Cold start: what to do when you have no data for the use case
The use case is new, so there is no historical labeled data to train on. That is a real constraint, not a dead end. There are four honest moves, and one of them is often to not build it yet.
Cold start: what to do when you have no data for the use case
Someone wants a model to do a thing the company has never done before. There is no history of that decision being made, so there is no labeled data to learn from. This is the cold-start problem, and pretending it is not there is how a pilot burns a quarter on a model that had nothing to learn.
The honest starting point
A product and operations leader who now advises mid-sized companies on AI starts every engagement from the business problem, not the technology, and is willing to tell a client that AI is not the right fit yet. Cold start is one of the moments that conversation earns its keep. If the company has never captured examples of the outcome it wants to predict, no amount of model choice fixes that. You cannot learn a pattern from data that does not exist. The first move is to say so plainly, before anyone spends the budget.
Naming the constraint is not giving up. It reframes the project from “train a model” to “get to a model,” which is a different plan with real options.
Option one: bootstrap the data on purpose
If the labels do not exist, you can create them. Have experts label a starter set by hand. Run the process manually for a while and record every decision as a labeled example. Generate synthetic cases for the situations you can describe. None of this is free, and a hand-labeled set of a few hundred good examples takes real time from people who have other jobs. But it turns “no data” into “a small, deliberate dataset,” which is often enough to test whether the idea works at all before committing to more.
Option two: keep a human in the loop from day one
When you cannot trust a cold model to run alone, you put a person in the path and let the system earn autonomy. A founder building a voice agent that separates good-intent leads from bad-intent ones keeps a human in the loop for exactly this reason. His buyers say the quiet part out loud: “I don’t know if it’s giving me the right data.” So the agent proposes and a person confirms, and every confirmation becomes a labeled example. The human is both the safety net and the data source. Over weeks the system accumulates the history it started without, and the human can step back as the evidence grows.
Option three: buy time with something off the shelf
Sometimes the fastest path to value is to not train anything yet. McKinsey’s 2025 read is that roughly 70% of enterprise AI use cases are adequately served by off-the-shelf tools. A generic model or an existing product may cover the use case well enough to ship now, deliver value, and, crucially, start generating the interaction data a custom model would need later. You solve the problem today and you collect the labels for tomorrow at the same time. For a cold-start use case, that is often the strongest opening move.
Option four: wait, and instrument for it
The last option is the one teams hate to hear. If the use case is not worth bootstrapping and no off-the-shelf tool fits, the answer might be to wait. Not wait idly. Wait while you instrument the manual process so it captures clean, labeled records of every decision, so that in six months you have the dataset you lack today. This is a deliberate choice to build the runway before the plane. It beats training a model on nothing and calling the poor results a model problem.
The wrong move is to pretend the data is there
The failure is not choosing any of these. It is skipping the question. A team assumes some usable history exists, builds a model on a thin or irrelevant scrap of data, and gets output that looks like an answer and behaves like a guess. Cold start is a design input, and the right response is a plan to acquire data, not a hope that the model will manufacture it.
How we approach it at Density Labs
In the AI Opportunity Assessment, our fixed two week, $2,500 engagement, one of the first checks is whether the data for the use case actually exists. If it does not, we say so, and we lay out the honest paths: bootstrap a labeled set, run with a human in the loop that captures data as it goes, ship an off-the-shelf tool now and collect labels for later, or wait and instrument first. We would rather find the cold-start gap in week one than watch a model train on data that was never there.
You cannot teach a model a lesson your data never recorded.