Arpan Saxena is the COO/CIO at basys.ai (based out of Harvard University), a leading healthcare AI solutions company.
Most health plans and provider groups already have the clinical knowledge they need to run well. The problem is where it lives. Some of it sits in medical policies and guidelines. The rest sits in reviewer notes, email threads and the heads of a few experienced nurses and physicians. When that knowledge is scattered, operations pay for it.
I see this often in prior authorization, utilization management, chart review and coding. Two reviewers look at the same case and reach different answers. A policy update takes months to show up in daily work. A new hire takes a year to make calls a senior reviewer makes in minutes. The root cause is usually fragmented knowledge, not the people doing the work.
Here is a framework clinical and operations leaders can use to put that knowledge to work.
Start with context.
“Stable” means different things in different service lines. In musculoskeletal care, it may mean a patient has completed conservative therapy and still shows functional decline. In surgery, it may mean the patient is fit for the procedure and the risks are documented. In diabetes, it may mean glucose control has held over several months.
So the first question is always context: Which service, which patient population, which decision and which policy applies? Get this wrong and good knowledge gets applied in the wrong place. Context also tells you which clinicians should shape the logic. The experts for spine surgery are not the experts for continuous glucose monitors.
Ask these six questions:
1. Identify: Where does clinical judgment matter? Start with workflows where a clinician’s decision changes the outcome or the cost, such as prior authorization, utilization management, chart review and coding. Not every step needs a clinician. Some do.
2. Codify: What knowledge drives the decision? Write down what reviewers actually use. That usually means guidelines, medical policies, coverage criteria and evidence, plus the informal rules experienced staff apply.
3. Standardize: What can become repeatable? Some decisions follow the same clinical logic every time. Those can be applied consistently across reviewers and shifts. Others carry real ambiguity and should stay with a clinician.
4. Embed: Where should knowledge appear in the workflow? Knowledge helps most at the moment of decision, inside the tool the reviewer already uses. A guideline in a shared drive does little for a nurse working a queue of 40 cases.
5. Supervise: Where must clinicians remain in control? Define checkpoints where a clinician reviews, approves or overrides. Place them where errors are most likely, most costly or hardest to reverse.
6. Learn: What should improve over time? Every override, appeal and escalation is feedback. Feed it back into the knowledge base so the gray areas get smaller.
Make variability visible.
Clinicians think differently, and that is often a strength. Two physicians can weigh the same evidence and land in different places for defensible reasons. The problem is variability that no one can see or explain.
This is where AI can help. It can define the gray area: cases where documentation is incomplete, criteria conflict or the pattern differs from similar past decisions. Instead of deciding those cases, it can nudge the reviewer to take a second look or route the case for a double-check. Clear cases move faster, and unclear ones get more attention.
Used this way, AI enables clinicians. Their expertise reaches more decisions with more consistency, and they spend their time where judgment is actually needed.
Define who does what.
Clinical teams are not uniform. In most organizations I work with, roles fall into three groups:
• Clinicians Who Execute: Reviewers and coders making case-level decisions
• Clinicians Who Own Policy: Medical directors and committees who define criteria and approve changes
• Clinicians Who Oversee Quality: Those who audit decisions, handle appeals and track outcomes
Each group needs different knowledge at different points. A framework that treats them as one audience will frustrate all three. When you map a workflow, assign each checkpoint to the role that should own it.
Put checkpoints where they matter.
Too many checkpoints slow work down and train people to click through. Too few let errors pass. Place them where the stakes justify the time: high-cost services, likely denials, cases with weak documentation and actions that are hard to undo.
Make responsibility clear.
When AI is part of a decision, everyone should know who is accountable for it. Each recommendation should show what knowledge it drew on and which clinician approved, changed or rejected it. That record supports appeals and audits. It also shows clinicians that their judgment is recorded and respected.
In a recent World Economic Forum piece, Dr. John Halamka of Mayo Clinic Platform, Amber Nigam and I described this as a human-in-command approach: AI earns authority step by step, oversight matches the risk and clinicians stay accountable for outcomes. The framework above is how that idea shows up in daily operations.
Where can you begin?
Pick one workflow where inconsistency is costing you, such as a high-volume prior authorization category. Run it through the six questions. Map the roles and checkpoints. Then measure what changes: turnaround time, reviewer agreement, overturn rates and appeal outcomes.
Healthcare does not lack clinical knowledge. It lacks a consistent way to use it. Organizations that build one will run faster and make better decisions, with clinicians firmly in command.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

