How to Evaluate AI Software for a Laboratory

1. Define the task precisely

• State exactly what the tool must produce, and what a correct output looks like
• Establish current performance and cost of the manual approach as a baseline
• Decide whether output influences a reported result, which raises the bar considerably
• Identify who remains accountable for the output

2. Test on your own data

• Use representative samples, including difficult and edge cases
• Include cases where the correct answer is already known
• Quantify errors: false positives, false negatives, and how bad each is for you
• Check behaviour when inputs fall outside what the tool was trained on
• Test reproducibility: does the same input give the same output over time
• Never accept vendor benchmark figures as evidence for your application

3. Data governance and confidentiality

• Where is data processed and stored, and under which jurisdiction
• Is your data used to train vendor models, and can that be disabled contractually
• Client confidentiality and contractual restrictions on sharing sample data
• Retention, deletion and export rights
• Security posture, access control and audit logging
• Include the tool in supplier assessment and, if regulated, data integrity review

4. Validation and control

• Treat it as a computerised system: intended use, risk assessment, validation and change control
• Establish what happens when the vendor updates the model version
• Define mandatory human review points and record them in procedures
• Set ongoing monitoring for drift and periodic re-verification
• Document limitations so users know when not to rely on it

5. Questions for vendors

• What was the model trained on, and how similar is that to our data?
• What performance metrics can you evidence, and on what test set?
• How is model versioning handled and communicated?
• Can we run a trial on our own data before committing?
• What happens to our data, and can we get it back?
• What does the tool explicitly not do well?

A note on pricing

We do not publish instrument or software prices. Real transaction prices depend on configuration, condition, accessories, service coverage, installation, region and negotiation, and published list figures are rarely what a laboratory pays. Where you need a figure to plan with, request quotations against a written specification.