Most enterprise AI roadmaps I review are a list of use cases with timelines attached. "Q2: Deploy customer service chatbot. Q3: Automate underwriting. Q4: Implement predictive churn model." The roadmap looks complete. It has dates. It has owners. It has executive buy-in.
What it usually doesn't have is any mention of the data that would make those use cases work. And that omission is why so many AI roadmaps slip from quarter to quarter until someone quietly moves them off the OKR list.
The Data Dependency Nobody Audited
Every AI use case has a data dependency. That dependency has three components: does the data exist, is it accessible, and is it clean enough to be useful?
These three questions are independent. Data that exists but lives in a decommissioned system is inaccessible. Data that exists and is accessible but contains three years of entry errors, inconsistent formatting, and missing values isn't useful for training or retrieval. The use case that looked straightforward on a roadmap slide can fail at any one of these checkpoints.
The predictive churn model needs eighteen months of labeled behavioral data. Where does that come from? The customer service chatbot needs a knowledge base of accurate, current policy information. Who owns that content, how often is it updated, and in what format does it exist? The underwriting automation needs structured data from documents that are currently processed manually. What does the document extraction pipeline look like, and who builds it?
These are not edge case questions. They are the central questions. And most AI roadmaps don't answer them because the people building the roadmap didn't ask them.
The Three-Layer Data Stack
Before committing to an AI roadmap, I push teams to audit their data stack against three layers:
Raw data availability: What data do we actually have? Where does it live? What are the access controls and data governance policies? Is there a data catalog? This is a discovery exercise, and it frequently reveals that the organization has much more (or much less) data than leadership assumed.
Data pipeline readiness: Can the data get from where it lives to where the AI system needs it, in the format the system requires, at the latency the application demands? A batch pipeline that refreshes daily works for some use cases and fails completely for others. Real-time feature serving is architecturally different from weekly model retraining. The pipeline question is an engineering question, not a data science question, and it's often overlooked in roadmap planning.
Data quality: Is the data accurate enough to train on or retrieve from? Quality has multiple dimensions: completeness (are the fields populated?), consistency (is the same concept represented the same way across records?), timeliness (is the data current enough for the use case?), and correctness (is the data actually right?). Quality issues compound in ML pipelines — a model trained on garbage data doesn't learn bad patterns, it learns garbage.
If you can't answer all three layers for a given use case, the use case isn't ready to be on the roadmap. It belongs in a data readiness backlog.
What a Honest AI Roadmap Looks Like
An AI roadmap that's grounded in data reality has a different structure. It starts with a data inventory: what datasets do we have, what's their quality, and what use cases could they support? It identifies data gaps and places data engineering work on the roadmap explicitly — not as prerequisites that happen offscreen, but as first-class deliverables.
It sequences AI use cases based on data readiness, not business excitement. The use case with clean, accessible, well-labeled data goes first — even if it's less strategically exciting than the use case that requires eighteen months of data cleaning. Shipping something small and real builds the organizational muscle and executive confidence that funds the harder work.
It also has an explicit feedback loop: what data does this AI system generate that we can use to improve it over time? Production AI systems that log their inputs, outputs, and errors can be continuously improved. Systems built without a data feedback loop are static — they perform as well on day one as they ever will.
The Conversation You Need to Have
If you're a technology leader trying to drive an AI roadmap forward, the most useful conversation you can have with your data team is not "what can AI do for us?" It's "what data do we have that's ready, and what use cases does that data support?"
Start there. The answer will constrain your roadmap in ways that feel uncomfortable in the short term and save you enormous amounts of time and credibility in the long term. An AI roadmap grounded in data reality is a roadmap that can actually be executed. And in AI, executed beats ambitious every time.
