Data-centric AI, explained as a purchasing decision
Holding the model fixed and fixing the data is often the cheaper route to accuracy, but only if the work is targeted, iterative, and budgeted in the right shape.
The phrase describes a shift in where the effort goes. Instead of iterating on the model and treating the dataset as fixed, you hold the model and the training recipe steady and iterate on the data. The interesting part for a buyer is not the philosophy. It is that data work has a different cost curve from model work, and it rewards a different purchasing shape.
Model work scales with hardware. More accelerators shorten a training run. Data work does not scale that way: two annotators do not produce one annotator's judgment twice as fast, and the throughput ceiling is set by how many people can make the judgment you need.
The three moves, in order of return
Most of the value comes from a small number of unglamorous actions, and they are worth ranking because teams tend to do them in the wrong order.
- Remove what is wrong. Mislabeled and corrupted examples put a ceiling on what any model can learn, because the model has no way to distinguish an error from a hard case. Finding and deleting a small share of bad rows is usually the highest-return work available.
- Add where the model fails. Examples drawn from an error analysis and concentrated in the region the model gets wrong are worth many times their number drawn uniformly from the distribution.
- Re-decide the boundary. Some errors are not labeling mistakes but unresolved questions in the guideline. Fixing them means making a decision and relabeling a defined band of cases, not hiring more annotators.
The workflow is error analysis, and it is not optional
The move that makes the others work is reading the failures. Run the model over the evaluation set, look at every case it got wrong, cluster those cases, count the clusters, and fix the largest one first.
Without that loop, data is bought on intuition, and intuition is drawn to whatever is easiest to describe rather than whatever is largest to fix. A per-slice scorecard is what makes the loop measurable: score by the slices that matter, and the cluster with the widest gap is where the next purchase goes.
One discipline keeps the loop honest. Change one thing at a time and re-measure. Changing the data and the training recipe in the same iteration produces a number that cannot be attributed to either, which is the same as having no number at all.
What it changes about what you buy
The unit shifts from hours or tokens to examples in a named region. That is a different product. A catalog is sold by the volume it contains; a targeted slice is sold by how precisely it hits the region you described.
Three practical consequences follow. A smaller dataset you can inspect in full is more useful than a large one you cannot, because the inspection is where the work happens. The per-file metadata — a record carrying speaker, session, environment, and label provenance — stops being paperwork and becomes the diagnostic instrument, since without it you cannot slice the data to find where the model is weak. And retention matters more than usual: keep the raw material and the guidelines, and labels can be regenerated; keep only the labeled output, and a guideline change means buying everything again.
Where data work stops paying
Two ceilings are worth knowing before committing a budget. The first is informational. No amount of cleaning creates a signal that is not present in the input. If the recording does not contain the distinction, the label cannot be recovered from it, and a relabeling project will spend its budget proving exactly that.
The second is human agreement. Past the point where qualified annotators disagree, the label is not a fact that better training can recover. It is noise. Measuring agreement first tells you whether the boundary you want to sharpen is defined at all: if two annotators cannot agree on a case, a model trained on either answer is being taught a coin flip.
Neither ceiling argues against the approach. They are the two questions to answer before starting: does the signal exist in the material, and is the boundary defined well enough that a person can apply it the same way twice?
Budgeting it in the right shape
Data-centric work is iterative by construction, so a single large purchase is the wrong instrument. What fits is a smaller initial buy with the ability to re-run a slice: the same pipeline, the same guideline, more items in a named region, on a short cycle.
Write that into the agreement rather than hoping for it. The clause to ask for is a priced option on further batches of the same slice at the same specification, with the turnaround stated — because the moment you need the second batch is the moment you have the least room to negotiate.