Diligence for Government-Backed AI Health Startups: Harrison.ai's Pivot
Test grant constraints, clinician incentives, and audit trails before trusting adoption metrics in a public-money AI health startup.
When a government-backed AI health startup pivots, the adoption story is only as strong as the grant's strings and the clinicians' pay. Harrison.ai, the Australian AI health startup, launched Frontier Radiology, a US teleradiology business that compensates clinicians for using its diagnostic AI platform. Its 2025 round included a $32 million Australian government investment from the federal government's National Reconstruction Fund, intended to keep its operations in Australia.
The company has raised about $333 million in total funding. Martijn Wilder, chair of the NRFC, said the fund's Harrison.ai stake was meant to keep the company's operations based in Australia. Public money can look like validation, but it is a set of conditions. The conditions decide where the company can put people, what it can sell, and how it reports back. If the conditions are not visible, the adoption metrics need a discount.
Harrison.ai moved Australian staff into hybrid roles in which they manage fleets of AI agents, and the change caused Australian redundancies. Affected Australian employees were told in April that the company would consult for one week before choosing which staff to make redundant. The sequence shows a pivot under pressure: the company was reorganizing around AI agents while the investment was still fresh.
Start with what the grant actually controls
A grant carries political and operational weight. The first diligence question is what the public investor can enforce, and whether that enforcement is written down. Did the investment require a headquarters, a local team, a data location rule, or a reporting schedule? If the answer is only a public statement, the constraint is weak.
A public investment comes with testable conditions. Request the investment terms, the board rights, the reporting obligations, and any repayment or exit language. If the company cannot show where the public money sits in the cap table and the operating plan, the adoption story is incomplete.
Clinician pay can manufacture adoption
Clinical AI adoption metrics are easy to inflate. If a doctor earns more when the system is used, usage rises. Usage does not prove clinical value. It proves the incentive worked.
Frontier Radiology's pay calculator projected that radiologists using Harrison.ai's tools would be 30% more productive and would receive a 25% bonus for the additional work attributed to the AI. The link is direct. Check whether the bonus is tied to volume, speed, or override behavior. If it is tied to volume, the adoption number is partly a compensation number.
Monash University's Michelle Lazarus warned that Frontier Radiology staff might feel pressured to rely on Harrison.ai's tools and follow their suggestions, creating an automation-bias risk: staff may accept the tool's answer without checking. A clinician who is paid to move faster may not be the best judge of whether the tool is safe. Determine whether the pay plan rewards careful review or merely speed.
The audit trail is the real product
In clinical AI, the model is only one part of the product. The record that shows what the model said, what the clinician did, and whether the final decision can be reconstructed is the other part. Without that record, the safety claim is a promise.
Ask for the override rate, not just the adoption rate. An override rate is the share of AI suggestions that clinicians change or reject. It is the closest observable sign of trust. If the company reports high usage but cannot show how often clinicians disagree, the usage number is thin.
Trace where the data flows. Patient images, metadata, and clinician decisions may move through multiple systems. The test is whether the vendor can map the path, identify who can access it, and show retention limits. If the answer is a diagram without timestamps, the audit trail is weak.
Run the three checks before the term sheet
- Request the grant or investment terms, including location, reporting, and exit clauses; if they are not produced, treat the public-money claim as unverified.
- Compare the clinician compensation plan to the adoption metric; if bonus depends on AI-attributed volume, discount the raw usage number.
- Get the most recent override logs, including clinician identity, timestamp, and final decision, and trace where patient data leaves the hospital, where the AI runs, and who can see the output; if the path is not mapped, the audit trail is incomplete.