The AI Deployment Lifecycle: AIDLC
By Prachi Mondaiyka, Deputy Chief Information Officer| Chief AI Strategy Officer, New York State Department of Health
CAIOs playbook built from what they ‘already’ know
Over the past five years, the Chief AI Officer (CAIO) role has evolved from AI evangelists promoting AI to leaders driving transformation with AI and moving it from pilots to enterprise-wide implementation. As the technology matured, so did the role. Today’s CAIO owns AI strategy, governance, and adoption end to end. Yet most CAIOs are promoted technologists, data leaders, or portfolio managers suddenly tasked with leading an enterprise-wide AI initiative, and then very quickly they have been asked to produce “the AI roadmap,” assuming such a thing already exists in a drawer somewhere, fully formed. It doesn’t. So where do these leaders turn, and can the disciplines they’ve practiced for years be repurposed into a new blueprint?
Unlike the digital transformation programs of the last decade, AI adoption doesn’t fit neatly inside a single discipline’s playbook; it sits at the crossroads of phase-gated rigor of the software development lifecycle (SDLC), the strategic framing of digital transformation, and the governance discipline of classic project management. And it adds a wrinkle none of those frameworks were built for: a model that keeps changing after it is built or bought and deployed.
A traditional software release, once tested and deployed, is done until the next planned iteration. An AI model is never quite done. Models drift as the real world diverges from training data, degrade as user behavior shifts, and require ongoing retraining just to hold their day-one accuracy.
So, here’s: AIDLC, the AI Deployment Lifecycle that borrows the best of SDLC and PM thinking while accounting for AI’s defining trait.
AIDLC (AI Deployment Lifecycle): an eight-stage model for AI adoption that borrows the phase-gate discipline of SDLC, the strategic framing of digital transformation, and the governance rigor of project management, built around one fact: an AI model, unlike software, is never truly “done.”
For CAIOs, CIOs, and digital leaders, that may be the difference between deploying AI—and building an enterprise that can continuously create value from it.
Stage 1: Discover and Envision
The opening stage looks much like a conventional transformation kickoff: build the business case. What problem is AI solving? What opportunity is on the table? Who owns the outcome, and what does success look like? What’s the ROI?
Stage 2: Strategize and Plan
Next comes governance: who sponsors and approves use cases, which guardrails (ethics, data privacy, cybersecurity, legal) apply, how regulatory requirements are addressed, and what level of rigor is required for monitoring and observability. The PM instinct to establish a clear charter before scope creep takes hold is essential. Because of the glamor of AI on the surface, AI pilots are notoriously prone to scope inflation, once stakeholders see a working demo; a charter with defined success metrics is the best defense against a pilot that never ends.
Stage 3: Data Readiness
This is the stage most initiatives quietly stall on. Clean, integrated data is the foundation of good decision-making even without AI; its absence is only amplified once AI enters the picture. Readiness means auditing sources, checking quality, resolving fragmentation across legacy systems, and building the pipelines that will feed a model. Skip this step, and organizations train sophisticated systems on unreliable inputs, a mistake that surfaces months later as poor performance and costs far more to fix retroactively than to prevent upfront. Lean in heavily into the data stewards performing true data governance.
Stage 4: Design and Prototype
Here AIDLC takes on its clearest SDLC flavor. Architecture decisions get made: build versus buy, which model family fits the use case, and how the system integrates with existing infrastructure and workflows. Once decided, pilots run in select environments against clean quality datasets.
Stage 5: Build and Train
When most people hear the term “AI project,” they immediately think of development and training, but this is only one link in an eight-stage chain. The SDLC parallel is useful here: version control for models and datasets, documented
experiment tracking, and a defined “done” for a training run not “until the demo looks good,” but against measurable criteria set back in stage one.
Stage 6: Test and Validate
AI Quality assurance carries everything traditional QA does, plus a category SDLC never had to worry about: bias and fairness testing. A model can pass every functional test and still produce systematically skewed outcomes for certain populations. In the public sector or a regulated industry, this stage is the difference between a defensible deployment and a future headline.
Stage 7: Deploy and Integrate
This is where change-management discipline earns its keep. A technically sound model that nobody trusts or knows how to use is a failed deployment regardless of its accuracy scores. This stage includes the technical work of release and integration, but the harder work is often organizational: training end users, adjusting workflows, and managing the very human resistance that comes with any system perceived as replacing human judgment.
Stage 8: Monitor, Govern, and Retrain
This stage has no clean equivalent in classic SDLC or digital transformation, and it’s the reason the AIDLC loops rather than ends. MLOps monitoring tracks production performance and watches for drift as real-world data shifts away from training assumptions. Governance tracks whether the model is still used as intended and whether new risks have emerged. Retraining, feeding the model updated data and re-validating it, happens on a cadence set by how fast the environment changes, not by a fixed release calendar. Crucially, this stage feeds back into stage four, not into a backlog. A model isn’t retired and replaced the way a legacy application eventually is; it’s continuously refreshed. The organization’s operating model has to budget for permanent capacity not just for a one-time project team.
Why the Loop Matters More Than the List
The eight stages are useful only if organizations resist treating them as a one-time checklist. The real value of AIDLC is the operating rhythm it creates. This modifies the main question for CAIOs. Rather than inquiring, “When will this AI project be finished?” leaders should ask, “Do we have the operating model, governance, data, talent, and funding required to keep this system trustworthy and valuable over time?”
That is the strategic shift AIDLC is designed to create.
It’s not another AI checklist. It is a management framework for making AI adoption repeatable, governable, and sustainable. For CAIOs, CIOs, and digital leaders, that may be the difference between deploying AI—and building an enterprise that can continuously create value from it.
Note: Generative AI was used as an editorial support tool to assist with organization and language refinement. The ideas, AIDLC framework, analysis, and conclusions presented in this article are the author’s own
About the Author:
Chief AI Strategy Officer, Deputy CIO: NYS Dept. Of Health
Prachi leads the AI governance and the digital strategy for the NYS Department of Health (NYS-DOH). She previously served as the Director of Strategy and Digital Transformation at NYS Housing and Community Renewal (HCR) and held various technology development and portfolio management roles across NYS agencies. Prachi is recognized for her expertise in leading strategic initiatives, managing large-scale service delivery, stakeholder relationship management, organizational change management, and driving digital transformation across diverse organizations. Prachi holds a bachelor’s degree in engineering from BIT, India, and a master’s in management from Rensselaer Polytechnic Institute (RPI), NY