GHADA ELKHODIRY / EXECUTIVE DECISION LABORATORY
A practical method for deciding what an AI system may do, who can stop it, and what evidence justifies its continued use. Designed for commercial leaders working with medical, legal, regulatory, data and technology teams.
An AI recommendation can be technically accurate and still be inappropriate for the context in which it will be used. The first question is not whether a model can produce an output. It is whether an accountable person has the information, authority and time to decide what happens next.
Two independent gates govern every proposed use: AUTHORITY — can a named, qualified owner challenge, reject, escalate and suspend the recommendation? EVIDENCE — has the team demonstrated a measurable workflow benefit against a credible non-AI alternative, with reliable performance and acceptable controls? Neither gate can compensate for failure of the other.
AUTHORITY READY + EVIDENCE READY → Consider controlled scale, subject to applicable approvals and monitoring.
AUTHORITY READY + EVIDENCE WEAK → Keep the pilot bounded; define the missing proof and retest.
AUTHORITY WEAK + EVIDENCE STRONG → Do not scale. Repair the review rights, workload, escalation and audit trail first.
AUTHORITY WEAK + EVIDENCE WEAK → Stop or fundamentally redesign. A persuasive demonstration is not a business case.
1. What decision or workflow changes, and who owns its consequences?
2. Does the system inform, recommend or execute—and what is it explicitly prohibited from doing?
3. Can a reviewer see limitations, reject an unsuitable output, escalate concerns and suspend use without penalty or workaround?
4. What end-to-end outcome improves against the current process or a simpler alternative?
5. Which reliability, privacy, medical, legal, regulatory and economic thresholds must be met before scale?
Imagine a fictional AI assistant proposing an HCP engagement action using approved information. The recommendation is plausible but unsuitable for the customer's permitted context. A reviewer should be able to identify the limitation, reject the action, record an appropriate reason and route a recurring issue to its owner. A high approval rate alone cannot prove meaningful oversight.
Test representative cases, including deliberate failures. Measure review burden and total workflow time, not just model speed. Record override reasons, recurring defects, approval ownership and operational cost. Where mandatory controls fail, pause or remediate before further deployment.
These published perspectives explain the two sides of the framework in more detail:
Who Is Accountable When AI Recommends an Action? · Which AI Projects Deserve Funding?For regulated uses, assess the applicable jurisdiction and current requirements with qualified specialists. These sources provide useful reference points, not a substitute for internal governance or legal review.
NIST AI Risk Management Framework · FDA Office of Prescription Drug Promotion FAQsTwo companion methods are being prepared for a later release: a Human Override Readiness Checklist and a Five-Gate AI Pilot Investment Scorecard. They are not yet available publicly.
Ghada Elkhodiry writes independently about commercial leadership, customer engagement and enterprise transformation in pharmaceutical organizations. Her focus is the operating decisions that connect strategy, markets, people and technology.
Read more about Ghada ElkhodiryIllustrative framework and scenario. The views I share are my own and do not represent those of my employer.
Authority and evidence matter only when they alter the way work is done. Before scaling an AI recommendation system, identify the manual step it replaces, the decision it improves, the person who owns the next action, and the exception that sends the process back for review. Track both the intended new workflow and the shadow work that persists alongside it.
A useful outcome is not more AI-generated recommendations. It is fewer unsupported decisions, fewer unowned handoffs, less duplicate work and clearer accountability—without compromising required medical, legal, regulatory or privacy safeguards.
Read the 30-day transformation reality checkSee how to repair one omnichannel handoffGhada Elkhodiry · Independent perspectives on pharmaceutical commercial leadership.
Explore all articles and ideasThe views I share are my own and do not represent those of my employer.