Healthcare administration
AI for healthcare administration that holds up under review
Health plans, TPAs, benefits administrators, and provider services teams are past asking whether to use AI. The questions now are which work deserves it and how to scale it without creating compliance risk or a fight over headcount.
What limits AI adoption
The blockers are rarely technical. They're readiness gaps that have to be named before any AI investment pays off.
No compliant path for the most valuable data: The highest-value use cases sit on regulated member-level data. Without an agreement covering the AI tool, intake review and claimant summaries are off-limits by default, not by choice.
No owner to sponsor a build: When process owners leave, the function and its handoffs go unowned. Without someone accountable for the process, nobody is positioned to authorize, fund, or stand behind an AI build on top of it.
Data and tooling too fragmented to build on: Every team has its own reporting method and its own system with no shared source of truth. Any use case that needs clean inputs hits that first.
Shadow AI, not managed adoption: People across the organization are already experimenting on their own. That's evidence of appetite, not readiness. The cost of maintaining it all grows with no clear read on whether it pays back.
Leadership wants a hard number first: Executives ask for a defensible "percent of work eliminated" before scaling investment and there's no baseline to produce one credibly. That's a measurement gap, not a results gap.
Where AI actually fits (and where it doesn't)
The highest-value use cases aren't “have AI do the work.” They're “have AI catch the gap before it becomes rework.” The fit only becomes clear once the real bottleneck is diagnosed.
Completeness checks on intake
Catching a missing field before it moves downstream is cheap. Catching it three handoffs later is not.
Cross-checking documents that should agree
A benefit configuration against the generated summary, pricing models against the proposal. Flag where they disagree instead of relying on one experienced reviewer's memory.
Surfacing reports nobody checks
Route an existing report to the team that needs it. Often that's a visibility problem, not a generation problem.
Insights from data and conversations
Call notes, email threads, and reports hold the patterns leadership keeps asking about. AI pulls them together from de-identified sources so decisions rest on evidence instead of anecdotes.
A strategy partner for leadership
Pressure-test a plan, model scenarios, and prepare for board and client conversations. AI works as a thinking partner for the people making decisions, not just a production tool.
When the expert is faster
AI can be slower than the expert doing it by hand. Testing that honestly is part of the work and it saves money.
Regulated member-level data
Anything touching it waits until the right compliance agreement is in place. We design the workflow now and switch it on when that's settled.
How I measure it
Never optimize for a single metric. Track the combination or you fix the wrong thing.
Three co-equal metrics, not one North Star: Speed (elapsed time, not just hands-on time), efficiency (touches and handoffs), and accuracy (error rate). A team that got faster but needs more review passes didn't actually win.
Before and after, stated as a range: Improvements are reported as a target being tracked, never as a claim already banked.
Unit economics, not just time saved: Cost per document, tracked from day one, so "we automated it" comes with a real number instead of a feeling.
Headcount stays in the conversation, honestly: When finance wants a bigger efficiency number than the analysis supports, the answer is to harden the underlying numbers (touches, handoffs, volumes, rework rates), not inflate the framing. Credible measurement compounds. Inflated claims don't.
How I work with healthcare teams
The same five steps, with the ground rules first. We design around PHI from day one: Member data stays out of AI tools unless the right agreements are in place. Until they are, we build and test on blank templates and de-identified examples. Where the work overlaps with engineering or another internal project, we map the boundaries before we build so nobody duplicates anybody.
Background
- Led strategy teams at CVS Health and Deloitte
- First product hire at a self-funded health benefits platform, scaled to nearly 10,000 members and over $100M in medical claims
- Built and shipped a GenAI clinical documentation workflow that cut report generation time by up to 75%
- Product leadership at several health tech startups, two of them acquired