Every AI strategy conversation I’ve been in this year has started with the same question: which model? Which vendor? Generative or predictive or agentic? Claude or GPT or Gemini? In-house or API? Fine-tuned or retrieval-augmented or both?
These are real questions. But they are not the first question. The first question is what you are going to feed the thing.
The Unsexy Truth About AI Performance
A well-understood pattern in applied machine learning is that the fastest path to better model performance is rarely a better model. It is better data. Cleaner labels. Fresher signals. Wider coverage. More structure. The model gets most of the credit in the press release, but the data is doing most of the work.
This is not news to data scientists. It is, in my experience, genuinely news to most executives making AI investment decisions. The pitch decks are about the models because the models are what vendors are selling. The data is assumed.
That assumption is tolerable in low-stakes environments where the downside of a bad output is an awkward email or a mistaken product recommendation. In compliance, it is not.
An AI model is only as intelligent as the data beneath it will allow. In compliance, that data layer is the real product — the model is the delivery mechanism.
Why Compliance Is Uniquely Unforgiving
A marketing team using AI to draft social posts can absorb a wide range of output quality. A weak draft gets edited or discarded. No one gets fined.
A sanctions screening system using AI to evaluate whether a beneficial owner appears on an OFAC list cannot absorb that range. A missed hit can result in a seven-figure enforcement action, a consent order, and in some cases a criminal referral. A flood of false positives cripples operations with alert fatigue and backlogs. The AI’s behavior — whatever architecture it runs on — is bounded by the data it sees.
If the PEP list the model screens against is six months stale, the model will miss newly designated individuals no matter how sophisticated its matching logic. If the adverse media feed is dominated by low-quality syndicated aggregation, the model will produce summaries that read as authoritative and are actually wrong. If beneficial ownership data is incomplete in a jurisdiction that matters to the portfolio, the model cannot reason its way out of the gap.
Five Dimensions That Actually Define “Good Data”
When I evaluate a compliance data layer — our own or anyone else’s — five dimensions matter more than all the others combined.
Coverage. Does the data include the people, entities, jurisdictions, and list types you actually need? “Global PEP coverage” means very different things across vendors. A feed that captures cabinet-level officials in G20 countries is not the same as one that captures sub-national politicians in emerging markets. Whatever the model does downstream, it cannot flag what it never sees.
Freshness. How quickly do changes in the real world propagate into the data? OFAC updates, new sanctions designations, and news of enforcement or conviction move constantly. A data layer that refreshes weekly is a liability in a world where a designation published Tuesday creates exposure by Wednesday morning.
Structure. Is the data atomic, normalized, and linkable — or is it a pile of text documents? Structure is what allows models to reason across entities, resolve aliases, connect family and corporate networks, and distinguish the two John Smiths. Unstructured data can be processed, but every processing step introduces uncertainty the model cannot undo.
Provenance. Can every data point be traced back to its original source? Regulators increasingly expect this, and examiners are learning to ask. A screening decision that cannot be defended on the record — “the AI said so” — is not a defensible program. Provenance is not a nice-to-have; it is the audit trail.
Governance. Who decides what goes in, what comes out, how disputes are adjudicated, and how errors are remediated? A data layer without governance is a data layer with unknown defects. At scale, unknown defects become regulatory findings.
The Hidden Costs of a Weak Data Layer
The cost of weak data in a compliance AI deployment rarely shows up as a single catastrophic failure. It shows up as a pattern. False positives climb, and analysts burn out triaging noise. Real hits get buried in the noise and missed. The program looks productive in dashboards while being functionally compromised. The cost of running the AI goes up — more compute applied to worse inputs — while the value goes down. Eventually, regulatory examiners start asking uncomfortable questions about testing and remediation.
None of these are model problems. All of them are data problems that the model inherits.
Questions to Ask Before the Model Questions
Before you evaluate an AI vendor, evaluate the data layer. Five questions separate serious compliance platforms from decorative ones.
What sources feed this system, and how are they verified? The answer should name specific authorities, publication cadences, and verification workflows — not a vague reference to “global sources.”
What is the end-to-end latency from a real-world event to a screening decision? Minutes and hours are different products from days and weeks.
How is the data structured, and which entities are resolved and linked? Entity resolution is expensive to build and impossible to retrofit cleanly.
Can every decision be traced to its underlying evidence? If not, the program is not auditable.
Who governs the data, and what is their change-management process? A vendor that cannot describe change management does not have any.
If the vendor cannot answer these confidently, the AI on top is decoration.
Conclusion
The competitive advantage in compliance AI is not going to come from having the best model. Model capabilities are rapidly becoming a commodity. Within eighteen months, most serious vendors will have access to approximately the same frontier models.
The advantage will come from what the model is allowed to see — how fresh it is, how structured it is, how complete it is, and how defensible it is. Models change every quarter. The data discipline endures.
If you are planning AI investment for 2026 or 2027, spend less time on the engine and more time on the fuel. That is where the performance actually lives.