Loading…
Fractional AI Product Manager · Philadelphia, PA
Directors of Digital Health at Jefferson Health, Penn Medicine, and the health tech companies selling into these systems face a specific problem: AI features for clinical users require specs that most AI PMs have not written before. A wrong AI suggestion in a clinical workflow is a patient safety event. The spec must define the confidence threshold for display, the override mechanism, the audit trail, and the failure mode before engineering starts.
Philadelphia also has GSK and Comcast, where AI product requirements are complex but not clinical. The fractional AI PM engagement covers both: clinical AI specs with the rigor patient safety requires, and non-clinical AI product work for pharma and media companies under the same fixed-scope model.
Fixed-scope engagement, with the price agreed before work starts based on the scope of the spec.
Tell us about your clinical AI product challenge.
Clinical AI features require five spec sections that consumer AI products never need. Confidence threshold for display: the spec names the minimum confidence score at which the AI suggestion appears to the clinician. Override mechanism: the clinician can reject any AI suggestion with a single interaction, and the rejection is logged. Audit trail: every AI output, every override, and every time the feature was displayed is recorded with user ID and timestamp.
These are not engineering defaults. They are product decisions. The AI PM writes them into the spec before the first line of code is committed. Features built without these specs retrofit them under pressure after a clinical review flags a patient safety concern.
Failure mode handling is where most clinical AI specs fall short. The spec must define what the feature does when the model returns a low-confidence output, when the Epic integration times out, and when the input data is missing a required field. In a clinical workflow, the feature must not block the clinician from proceeding. It must fail silently or display a neutral fallback.
Nursing workflow constraints are a separate section. The spec documents where in the nursing workflow the AI output appears, how much time a nurse has to evaluate it, and what the tap target size requirement is in a clinical environment where gloves are worn. These are product requirements, not design requirements.
These are the two spec sections that patient safety reviews focus on first. Both require a product decision before engineering can start.
The confidence threshold for display is set by analyzing the consequence of a wrong suggestion at each confidence level. A threshold too low creates noise. A threshold too high withholds useful suggestions. The AI PM documents the rationale for the chosen threshold in the spec.
The clinician rejects an AI suggestion with one tap or one click. No confirmation dialogs. No reason codes required unless the health system specifically requests audit data at that granularity. The override is logged automatically.
The feature runs in production with outputs visible only to the product team before clinical go-live. This reveals how often the feature would have appeared and how often it would have been correct, without any risk to patients.
The spec defines what the AI feature tells the clinician it is doing, in plain language, at the point of use. At Penn Medicine and Jefferson Health, this language goes through clinical informatics review before go-live.
The AI PM runs three stages of validation. Each one is documented before proceeding to the next.
Step-by-step documentation of the current clinical workflow. Exact placement of the AI output in the workflow is confirmed with nursing or physician stakeholders before engineering starts.
Feature runs in production with outputs visible to the product team only. Minimum 200 cases before proceeding. Reveals accuracy, frequency of display, and contradiction rate against actual clinical decisions.
Live for a selected group of clinicians who document any case where the AI output was confusing or introduced workflow friction. Validation ends when 200 cases have been observed with no unresolved concerns.
Discovery call, written scope, fixed price. You approve before we start. No surprise invoices, no scope drift. Clinical engagements and non-clinical engagements are scoped separately.
Shared Slack channel, written updates every Friday, and Loom walkthroughs for every milestone. Works for Jefferson Health vendor teams and distributed GSK product teams equally well.
No recruiting cycle. Typically starts within two weeks of a signed agreement. The engagement ends with documents your team owns and can use with any health system.
Describe the clinical AI feature you are building.
Tell us the clinical workflow it touches, the health system you are targeting, and where the spec is currently incomplete. We reply within one business day.