This research project was carried out by the Swedish Dairy Association as part of the development of a ‘Scheme for Animal Welfare’. As a first step, on-farm, animal-based measurements on calves, young stock and cows in 55 randomly selected herds were performed. Nine animal-based measurements formed the basis for a classification of welfare at the herd level: cleanliness and body condition in calves, cows and young stock, in combination with lameness, injuries/inflammations and rising behaviour which were recorded for cows only. The threshold (gold standard) for being a case herd with poor welfare, was a score lying among the worst 10% on two or more of the nine welfare measurements. Thirteen of the 55 herds were cases fitting this criterion. As a second step, 65 potential welfare indicators from seven different focus areas in a pre-collected data register were identified by expert opinion. In the final step, the extent to which suggested potential welfare indicators predicted farms’ risk of having poor welfare according to the definition was assessed. The final set of welfare indicators, taken from the national dairy database, included two fertility measures and calf mortality, and it correctly classified 77% of the herds, with a sensitivity of 0.62. The inclusion of cow and young stock mortality led to it correctly classifying 76% of the herds with a sensitivity of 0.77. We propose that this approach could be useful in helping to allocate advisory services to farms at risk of poor welfare.