Ninety days from today, you can be standing in front of a governance committee with a single AI use case that went live safely, with success metrics that improved, safety metrics that held, equity metrics that were watched, and a record that proves a human was accountable for every output that touched a patient. Or you can be exactly where you are now, having read a great deal and changed nothing. The difference between those two futures is not budget, and it is not the arrival of some better model. It is whether, in the next ninety days, you do the specific, unglamorous, sequenced work this final lesson lays out. This is the send-off: everything the program taught, turned into a plan you can start Monday.

What the Program Actually Gave You

Before the plan, take stock of what you now carry, because the plan is just this knowledge put in motion. Across five levels, the program handed you a compact and durable toolkit. You learned the four machines: the ways clinical AI actually works, from the ambient scribe that drafts your note, to the predictive model that scores risk, to the imaging AI that reads a study, to the assistant that drafts a message or a packet, each with a different shape and a different failure surface. You learned the failure families: fabrication, omission, and bias, the three ways a model's output goes wrong, each a chart-review, malpractice, fraud, privacy, or survey finding waiting to happen. You learned the human-in-the-loop discipline and why automation bias makes it the load-bearing wall of everything. You learned verification by stakes: that you match the depth of your checking to what happens if the output is wrong and how hard it is to undo. You learned governance: how an organization holds all of this together instead of leaving it to individual willpower. And you learned the seven CHAI elements from the Joint Commission and CHAI's Responsible Use of AI in Healthcare guidance: AI policy and governance, patient safety and quality, a designated governance structure, risk and bias evaluation before and after deployment, vendor disclosure of known limits, validation on representative data, and workforce training.

That is the whole apparatus, and it is enough. You do not need a new framework to start. You need to take the one you have and apply it to exactly one thing, well, in ninety days. The plan that follows is that toolkit, sequenced. It is written for the executive who has to make it real: the CMIO, the CNIO, the chief quality officer, the VP of clinical operations, or the physician leader who has been handed the AI file and told to produce something defensible. What follows is not inspiration. It is a sequence of named workstreams, with the mechanics of each, so that on day ninety you can point to work, not slides.

The 90-Day Map: Named Workstreams, Sequenced

The plan runs six workstreams across three thirty-day phases. The workstreams do not run one after another; several start early and continue throughout. What changes phase to phase is the center of gravity. The table below is the map. Treat every number in it as a placeholder for a number you will set with your own leaders, not a target to copy: the point of a sequenced plan is that you decide the thresholds in advance and in the open.

Read the columns as centers of gravity, not walls. Governance and the AI inventory begin in week one because nothing safe can happen without them. Workforce literacy begins in week one and never ends. The pilot itself lives in the final phase, but it is only judgeable because the two phases before it built the accountability, the bounded scope, and the pre-committed metrics that make its evidence trustworthy. Change the order and you get a rollout wearing a pilot's clothes.

Days 1 to 30: Stand Up Governance, Inventory What You Have, Choose One Use Case

The first month is not about technology. It is about the decisions that determine whether anything you deploy will be safe: who governs, what is already in the building, and what you deploy first.

Stand up governance. Before any tool goes live, there must be a named, standing body that owns AI in your organization, the designated governance structure that is the third of the seven CHAI elements. It does not need to be large, but it must be real: a defined group, with clinical, informatics, quality, legal, and equity voices at the table, that meets on a schedule, keeps an inventory of AI in use, and has the authority to approve, monitor, and stop a deployment. The single most common failure in health-system AI is not a bad model; it is a good model deployed with no one accountable for it. The first thing you build is the accountability, because everything else hangs from it.

What does "real" look like in practice, concretely? It looks like a charter of one page that names the members and their roles, a standing meeting on the calendar, an inventory spreadsheet that lists every AI touching a patient or the record with its intended use and its owner, and a written statement that no AI goes live in the organization without this body's sign-off. It does not require a consultant, a new software platform, or a reorganization. It requires a decision that someone is accountable and a calendar invitation that makes it true. The organizations that stall here are usually the ones waiting for the perfect governance structure; the ones that succeed stand up a good-enough body in week one and improve it as they learn. Do not let the search for the ideal committee delay the existence of any committee.

Inventory the AI already in the building. Almost every executive underestimates how much AI is already live in their organization, unowned and ungoverned. Roughly seventy-one percent of hospitals report predictive AI already embedded in the EHR, and that is only the part you can see; underneath it sit sepsis alerts, readmission scores, inbox triage, imaging tools switched on by a vendor default, and staff quietly pasting notes into consumer chatbots. Treat that figure as a number to verify in your own building, not a statistic to repeat, and then go find your real number. Your first governance deliverable after the charter is an honest inventory: every model touching a patient or the record, its intended use, its owner, whether a business associate agreement covers any protected health information it sees, and whether the vendor disclosed its known risks and limits. You cannot govern what you have not counted, and the inventory almost always surfaces an undisclosed or unowned tool that becomes your first real governance decision.

Pick one high-value, bounded use case. Resist the urge to transform everything. Choose a single use case that is high-value (it solves a real, expensive, felt problem) and bounded (its intended use is narrow, its stakes are legible, and its failure modes are ones you can name and catch). Prioritize on two axes at once: value and risk. High value with catchable, bounded risk is where you start; high value with unbounded or invisible risk is where good programs get hurt. Ambient documentation is the archetypal good first choice for many systems: the value is enormous because documentation is the top driver of a roughly forty-two percent burnout rate, the failure modes are known (confabulated findings, wrong laterality, dropped pertinent negatives, an abnormal value hidden in a clean summary), and the verification workflow, a clinician reading and attesting the note, is one clinicians already understand. The point is not that it must be ambient documentation. The point is that your first use case should be one where you can clearly state the value, the failure modes, and where the human verifies. If you cannot state all three in a sentence each, it is the wrong first use case.

The most common failure in health-system AI is not a bad model. It is a good model deployed with no one accountable for it. Build the accountability first; everything else hangs from it.

Days 31 to 60: Define Success, Safety, and Equity, and Build the Funding Story

A use case without metrics is a hope, not a plan, and a use case with only success metrics is how disparities and safety events go unseen. The second month is where you decide, in advance, what you will measure, because metrics defined after go-live are metrics designed to flatter the decision you already made.

Define three linked sets of metrics, matching the one-system principle from the previous lessons. Success metrics answer whether the tool delivered its value: for ambient documentation, time in the note, after-hours EHR minutes, clinician-reported burden, throughput. Safety metrics answer whether it is holding the line on quality: the rate of errors caught in verification, the rate of edits clinicians make to AI drafts, near-misses reported, any signal that unverified output is reaching the record. A rising rate of caught errors is not a failure of the program; it is the verification working, and you want to see it. Equity metrics answer whether the value and the safety hold across the patients you serve: performance and adoption broken out by subgroup, so a good average cannot hide a population the tool is failing. Define all three before go-live, decide who reviews them and how often, and decide in advance what result would make you stop. A metric with no threshold that triggers action is decoration. A pilot you are unwilling to halt is not a pilot; it is a rollout you have not admitted to.

Building the Funding Story and the ROI You Can Defend

Somewhere in the second month, a finance leader or a board member will ask what this costs and what it returns. You need an answer that survives scrutiny, which means building the funding story before you need it, not scrambling for it after. Start with honest cost categories, because AI is rarely a single line item. There is the tool itself (licensing, per-clinician or per-encounter fees), the integration into the EHR, the validation and testing work, the verification labor that never disappears (a human still reads and attests), the training and change-management effort, and the ongoing monitoring that keeps the deployment safe after go-live. A funding story that counts only the license and forgets the verification labor is a story that will collapse the first time someone audits it.

On the return side, resist the temptation to launder a vendor's headline into your board deck. The category is real: a 2025 multi-system study reported burnout falling from roughly fifty-two percent to thirty-nine percent within thirty days of an ambient scribe, and reclaimed after-hours documentation time has genuine value. But those are numbers to verify in your own population, not numbers to repeat blindly. Your defensible ROI is the one your own baseline and your own pilot produce: measure after-hours EHR minutes and clinician burden before you start, measure them again during the pilot, and report the delta you actually observed with your people. When the board asks how you know the return is real, "our own before-and-after measurement, reviewed by governance" is an answer that holds. "The vendor says ninety-six percent" is not. The strongest funding story frames the investment as buying down a felt, expensive problem (burnout-driven turnover, after-hours burden, documentation risk) with a return you commit to measuring honestly, including the possibility that the pilot shows the return is smaller than hoped and the money is better spent elsewhere.

Days 61 to 90: Run a Disciplined, Gated Pilot to Evidence

The final month is the pilot: a small, controlled, closely watched deployment whose purpose is not to prove the tool works but to find out whether it does, in your workflow, with your patients. The distinction matters. A pilot designed to succeed learns nothing; a pilot designed to surface the truth is the only kind worth running.

Keep it small and bounded: a defined group of willing clinicians, a defined patient population, a defined time window, with the verification workflow live and staffed from day one, not added later. Watch the three metric sets you defined. Collect the evidence honestly, including the uncomfortable parts: the errors verification caught, the workflow friction, the subgroup where performance lagged. That evidence, not the vendor's brochure and not the headline accuracy number, is what your governance body uses to decide whether to expand, adjust, or stop. Every accuracy and ROI number you were handed is a number to verify in your own population, and the pilot is where you verify it.

When the safety or equity gate fails, the gate is doing its job. Plan for the pilot to fail its gate, because a fraction of good pilots should. If a confabulated clinical finding reaches a signed note, if the edit rate climbs to where clinicians are effectively rewriting every draft, or if one subgroup's performance or adoption lags the rest, the disciplined response is not to explain it away or to widen the tolerance so the number passes. It is to pause, feed the breach to governance, fix the workflow or the training or the scope, and only then decide whether to resume. Gaming the gate, moving the threshold after the fact so the pilot "passes," is the single fastest way to turn your safety program into theater. The whole value of setting thresholds in advance is that they are allowed to stop you. A gate you would never let stop you was never a gate. At ninety days you will not have transformed the enterprise. You will have something far more valuable: one use case that went live safely, real evidence of what it does in your hands, a governance body that functioned, and a repeatable pattern you can run again on the next use case, and the next. Transformation is not one heroic leap. It is this loop, run with discipline, again and again.

A Worked Ninety Days: One System Sequencing Its First Quarter

Make it concrete with one honest walkthrough. A mid-sized system decides ambient documentation is its first use case. In the first month, it convenes an AI governance group of a CMIO, two frontline physicians, a nurse leader, a quality lead, a compliance officer, and a health-equity representative, puts a monthly meeting on the calendar, and writes a one-line rule that nothing goes live without the group's approval. Its second act is the inventory, and the inventory is where the surprises live: the team finds a sepsis-prediction model running in the EHR that no one owns, an imaging triage tool a vendor enabled by default, and a handful of clinicians pasting notes into a consumer chatbot with no business associate agreement. That last finding becomes the group's first real decision, made before the pilot even starts. Then it selects ambient documentation and states, in three sentences, the value (documentation burden and after-hours EHR time), the failure modes (confabulated findings, wrong laterality, dropped pertinent negatives, an abnormal value hidden in a clean summary), and where the human verifies (the clinician reads and attests the note before signing).

In the second month, it defines its metrics: success as after-hours EHR minutes and a clinician burden survey; safety as the clinician edit rate on drafts, errors caught before signing, and any note signed without evidence of review; equity as adoption and edit patterns across clinician groups and any signal the tool serves some patient populations worse, for instance where language or complexity differs. It sets its gate and its stop rule in advance: if a confabulated clinical finding reaches a signed note even once without being caught, the pilot pauses and the workflow is fixed before proceeding. In parallel, the finance partner builds the cost picture (license, integration, the verification time that stays, training, monitoring) and records a clean baseline of after-hours minutes so the eventual ROI is measured, not asserted.

In the third month, ten willing physicians run the tool on their own patients for six weeks, with the attestation workflow live from day one and a weekly fifteen-minute huddle to surface problems. The evidence comes back mixed and honest, which is exactly what a real pilot produces: after-hours time dropped meaningfully against the baseline they recorded, most clinicians loved it, the edit rate was healthy and stable, verification caught a handful of confabulated findings before they were signed (the workflow working as designed), and one signal emerged that notes for patients seen with an interpreter needed more editing, a possible equity flag the governance group now investigates before expanding. The equity gate holds the expansion until that flag is understood, which is the gate doing exactly what it was built to do. At day ninety the system has not transformed itself. It has one use case live and safe, a governance body that met and functioned, a cleaned-up inventory, a metrics dashboard it trusts, a defensible ROI story built on its own before-and-after data, a known issue it is addressing rather than ignoring, and a template it can now run on its next use case. That is what success actually looks like, and it is repeatable.

The Workforce Foundation That Runs Underneath

One thread runs under all ninety days and does not stop at the end of them: building the workforce-literacy foundation, the seventh CHAI element. A tool is only as safe as the people using it, and automation bias guarantees that even excellent clinicians will, under load, drift toward deferring to an authoritative output. The defense is a workforce that understands what the tool can and cannot do, knows the failure modes to watch for, knows where and how to verify, and knows it is legitimate, even expected, to question the AI. This is not a one-time in-service. It is a standing capability you begin building in the first thirty days and never finish, because every new tool and every new hire reopens it. The plan deploys a use case; the literacy is what lets the next hundred use cases be deployed safely. Invest in the people, or the best-governed tool in the world will still be undone at the point of use.

The Iron Rule That Sends You Off

Every level of this program, whatever its subject, has been carrying one sentence, and now that you are at the end, you can see that the entire curriculum was an elaboration of it. The four machines are the things it governs. The failure families are why it is needed. The human-in-the-loop, verification by stakes, governance, and the seven CHAI elements are the structures that keep it true. The regulators and accreditors are the world's attempt to enforce it. And the AI-native health system is what it looks like built at scale. The sentence is this: AI assists, the clinician decides, the record proves it.

Sit with each clause one final time, because you will carry them into every deployment of your career. AI assists: the tool is a powerful instrument that drafts, scores, reads, and suggests, and you should use it, because the burnout it can lift and the access it can extend are real and the patients are waiting. The clinician decides: the accountability never transfers to the model, because "the AI said so" is not a defense to a board, a plaintiff, a family, or a surveyor, and the moment you forget that is the moment automation bias turns a model's error into a patient's harm. The record proves it: a decision that is not documented as a human decision is indistinguishable, later, from a machine acting alone, and the short note explaining why you agreed or disagreed with the AI is what makes your judgment defensible and your care traceable. Three clauses. Every one load-bearing. Together, the whole discipline.

You did not enroll in this program to learn a tool. Tools change; the ambient scribe of today will be superseded, the models will change hands, the interfaces will be redrawn. You enrolled to learn how to stay the accountable human at the center of a system that is racing to automate around you, and to do it without either rejecting the technology's real gifts or surrendering the judgment that only you can provide. That is a discipline, not a product, and it will outlast every tool you ever touch. So take the ninety-day plan and start. Stand up the governance. Count what is already in the building. Choose the one use case. Define success, safety, and equity, and build the funding story you can defend. Run the honest, gated pilot. Build the literacy. And carry the rule into all of it, the way you carry every other thing that keeps a patient safe: as second nature, as professional reflex, as the quiet promise underneath the work. AI assists. You decide. The record proves it. Now go build it.

Key Takeaways

• Ninety days is enough to change your reality: one AI use case live and safe, with metrics that improved, safety that held, equity that was watched, and a record proving human accountability, versus having read much and changed nothing. The plan is the program's toolkit (the four machines, the failure families of fabrication, omission, and bias, human-in-the-loop, verification by stakes, governance, and the seven CHAI elements) sequenced, not a new framework.
• The plan runs six named workstreams across three phases: governance, an inventory of the AI already in the building, use-case prioritization by value and risk, a safety-and-equity gate, workforce training, and monitoring with ROI. Several start in week one and continue throughout; only the center of gravity moves phase to phase.
• Days 1 to 30: stand up a real, named governance body first (the most common failure is a good model with no one accountable), inventory every AI already live and unowned (verify your own number rather than repeating that roughly seventy-one percent of hospitals run predictive AI in the EHR), and pick one high-value, bounded use case where you can state the value, the failure modes, and where the human verifies.
• Days 31 to 60: define three linked metric sets before go-live: success (did it deliver value), safety (edit and caught-error rates), and equity (subgroup breakdown so a good average cannot hide a failing population), each with a threshold that triggers action.
• Build the funding story in advance with honest cost categories (license, integration, validation, the verification labor that never leaves, training, monitoring) and a defensible ROI measured from your own before-and-after baseline, not a vendor's headline, since every accuracy and ROI number is one to verify in your own population.
• Days 61 to 90: run a small, honest, gated pilot designed to surface the truth, not to succeed, with the verification workflow staffed from day one, and let the real evidence drive the decision to expand, adjust, or stop. When the safety or equity gate fails, the gate is working; pause and fix rather than moving the threshold, because gaming the gate turns a safety program into theater.
• Workforce literacy runs under all ninety days and never ends: a tool is only as safe as the people using it, and building the standing capability to understand, verify, and legitimately question the AI is what lets the next hundred use cases deploy safely. A pilot you are unwilling to halt is not a pilot, and transformation is not a leap but this loop run again and again.
• The iron rule that ran through every level is the whole discipline in one sentence: AI assists (use the tool), the clinician decides (accountability never transfers to the model), the record proves it (an undocumented decision is indistinguishable from a machine acting alone). Carry it into every deployment, and go build it.