aceFrom accuracy to service: deciding what artificial intelligence outputs may do in veterinary diagnostic laboratories
Abstract: 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 decisi…
Hospital Universitario Doctor Peset, Valencia 01 by 19Tarrestnom65. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0
On August 24, 2026, a commentary in the Journal of Veterinary Diagnostic Investigation argued that veterinary diagnostic laboratories should decide what AI outputs are permitted to do in service, not just how accurate models are. It defines service entry as the moment an output can influence triage, interpretation, draft reports, or result release, and proposes a standard operating procedure covering intended use, reviewer and signer roles, disclosure, input compatibility, refusal conditions, pathologist override, QC monitoring, and stop rules.
The distinction matters because a deep-learning example for canine lymphoma cytology performed strongly for lymphoma versus reactive hyperplasia but less favorably for B-cell versus T-cell classification, which usually requires ancillary immunophenotyping. The authors use this to argue laboratories must ask whether a Romanowsky-stained image holds enough information for the intended claim and ensure interpretive outputs reach clients only through a pathologist-reviewed and pathologist-signed report, leaving uncertainty about how labs will validate local compatibility and enforce stop rules in practice.
- Commentary defines service entry as the point an AI output is allowed to influence triage, interpretation, draft report, or result release.
- Authors propose documenting 7 decisions in an SOP: intended use, reviewer/signer and disclosure, input compatibility, refusal conditions, pathologist override, QC monitoring, and stop rules.
- Canine lymphoma example shows strong performance for lymphoma vs reactive hyperplasia but weaker performance for B-cell vs T-cell classification requiring ancillary immunophenotyping.
The same canine lymphoma cytology AI performed less favorably on B-cell versus T-cell classification, prompting concern that allowing outputs to influence reports without pathologist review and without assessing whether the specimen supports the claim could lead to unsupported clinical claims.
The rundown
The commentary frames the decision point as service entry, defined as when an AI output is allowed to influence case triage, interpretation, a draft report, or result release, and argues accuracy alone should not determine permission.
Before service entry, staff should assess whether the submitted specimen can support the question, then document intended use, who reviews and signs and whether AI use is recorded or disclosed, input compatibility, refusal conditions, override, QC monitoring, and stop rules.
Using canine lymphoma cytology, the authors contrast a task close to routine cytomorphology (lymphoma versus reactive lymphoid hyperplasia) where the model performed strongly, with B-cell versus T-cell classification where performance was less favorable and typically needs ancillary immunophenotyping.
Sources
- Peer-reviewedJournal of Veterinary Diagnostic Investigation2026-08-24
How should this claim be treated?
The debate