Part 8 of 9 — Data Foundation
At Aurelia, the institutional knowledge that made the company work lived in roughly 14,000 documents that no system formally owned. Show binders at the parks with hand-annotated cueing notes for every parade, fireworks show, and stage performance. Personal spreadsheets at the resorts tracking VIP guest preferences — which celebrity guest preferred which suite, which annual passholder family had a child with a peanut allergy and a fascination with a specific character, which corporate event planner needed the same set of accommodations every year. Production binders at the studios with talent quirks, location vendor relationships, and the unwritten rules of how a particular director liked his post-production handoffs. Continuity bibles at Aurelia Live tracking costume, choreography, and song variations across cities. Cruise itinerary lore at Aurelia Voyages documenting which port had which berth issues at which tides.
When Aurelia tried to build agents on top of that data, the agents hallucinated about 1 in 4 guest interactions. The fix was not a better model. The fix was the boring, expensive work of fixing the data.
This is the focus area where the most resistance lives. Strategy is glamorous. Operating models are necessary. Platforms are tangible. Data foundation work is unglamorous, expensive, slow, and rarely the thing executives want to fund. It is also, in my experience, the single highest-leverage investment in the entire agentic transformation. Every other focus area is more effective when the data foundation is solid. Every other focus area is undermined when it is not.
This article is about the five capabilities that make up the Data Foundation focus area, and the work Aurelia did that nearly didn't get funded but ended up being the work that made everything else possible.
The Five Capabilities of Data Foundation
Five capabilities, all foundational:
Knowledge Management — the organization of enterprise knowledge for agent reference, and the quality of that knowledge base for AI consumption (which is different from human consumption)
Data Availability — the breadth of data sources accessible to agents: real-time, historical, structured, unstructured, internal, external
Data Management & Governance — how critical data entities are managed across systems, plus the AI-specific data governance, security, and compliance overlay
Data Quality — the consistency, completeness, accuracy, freshness, and hygiene of data across the organization
Data Integration & Performance — the effectiveness of data integration, transformation, and platform scalability for the data volumes and latencies AI workloads require
These look like classic data engineering disciplines, and they are. The difference is that AI workloads put novel pressure on each of them. Hallucinations, retrieval failures, and inconsistent agent behavior almost always trace back to one of these five capabilities being undercooked.
What Level 1 Looks Like
Low-maturity Data Foundation has the most universal symptoms of any focus area:
Critical knowledge lives in show binders, email threads, and the heads of long-tenured cast members
"Data availability" means whatever the data engineering team has gotten around to integrating, on whatever timeline they got around to it
Master data management is something the company started in 2017, declared done in 2019, and quietly let drift
Data quality is measured in dashboards that nobody reads, with thresholds nobody enforces
Data integration uses an ETL pipeline built five years ago that runs nightly and falls over once a quarter
"Real-time" data is a phrase used aspirationally
When an agent hallucinates about a guest's preferences or a show's running order, the first response is to blame the model
If three of those describe your enterprise, your agents are not failing because the model is bad. They are failing because the substrate is bad.
Aurelia's Starting Point
The 14,000 documents were not a metaphor. They were a literal inventory across the company:
About 3,200 show running binders and operational notes across the four resorts
About 4,800 personal spreadsheets at the resorts and Voyages tracking VIP and frequent-guest preferences
About 2,400 production binders inside Aurelia Pictures and Lantern Animation containing talent, vendor, and location intelligence
About 1,600 continuity bibles across Aurelia Live productions
About 1,500 maintenance lore documents inside the parks' attractions engineering teams
About 500 schedule and ops continuity documents inside Aurelia Skies and Aurelia Voyages
The senior cast members used them. The new cast members did not know about them. The Imagineers used them. The IT department had no idea most of them existed. They were "personal productivity tools" — even though, collectively, they were the company's most important repository of operational and guest intelligence.
When the Guest Concierge agent was first deployed at Aurelia Kingdom, it could only access the structured data in the resort's reservation and guest profile systems, which was a small fraction of what the senior concierges actually relied on. The agent's responses were technically correct and operationally useless. It told a returning guest about restaurants instead of telling her that her usual table at her usual restaurant was being held the way it had been held every year for eight years. Cast members ignored it. Guest adoption was 7%.
The CDO commissioned a data foundation diagnostic. The findings were uncomfortable. The plan to fix it was an $18M, 16-month investment with no immediate visible output for the first 8 months. Funding it required real conviction.
The Big Rocks Aurelia Had to Move
Rock 1: The "tribal knowledge" migration
The most basic, most boring, most expensive rock. Fourteen thousand documents had to be inventoried, classified, and either retired, replaced, or migrated into systems that an agent could safely use. The work was done in three waves:
Wave 1: 5,200 documents identified as containing guest-relationship intelligence, migrated to a managed knowledge graph with structured entities (guests, parties, preferences, prior visits, dietary needs, accessibility needs, character interests) and semi-structured notes — with strict access controls aligned to the Four Zones data model
Wave 2: 5,800 documents identified as operational continuity tools (show binders, production binders, cruise itinerary lore, ride maintenance notes), digitized and structured into domain-specific knowledge stores with proper versioning
Wave 3: 3,000 documents identified as personal ad-hoc analysis, retired with the cast members retrained on sanctioned tools
The work took 13 months. It was unglamorous to the point of being almost invisible. Senior concierges, master Imagineers, and Live production stage managers complained constantly. Three different VPs tried to get the program paused. The CDO held the line.
By month 14, the Guest Concierge agent could access roughly 80% of the context that senior concierges actually used. Adoption climbed from 7% to 52% in the next quarter, and to 84% by month 18.
Rock 2: Master data management for AI
Traditional master data management is built for transactional systems. AI workloads need something different. Agents need to reason about entities — guests, characters, attractions, vessels, aircraft, productions, properties — across disparate sources, often probabilistically, often across spelling variations, name changes, and historical aliasing.
Aurelia invested in an AI-aware MDM layer that maintained: a canonical entity graph, a fuzzy match layer for resolution, a confidence score for every match, and a feedback loop where corrections from agent users improved the match quality over time. The entity coverage spanned guests (with their household groupings, frequent traveler IDs across BUs, and consent states), characters and IP (with rights and licensing context), attractions and shows (with their operational schedules and capacity profiles), and the company's hard assets (resorts, ships, aircraft).
This was the kind of plumbing investment that would not show up in any quarterly board update. It was also the difference between an agent that knew that "the Smith family from Ohio" who booked a Pacifica trip and "James and Margaret Smith, frequent guests" with the loyalty profile that had a peanut allergy on it were the same family — and an agent that did not. The hallucination rate is correlated almost perfectly with the quality of this layer.
Rock 3: Data quality SLAs by domain
You cannot manage data quality at the enterprise level. The standards are different for different domains, and the tradeoffs are different. Real-time park ride throughput data needs near-perfect accuracy and very low latency. Historical guest preference data needs completeness and provenance more than freshness. Maritime route data has its own freshness curve. Streaming engagement data has yet another.
Aurelia established data quality SLAs by domain: completeness, accuracy, freshness, consistency, and lineage targets, owned by named domain stewards. The stewards were not in IT. They were in the business — the heads of Resort Operations, the chief commercial officer, the head of Studios production, the head of Skies operations — given a quarterly target on data quality and given the budget and authority to hit it.
This was a significant cultural shift. Domain leaders had previously thought of data quality as IT's problem. After the shift, it was theirs, with measured outcomes attached to performance reviews. The data quality measured in the 17 most-used data domains improved by an average of 41% in the first year.
Rock 4: Knowledge management for agent consumption
Human-readable knowledge bases are not the same as agent-readable knowledge bases. A document written for a human can use implication, context, and judgment. An agent needs structured signals: entity references, claim provenance, recency markers, and explicit relationships.
Aurelia invested in a knowledge management overlay that took the company's roughly 8 million internal documents — including the migrated 14,000 — and produced agent-consumable representations: chunked, embedded, entity-tagged, source-linked, and freshness-stamped. The investment included a continuous curation process — about 18 FTEs across the COE and the embedded pods — whose job was to maintain the quality of this layer.
This was the investment most often questioned by the executive committee. It was also the one the agents needed most.
What "Leading" Looks Like
Eighteen months in, Aurelia's Data Foundation maturity looks like this:
The 14,000 tribal knowledge documents are gone, replaced by structured systems and a curated knowledge graph
The MDM layer maintains canonical entities for guests, parties, characters, IP, attractions, productions, and hard assets, with measured match quality
17 critical data domains have published SLAs and named domain stewards with quarterly accountability
The knowledge management overlay covers roughly 8 million documents, with a curation team of 18 FTEs maintaining quality
Agent hallucination rates are measured per-agent, per-domain, with weekly reporting; the Guest Concierge agent's hallucination rate has dropped from 25% at launch to under 4%
Real-time data integration is a discipline, with about 50 data flows running at sub-second latency for time-sensitive use cases (park ride wait times, flight ops, port operations, broadcast traffic)
Data integration costs have dropped by about 25% even as data volumes have grown 5x
The shift here is not just technical. It is organizational. Data quality has stopped being the data team's problem and started being everybody's problem, with named owners and measured outcomes.
The RPI at the Other End
The cast members and guests at Aurelia did not care about the knowledge graph. They did not care about the MDM layer. The senior concierge cared about whether the agent would tell a returning guest the right thing about her usual table. The Live stage manager cared about whether the agent could surface the right cueing variation for tonight's market. The frequent flier on Aurelia Skies cared about whether the agent knew she preferred aisle seats and didn't ask her every time.
The day the agent stopped being wrong about those things, cast members stopped fighting it.
Adoption tells the story. The same Guest Concierge agent, with the same model, with the same prompts, went from 7% adoption to 84% adoption among guests at the flagship park — driven almost entirely by data foundation work. The model was not the bottleneck. The substrate was. When the substrate got fixed, the agent stopped feeling like a corporate gimmick and started feeling like a colleague who had read the same notes the senior concierge had been keeping for years.
That is the Data Foundation RPI signal. Not data quality scores. Whether the humans on the other end have stopped checking the agent's work because they trust it.
What to Do This Week
If you want to test where your Data Foundation maturity sits, run this diagnostic:
Pick the most important agent in your portfolio. Have a domain expert independently review 50 of its recent responses for accuracy. Calculate the error rate.
For the errors, do a root cause analysis. What percentage of errors trace to model issues? What percentage trace to the underlying data being missing, inconsistent, stale, or wrong?
Look at where your data investment is going. Is it proportional to where the errors are coming from?
In almost every assessment I have done, the answer to question 3 is no. Most companies overinvest in models and underinvest in data. The maturity gap usually closes itself once that imbalance is corrected.
Next article: the recap. What Aurelia taught me about the pattern that holds across all seven focus areas, and the meta-lessons I keep coming back to.
This is Part 8 of a 9-part series on agentic enterprise maturity.