care about. Run that comparison honestly and it will sometimes come out against training. That is a useful result rather than a disappointing one.
If a tuned model does not beat retrieval plus a good prompt, you have bought complexity and nothing else. Complexity has a running cost, and it is paid by whoever maintains the thing after everyone has moved on.
Most never get there, and that is fine
Plenty of organizations run well on retrieval and careful design for years, and never need a tuned model at all.
The phases are only worth running if each one is useful on its own terms. If phase one is useful, stopping there is a decision rather than a shortfall.
Which is the thing worth saying to the person who studied hardest. Nothing in the reading was wrong. It just describes the part of the work that comes last.
A conversation we have in most engagements, and the hardest version of it is with the customers who have prepared most.