AI exposes chaos: turning automation pilots into operational wins
Healthcare talks a lot about AI for diagnostics. The faster wins are in operations-where time, money and morale leak through inefficient processes. At the HIMSS Global Health Conference & Exhibition, Michael Vipond, head of Healthcare Provider Go-to-Market at ServiceNow, didn't mince words: "AI exposes chaos."
Once you start automating, fragmented data and broken workflows move from background noise to bright red alerts. As Vipond put it, "The intelligence is the easy part with foundational LLMs." The harder part is wiring insights into daily work so they drive action-not reports.
Why AI pilots stall
Too many projects become what Vipond calls "very expensive advice." They produce recommendations that never connect to a ticket, a task or a change in the system of record.
The fix is boring and essential: governance, ownership and integration. AI must resolve something, update something, or route something-automatically-inside the platforms your teams already use.
Northwestern Medicine's playbook: adoption first
Teresa Incesi, director of IT at Northwestern Medicine, was blunt: "Getting adoption was our No. 1 challenge." The team led with education and clear communication so staff understood AI would support their work, not replace it.
They started by cleaning up information structure and improving search and knowledge management across systems. From there, they prioritized high-volume workflows in clinical, administrative, HR and finance operations where automation could show quick impact.
Clinical teams leaned in with ideas like surfacing radiology findings and assisting incident management. IT applied AI to service management: automated resolution notes and support workflows. Much of this ran on Microsoft technologies, Epic and ServiceNow, with targeted custom models where it made sense.
"We knew we had to get ahead of AI," Incesi said. Trust grew by inviting frontline teams into pilots, staying transparent about limits, and letting early users test in real workflows. "That's how you develop evangelists."
Make AI useful: connect insight to action
- Define the "last mile" before you start: where will AI write, route, update or resolve work?
- Instrument the workflow: every recommendation needs a destination (ticket, case, order, message).
- Codify ownership: who approves, who monitors, who retrains, who shuts it off?
- Keep humans in the loop where risk is high; auto-resolve low-risk, repetitive tasks.
High-yield use cases for operations
- IT service management: summarize incidents, generate resolution notes, auto-route tickets.
- Contact center ops: intent detection, next-best action, knowledge article suggestions.
- Clinical operations: surface radiology findings to the right queue; triage incident reports.
- Revenue cycle: draft denial appeal letters; flag documentation gaps early.
- HR and onboarding: policy Q&A, automated checklist progression, case deflection.
- Supply chain: exception handling, vendor inquiry summarization, reorder suggestions.
- Scheduling: preference-aware templates; fill-rate predictions and escalations.
Governance that scales
Set up a lightweight but firm structure so pilots don't drift. Use a standard risk tiering, approval path and monitoring rubric. If you need a reference, the NIST AI Risk Management Framework is a solid starting point.
- Catalog: one inventory for models, prompts, data sources and owners.
- Policy: define what can auto-execute vs. require review; document data boundaries.
- Quality: measure accuracy, false positives/negatives, and drift monthly.
- Safety: red-team sensitive prompts; log decisions; enable quick rollback.
Platform over patchwork
Northwestern Medicine standardized on enterprise platforms already in use-Microsoft, Epic and ServiceNow. That let teams reuse skills, plug into existing data and scale without multiplying tools.
"The platform approach was really the only decision I had," Incesi said. Consolidation reduces integration pain, shrinks support load and keeps governance enforceable.
Change management that wins trust
- Message clearly: "AI will assist you, not replace you."
- Put staff in the pilot seat: let them test on live work and shape the rollout.
- Be transparent: disclose where AI helps, where it struggles and how feedback is used.
- Close the loop: publish weekly wins, fixes and next steps to build momentum.
Avoid these traps
- Automating chaos: fix data fragmentation and broken handoffs first.
- Pilots with no pathway to production: define integration points up front.
- Shadow AI: centralize access, auditing and model updates.
- Over-customizing early: start with platform-native features, then add bespoke models.
- Ignoring knowledge bases: train models on vetted content and keep it fresh.
- Security as an afterthought: set permissions and PHI handling on day one.
Metrics that matter to operations
- Time to resolution and first-contact resolution
- Queue backlog and average handle time
- Clinician/admin time returned (hours/week)
- SLA adherence and exception rates
- Model performance (precision/recall) and drift
- User satisfaction and adoption by role
Your 90-day rollout plan
- Weeks 1-2 (Discovery): Map one high-volume workflow; define the last-mile action; set governance and metrics.
- Weeks 3-8 (Pilot): Ship a human-in-the-loop version; integrate with your platform (Epic/Microsoft/ServiceNow); publish weekly metrics.
- Weeks 9-12 (Scale): Promote 2-3 low-risk actions to auto; train "evangelists"; expand to the next adjacent workflow.
AI will surface every brittle handoff and scattered dataset you've learned to live with. That's not a problem-it's the blueprint. Fix what it reveals, wire insights into action and you'll reclaim time for patient care at scale.
Want more practical examples and training for health system teams? Explore AI for Healthcare.
Your membership also unlocks:
