AI Best Practices

| PWH Technology Committee

AI Best Practices for Professional Women in Healthcare

By: PWH Technology Committee


Purpose

To provide practical, responsible, and empowering best practices for using Artificial Intelligence (AI) in healthcare and healthcare-adjacent roles. The goal is to help professional women leverage AI to work smarter, lead confidently, and stay ahead—without compromising ethics, privacy, or credibility.

Guiding Principles

1. AI Is an Assistant, Not an Authority

AI accelerates thinking; it does not replace judgment. Final decisions—especially clinical, financial, or people-related—must always be human-led.

Rule of thumb: If you wouldn’t blindly trust a junior analyst with it, don’t blindly trust AI.

2. Protect Patient, Employee, and Organizational Data

Never input: - PHI or PII - Confidential contract or pricing data - Confidential Corporate Sponsor data – Confidential Member data - Internal-only strategic materials - Anything you wouldn’t be comfortable seeing on the front page of the Wall Street Journal. Use de-identified, hypothetical, or generalized data whenever possible.

3. Use AI to Eliminate Low-Value Work

High-impact use cases include: - Drafting emails, presentations, and talking points - Summarizing articles, policies, or long documents - Brainstorming ideas, frameworks, or meeting agendas - Translating complex information into executive-friendly language. AI should buy back your time, not create more work.

4. Maintain Professional Voice and Credibility

AI-generated content should always be: - Reviewed - Edited - Personalized. Ensure tone, accuracy, and intent reflect you and your organization. Authentic leadership cannot be automated.

5. Beware of Bias (Including Polite-Sounding Bias)

AI systems are trained on historical data—which may reflect: - Gender bias - Racial or socioeconomic bias - Traditional power structures in healthcare. Question outputs. Ask: - Who benefits from this recommendation? - Who might be missing from the picture? - Does this reinforce the status quo—or challenge it productively?

6. Validate Before You Amplify

Before sharing AI-generated insights: - Fact-check statistics - Confirm sources - Sense-check logic - Pressure-test assumptions AI is confident—even when wrong.

7. Be Transparent, Not Apologetic

You don’t need to hide AI use—but you also don’t need to over-disclose it. Good framing: - “I used AI to accelerate a first draft.” - “This helped me synthesize multiple perspectives quickly.” Using AI is not cutting corners; it’s modern leadership.

8. Stay Within Organizational and Regulatory Guardrails

Follow: - PWH Bylaws - Employer AI policies - HIPAA and regulatory requirements - Legal and compliance guidance. If policies don’t exist yet, default to conservative, ethical use—and advocate for clarity.

9. Continuously Upskill

AI literacy is becoming a leadership competency. Recommended habits: - Experiment with prompts weekly - Share effective use cases with peers - Stay informed on healthcare-specific AI regulations and risks. Confidence comes from competence.

What AI Is Not Appropriate For

  • Diagnosing patients
  • Making autonomous clinical or organizational decisions
  • Final hiring, firing, or disciplinary decisions
  • Replacing human empathy, judgment, or accountability
  • Consistency at scale
  • Repeatable workflows
  • Embedded institutional knowledge
  • Guardrails for high-risk or high-visibility outputs
  • Same type of deliverable
  • Same structure
  • Same inputs
  • Same standards
  • Quarterly Business Review slide drafting
  • Contract review summaries
  • RFP response frameworks
  • Executive briefing templates
  • Policy interpretation assistant
  • Your org’s language
  • Your methodology
  • Your service lines
  • Your pricing models
  • Your compliance standards
  • Regulatory language matters
  • Legal/compliance positioning matters
  • Financial assumptions must be standardized
  • Brand voice must stay consistent
  • Controlling tone
  • Defining exclusions
  • Setting clear boundaries
  • Providing approved frameworks
  • Use it weekly
  • Produce standardized outputs
  • Benefit from speed + alignment
  • The use case is vague (“help with strategy”)
  • The task changes every time
  • You don’t have standardized source material
  • You don’t know what “good output” looks like
  • It’s a one-time project
  • It requires constant manual oversight
  • Required inputs
  • Optional inputs
  • Format standards
  • Data assumptions
  • Client name
  • Spend category focus
  • Savings performance YTD
  • Engagement score
  • Identified pipeline opportunities
  • Length
  • Tone
  • Structure
  • Formatting
  • Required sections
  • “Do not include” constraints
  • Executive tone
  • Bullet-based slides
  • No emojis
  • No speculative financial claims
  • Clear call-to-action at end
  • Best past examples
  • Templates
  • Approved decks
  • Policy documents
  • Messaging frameworks
  • What it should NOT do
  • What assumptions it cannot make
  • Where it must flag uncertainty
  • Where human review is required
  • Incomplete data
  • Conflicting data
  • Edge-case scenarios
  • Worst-case compliance scenarios
  • Version updates
  • Improvement logs
  • Feedback loops
  • Clear owner
  • Onboarding accelerators
  • Knowledge retention engines
  • Methodology enforcers
  • Institutional memory systems
  • Change management tools
  • Is the work repeatable?
  • Is there a clear “gold standard” output?
  • Will multiple people use it?
  • Does it reduce risk or save meaningful time?
  • Do we have high-quality source material?
  • Standardization
  • Risk reduction
  • Speed with alignment
  • Scaling expertise without scaling headcount

When creating a Custom GPT is Beneficial 

A custom GPT should be built when you need:

  • Consistency at scale
  • Repeatable workflows
  • Embedded institutional knowledge
  • Guardrails for high-risk or high-visibility outputs

It should not be built just because “AI is cool” or because a team wants a novelty tool. If it doesn’t save time, reduce risk, or increase quality in a measurable way — don’t build it.

Part 1: The Best Scenario to Create a Custom GPT

✅ Ideal Use Case Characteristics

Create a custom GPT when the work:

1. Is Repetitive and Pattern-Based

  • Same type of deliverable
  • Same structure
  • Same inputs
  • Same standards

Examples:

  • Quarterly Business Review slide drafting
  • Contract review summaries
  • RFP response frameworks
  • Executive briefing templates
  • Policy interpretation assistant

If you’re rewriting the same instructions to ChatGPT over and over — that’s your signal.

2. Requires Institutional Context

If the GPT needs:

  • Your org’s language
  • Your methodology
  • Your service lines
  • Your pricing models
  • Your compliance standards

That’s custom GPT territory. Generic AI doesn’t know your playbook. Custom GPT = embeds your playbook.

3. Needs Guardrails

Create a custom GPT when:

  • Regulatory language matters
  • Legal/compliance positioning matters
  • Financial assumptions must be standardized
  • Brand voice must stay consistent

A custom GPT reduces risk by:

  • Controlling tone
  • Defining exclusions
  • Setting clear boundaries
  • Providing approved frameworks

Think: “AI with bumpers.”

4. Must Be Used by Multiple People

If only one person needs it occasionally → not worth it. If 10+ team members will:

  • Use it weekly
  • Produce standardized outputs
  • Benefit from speed + alignment

Now you’re building leverage.

5. Has Clear ROI

Ask this blunt question: Will this save at least 5+ hours per month per user? If not, skip it.

Part 2: When NOT to Create a Custom GPT

Do NOT build one if:

  • The use case is vague (“help with strategy”)
  • The task changes every time
  • You don’t have standardized source material
  • You don’t know what “good output” looks like
  • It’s a one-time project
  • It requires constant manual oversight

If the process is chaotic, AI will only automate the chaos. Fix the process first.

Part 3: How to Create a High-Performing Custom GPT

Step 1: Define the Job

Be painfully specific.

Instead of: “Helps with QBR prep.”

Define: “Drafts a 10–15 slide executive-level QBR narrative aligned to Vizient’s spend management framework, using structured financial and engagement data inputs.”

Clarity = quality.

Step 2: Identify Inputs

Document:

  • Required inputs
  • Optional inputs
  • Format standards
  • Data assumptions

Examples:

  • Client name
  • Spend category focus
  • Savings performance YTD
  • Engagement score
  • Identified pipeline opportunities

Garbage in = garbage out.

Step 3: Define Output Format

Never leave output ambiguous. Specify: 

  • Length
  • Tone
  • Structure
  • Formatting
  • Required sections
  • “Do not include” constraints

Examples:

  • Executive tone
  • Bullet-based slides
  • No emojis
  • No speculative financial claims
  • Clear call-to-action at end

Custom GPTs perform best with tight constraints.

Step 4: Load High-Quality Reference Material

This is where most people underinvest. Upload:

  • Best past examples
  • Templates
  • Approved decks
  • Policy documents
  • Messaging frameworks

Bad source material = scaled mediocrity.

Step 5: Build Guardrails

Explicitly state:

  • What it should NOT do
  • What assumptions it cannot make
  • Where it must flag uncertainty
  • Where human review is required

Example: If financial data is missing, prompt user instead of estimating.

Build skepticism into the GPT.

Step 6: Test Like a Skeptic

Pressure test with:

  • Incomplete data
  • Conflicting data
  • Edge-case scenarios
  • Worst-case compliance scenarios

If it survives that, it’s production-ready.

Step 7: Version Control

Treat your GPT like a product:

  • Version updates
  • Improvement logs
  • Feedback loops
  • Clear owner

If no one owns it, it decays.

Part 4: Advanced Strategy — Think Bigger

Here’s where it gets interesting. Custom GPTs are not just productivity tools. They can become:

  • Onboarding accelerators
  • Knowledge retention engines
  • Methodology enforcers
  • Institutional memory systems
  • Change management tools

Instead of asking: “What task should AI help with?”

Ask: “What capability do we need to scale without hiring 5 more people?”

That’s executive-level thinking.

Final Litmus Test

Before building a custom GPT, ask:

  • Is the work repeatable?
  • Is there a clear “gold standard” output?
  • Will multiple people use it?
  • Does it reduce risk or save meaningful time?
  • Do we have high-quality source material?

If you can’t confidently answer yes to at least four — pause.

Bottom Line

Custom GPTs are not about automation.

They’re about:

  • Standardization
  • Risk reduction
  • Speed with alignment
  • Scaling expertise without scaling headcount

Build them intentionally. Govern them seriously. Measure their impact. And if it doesn’t move the needle — don’t build it.

Final Thought

AI won’t replace women in healthcare—but women who understand AI will replace those who don’t.

Use it wisely. Use it boldly. And always stay human.

Use the Best Practices Checklist to put this guidance into action. 

Prepared for: Professional Women in Healthcare