Your next AI purchase needs an operating case

The CII–PwC report published for HospitalTech on 8–9 September 2026 offers a useful starting point for a hospital's next AI purchasing discussion. Its August pulse survey collected one response per organisation from 28 senior leaders across care providers, medtech, diagnostics and dialysis.

Put one task on the purchase order

The following is Lifeline's applied analysis. A hospital owner needs a decision that can survive a disappointing demonstration as well as an impressive one. Write down the task the proposed tool will perform, the person accountable for its output and the evidence needed to continue paying for it.

Take a hypothetical tool that checks an insurance claim packet for missing documents before submission. Give it a bounded job: flag omissions for the claims executive to review. Keep decisions about treatment, coding and the accuracy of the record with the responsible hospital staff.

For illustration only, suppose the team handles 20 packets a day and the tool removes three minutes of searching from each. That creates 60 minutes of gross capacity. If checking its suggestions takes another 30 minutes, the apparent benefit falls to half an hour. These are invented inputs for a worked example, not a result achieved by a client or vendor.

Measure the whole packet, including corrections and repeat submissions. A quicker first check may simply move work to another desk. Nor is time released automatically money saved: the hospital must identify what useful work fills that time, or what actual expenditure changes.

What this means for your hospital

Establish your own starting point. Before a pilot, record staff minutes per packet, missing-document returns and the time from a complete record to submission. Separate payers and case types where their requirements differ. Compare the same kinds of work before and during the trial.

Test the inconvenient cases. Include incomplete scans, inconsistent names and revised bills in an approved test set. Count omissions the tool misses and unnecessary alerts staff dismiss. A polished demonstration should not determine the acceptance criteria. Agree those criteria in writing before testing begins.

Price the work around the licence. Ask who pays for connection to your hospital information system, configuration, training, support and subsequent changes. Identify where patient information goes, who can access it and how it can be removed or exported. Require an answer your team can verify.

Name the person who can stop it. Assign a claims lead to review errors and pause use when the agreed threshold is breached. Keep a usable manual route. Tie the next payment or expansion decision to the observed results and the terms you agreed with the supplier.

These questions belong in the technology and data requirements review, before vendor selection. For the claims example, also map the actual reasons your hospital's bills are reduced. The useful purchasing question is which recurring piece of work will improve, and how your team will know.