From accuracy to service: deciding what artificial intelligence outputs may do in veterinary diagnostic laboratories

Artificial intelligence (AI) tools are entering veterinary diagnostic laboratory service, but reported model accuracy does not determine what the laboratory staff should allow an output to do. This Commentary defines service entry as the point at which an AI output is allowed to influence case triage, interpretation, a draft report, or result release. Before that point, the laboratory staff should first decide whether the submitted specimen can support the question being asked. They should then document 7 decisions in a standard operating procedure: intended use; who reviews the output, who signs the report, and whether AI use is recorded internally or disclosed in the report; local input compatibility; refusal conditions; pathologist override; QC monitoring; and stop rules for removing the tool from service. A canine lymphoma cytology example is used here to illustrate the distinction between model performance and permitted service use. A deep-learning system performed strongly for lymphoma versus reactive lymphoid hyperplasia, a distinction close to a routine cytomorphologic question, but less favorably on B-cell versus T-cell classification, which usually requires ancillary immunophenotyping. Lower performance should prompt a specimen-inference question: Does a Romanowsky-stained cytology image hold enough information for the intended clinical claim? For interpretive outputs such as cytology or hematology, the output should reach the client only through a pathologist-reviewed and pathologist-signed report.