Day 47 is when rollout plans get honest. The enthusiastic physician is still using the new workflow. Two colleagues have quietly returned to the old one. The front desk has a workaround nobody wrote down, and the project lead is still calling the launch a pilot.
A 90-day plan should make that drift visible early. It needs one owner, one workflow at a time, written stop rules, and a way back that staff can use without improvising. Ninety days is a planning box, not a vendor promise. Extend it when the evidence is incomplete.
Key takeaways
- Baseline the old workflow before touching the new one.
- Pilot one bounded job with willing users and representative cases.
- Add modules only after the current workflow meets its pass conditions.
- Rehearse downtime and rollback before broad use.
Days 1-15: write the operating brief
Name the sponsor and the working owner. The sponsor can approve budget and remove blockers. The working owner watches the queue, answers staff, records exceptions, and pauses the pilot when a stop rule is met. Those are separate jobs even when one person holds both.
Choose the first workflow. “Implement AI” is not a workflow. “Prepare a reviewable draft note for routine follow-up visits, then complete the signed note in the designated EHR” is.
Record the old numbers for at least one representative week:
- notes completed before the clinician leaves;
- correction or catch-up minutes;
- unsigned or abandoned notes;
- transfers between systems;
- support requests and downtime workarounds.
NIST’s AI Risk Management Framework organizes AI risk work around Govern, Map, Measure, and Manage (NIST AI RMF). The companion Playbook offers suggested actions rather than a checklist that guarantees compliance. For a small practice, the useful translation is plain: assign authority, map the workflow, choose evidence, and decide how the team responds.
Days 16-30: settle contracts, access, and data
No real patient data yet.
Confirm the business associate agreement and the product’s role in the workflow. Document what data enters, where it is processed, what is retained, who can access it, and how the practice obtains or deletes it. The HHS Office for Civil Rights describes risk analysis as foundational to Security Rule compliance and expects an accurate assessment of risks to electronic protected health information (HHS risk-analysis guidance).
Build roles with least access needed for the pilot. Test joiners, leavers, password recovery, and an account with the wrong role. Verify support escalation and incident contacts. If the product will work beside an EHR, watch the handoff with fictional data. If it will replace the EHR, do not squeeze migration into a documentation pilot.
Write stop rules before optimism gets involved. Examples:
- wrong-patient content or identity mismatch;
- draft and final status cannot be distinguished;
- a required field repeatedly disappears;
- an output reaches a patient or external party without the intended approval;
- the team cannot recover the workflow during downtime;
- access cannot be removed promptly.
A stop rule is not a declaration that the product is unsafe forever. It is permission to pause while evidence is still cheap.
Days 31-45: run a shadow pilot
Select two or three willing clinicians and a narrow set of visit types. Keep the established workflow authoritative. Use the copilot output as a comparison draft until the team understands its failure patterns.
Include ordinary and awkward cases: a long history, background noise, a late correction, two similar names, and a visit whose template does not fit. Track missing information and invented or misplaced content. Do not average those differences into one satisfaction score.
The AMA’s 2025 principles call for physician consent, final review, and a clear protocol when AI-generated content enters the medical record (AMA principles). Translate that into the screen and the policy: users see that the note is a draft, they know what to check, and the sign-off event is recorded.
At day 45, decide. Continue only if the owner can explain the common exceptions and the measured workflow is at least as controllable as the baseline.
Days 46-60: make the first workflow real
Move the bounded workflow from shadow mode into ordinary use for the pilot group. Keep daily office hours short and predictable. Ten minutes at the end of clinic can surface more than a weekly survey because the failed note is still fresh.
Review exceptions by type, not anecdote. Separate transcription errors, missing context, template mismatch, wrong destination, user confusion, and product outage. Each category needs a different remedy.
The ASTP/ONC SAFER Guides address EHR areas such as patient identification, system configuration, result follow-up, and contingency planning (SAFER Guides). Use the relevant guide to challenge the new route. An AI feature should not weaken a safety practice that already exists in the record system.
Days 61-75: add one adjacent module
Only add a second module if the first workflow has met its pass conditions for two consecutive review periods and the exception owner is not overwhelmed.
Pick an adjacent job with a visible boundary. Scheduling may be a better second step than patient messaging because staff can test availability, double booking, cancellations, and takeover without creating clinical communication. Your order may differ. The rule is one controlled change at a time.
Patient Square’s Practice Copilot lists documentation, scheduling, messaging, EHR, receptionist, follow-up, and billing-related modules across the product and plan set. Treat each module as a separate rollout row. A bundle is not one workflow just because it has one invoice.
Days 76-90: rehearse failure and choose the steady state
Turn off the route in a planned exercise. How do staff document? Where do messages wait? Who restores access? How does delayed work return without duplication?
Then test rollback. Export pilot records, remove a user, disable the new module, and return to the established path. A reversible pilot creates negotiating room. A pilot that cannot unwind has already made the purchasing decision.
At day 90, write one page:
| Decision | Evidence required |
|---|---|
| Expand | Pass conditions met, exception load staffed, recovery rehearsed |
| Hold | Workflow useful but unresolved fit, training, or contract work remains |
| Roll back | Stop rule persists, control is weaker, or benefit does not cover the operating cost |
Do not call “hold” a quiet expansion. Freeze the user group and scope until the open items close.
The weekly project note
Keep the project honest with six lines:
- workflow in scope;
- users active;
- pass measure and current result;
- top exception by count;
- open risk and owner;
- next decision date.
The note is intentionally dull. It prevents a good demo, one vocal champion, or a large feature list from substituting for operating evidence.
If Patient Square is on the shortlist,