Overview
.svg){kind=small}
One AI
Data Mesh
.svg){kind=icon}
One AI Assistant
Data Stories
Enterprise AI
One AI Data Intelligence
Employee Engagement
.svg){kind=small}
Talent Retention
-1.svg){kind=small}
HR Dashboards
Succession Planning
.svg){kind=small}
Talent Acquisition
Workforce Analytics
Why One Model
-2.svg){kind=small}
Whitepapers
Blog
-2.svg){kind=icon}
Events & Webinars
.svg){kind=icon}
Product Innovation Blog
-1.svg){kind=small}
Roles in People Analytics
.svg){kind=small}
Contact Us
-3.svg){kind=small}
Partners
About US
-3.svg){kind=icon}
Careers
-1.svg){kind=icon}
News
Since its inception, enterprise leaders have been tempted by the siren song of plug-and-play AI, the idea that you can just point an LLM at your raw, sprawling data warehouse and let it answer complex business questions. We’ve repeatedly cautioned against this and championed a foundational truth: Without a governed data foundation, AI answers are impossible to trust.
Anthropic, the AI giant behind Claude, just validated this point beautifully. In June 2026, they detailed how they enabled a self-service analytics engine using Claude. Their experience underscored what One Model has been saying all along: before they built that robust data foundation, their AI made assumptions, answered without context, and barely cracked a 20% accuracy rate. After they built it, their AI reached 95% accuracy.
Here, we’ll explore why they needed that data foundation and how One Model offers the same thing for orgs who prefer not to build it themselves. Finally, we’ll unpack why even 95% isn’t good enough for executive decision-making, and what it actually takes to tackle that 5% of residual error.
But First, What Is a Governed Data Foundation?
At its core, a governed data foundation is a bridge between your raw data sources and the tools that use them, like AI agents, dashboards, and reports.
Instead of letting an AI agent or analyst query messy, disparate databases directly, a governed data foundation cleanses, structures, and unifies that data into a single source of truth. It defines the business rules, metrics, security parameters, and relationships once, ensuring that no matter who or what asks a question, the answer is calculated the exact same way every single time.
Think of it as the ultimate translation and safety layer for enterprise data. It generally consists of three core components:
- A Unified Data Layer: Standardizes organization, connects disparate systems, and handles the heavy lifting of data modeling and pipeline maintenance.
- A Semantic Layer: Translates complex data tables and cryptic field names into standardized, human-readable metrics (e.g., standardizing how "headcount," "turnover," or "active employee" are defined across the enterprise).
- Governance & Security Protocols: Enforces strict role-based access controls, auditing, and compliance standards directly at the data level so sensitive information never ends up where it shouldn't.
Without a governed data foundation, AI operates on wild assumptions and raw chaos. With one, AI gains the context, boundaries, and precision it needs to deliver answers you can actually take to the C-suite.
Data foundations don't just happen. They have to built or bought.
Their Story: How Anthropic Achieved 95% AI Accuracy With a Custom Built Data Foundation
Anthropic wanted to empower their team with self-service analytics, hoping to free their data scientists from repetitive internal requests. When you remove data scientist supervision, you need a tool that produces unquestionably accurate results.
They knew it was more complicated than just “pointing Claude at a [data] warehouse and letting the agents execute,”–which, as they wrote, “can create a false sense of precision.” But why can’t AI query raw data? Because AI fundamentally lacks context on the data it’s accessing, and without context, it guesses and invents. In analytics, questions typically have just one correct answer, which the AI must be diligently shepherded toward.
To accomplish that, they had to build a governed data foundation and semantic layer–the very foundation that One Model has always championed as essential to AI success. Their complex build consisted of a custom agentic data stack powered by Claude. Once they built the data foundation and semantic layer, AI accuracy exploded from 21% to 95%. However, most organizations simply don’t have the time or resources for this kind of mammoth in-house build, even when armed with Claude or similar AI tools.
Their experience underscored a simple truth: you can’t make AI accurate or trustworthy without a data foundation. And if organizations want to avoid the intense effort of building and maintaining a data foundation, they can simply purchase it from One Model.
Not So Happily Ever After: The Hidden Downside of DIYing a Data Layer
Anthropic achieved 95% accuracy, but maintaining that victory takes rigorous upkeep. Without active maintenance, Anthropic watched Claude's 95% accuracy sink to 65% in just one month.
This is not a skill issue; Anthropic possesses world-class AI and data engineering capabilities. It’s simply the inevitability of dynamic, changing organizations. Business logic changes. The data model changes. Their build has to adapt with them, or it stagnates and becomes unusable. That means continuous auditing of their system and rules.
You could say that relentless updates are to data layers what oil is to automobiles. If you can’t dedicate a team to maintaining data pipelines, you can’t maintain 95% AI accuracy long term. For most companies, the immense investment required to achieve that keeps them focused on data plumbing, rather than insight mining.
The fluidity of the company and data environment mean that the data layer is always staggering to keep up with it. That burden will always fall on the data engineers who undertake in-house builds. Companies who choose to buy these features with One Model (along with the expert team, platform maintenance, and technology that comes with it) don't have to worry about the continuous upkeep because their workforce data is transformed into a governed, decision-ready foundation before AI is ever introduced.
Why Does AI Need a Governed Data Foundation?
It bears asking: why does an industry-leading LLM fail nearly 80% of the time without a data foundation? Anthropic identified 3 fundamental errors inherent to what we’ll call “raw” AI analytics (that is to say, pointing AI at a data warehouse with zero data foundation).
1. Concept Ambiguity: The AI Doesn’t Know Know What You Mean
Raw data and query language are rife with ambiguity. So the AI fills in the gaps and gets “creative.” But creativity, in this environment, fosters inaccuracy. As Anthropic writes, “With hundreds of viable options in a data model (out of potentially millions of fields), the agent is unable to choose the correct fields that best answer a user’s question. For example, in measuring the number of active users: what actions constitute being “active”? Do you include fraudulent users? What lookback window do you use?”
Consider the “Coffee Shop Analogy.” The same principle applies when you walk into a coffee shop and say, "I'd like a coffee." The barista won't just turn around and hand you a cup–they will ask: “Cappuccino? Latte? Black?” Vague inputs demand clarification, but LLMs don't always pause to ask. (Credit: Retinder Labana, LinkedIn)
One Model’s Solution: Rather than forcing an AI to guess which fields to join or how to define a metric on the fly, One Model codifies your business logic directly into a governed data foundation. When the AI agent queries the system, it simply calls One Model's pre-defined, mathematically bounded metrics and guardrails (e.g., locking in precise logic for headcount, turnover, or active users). This kind of precision at the input guarantees reliability at the output.
2. Data Staleness: The Data No Longer Reflects Your Current Organization
Businesses are dynamic environments. Metric definitions change, schemas shift, and underlying business rules evolve continuously. Without active, explicit maintenance, an AI’s knowledge base rots over time. Yesterday’s right answer silently turns into today’s wrong one.
One Model’s Solution: One Model acts as an active, living semantic layer that insulates AI from underlying schema drift and operational changes. When business logic evolves, like a new department structure or updated turnover definitions, the logic is updated once within One Model's centralized platform. Instead of requiring data engineers to constantly rewrite prompt templates or custom skills, One Model maintains a continuous, audit-ready source of truth.
3. Retrieval Failure: The AI Can’t Find What It’s Looking For
When AI has to search a million-field warehouse for one metric, it may simply fail to find it. Anthropic found that even when a metric was properly annotated, failure could still occur.
One Model’s Solution: One Model tackles retrieval failure structurally, not by better prompting: it collapses the search space to a curated, admin-approved set of governed metrics and dimensions, resolves a natural-language question to those governed entities via its query engine and a vector database, and enforces definitions plus role-based security on every query (including through its MCP server into Claude/Copilot/ChatGPT).
The Key Takeaway: AI accuracy always comes down to the structure of the data behind it. You have to treat your workforce data like raw material that needs to be prepared, structured, and curated, not a ready-to-use source of truth that your AI can read directly.
The Deeper Problem: Why 95% Accuracy Isn't Good Enough
Even if you were to build something similar to Anthropic’s agentic data stack, assume a heroic level of ongoing maintenance, and prevent performance atrophy, there is still a fundamental flaw to be reckoned with: the missing 5 percentage points of accuracy.
As One Model CEO Chris Butler puts it, "95% accuracy is not a win. It means that one in twenty numbers is wrong, with no way to know which."
In executive decision-making, a 5% error rate is fatal to organizational trust. To quote data professional Jakob Berglund, "The best way to ruin a company running on data is not to delete the data, but instead fill it with small errors that are nearly impossible to catch. That makes the data completely unusable in terms of trust and reliability."
Anthropic overcame inaccurate AI with their bespoke agentic build. It included a proprietary data foundation, narrow sources of truth, AI Skills, and meticulous ongoing validation, skyrocketing them to 95% accuracy.
One Model helps organizations like yours do the same, without having to do it yourself. The hard truth is that no one gets to 100%. But One Model attacks the structural sources of error, while showing its work and ensuring human sign-off–converting any remaining error from silent and imperceptible into visible and traceable.
Defensible Accuracy Starts with One Model
Anthropic’s experience is a victory for the data community because it dispels the myth that AI models alone can solve enterprise analytics. Accuracy is a context and governance problem, not an LLM problem. A governed data layer solves that.
Building a data layer internally requires endless data engineering hours, continuous evaluation pipelines, and constant human oversight just to keep the system from falling apart.
At One Model, we deliver the exact same governed architecture they painstakingly engineered, but pre-built and purpose-made for the world of sensitive HR data. By centralizing, modeling, and governing your workforce data in a dedicated layer, we ensure that every AI agent, dashboard, and user works from the same governed definitions and traceable logic, so when a number gets challenged, there's an answer.
