Your Business Needs One Shared Model of Itself

The Decision in Front of You

Most companies have spent heavily on data in the past decade: a warehouse or lakehouse, a reporting tool, an integration platform, and now AI pilots. Each of those purchases came with a project, and each project defined the business's core objects for itself.

The decision here is whether to keep paying for meaning one project at a time, or to fund a single model of the business that everything shares. That model is usually called an ontology. In plain terms, an ontology is a written, machine-readable agreement about what the business's things are (customer, order, asset, site, contract), how they connect to each other, and what may be done to them. The software that holds those definitions between the source systems and the tools that read them is called a semantic layer.

Two earlier papers in this series cover preparing data for one AI agent and funding AI work in the right order; this paper concerns the model the whole business shares, whether or not AI is involved.

What's Happening

Leaders do not trust their own numbers

The Precisely and Drexel University LeBow College of Business survey of more than 550 data and analytics professionals, published in November 2024, found 67% did not completely trust the data they use for decision-making, up from 55% the year before ¹. Data quality was the top data integrity challenge, cited by 64% ¹. Salesforce's State of Data and Analytics report, drawn from nearly 8,000 executives, found data and analytics leaders estimate 26% of their data is untrustworthy ². Both firms sell data software.

Ownership is the recurring gap. dbt Labs' 2026 State of Analytics Engineering report, a survey of 363 data practitioners and leaders published in April 2026, found 41% named ambiguous data ownership as a governance obstacle, and 71% named incorrect or hallucinated outputs reaching stakeholders as a top concern ³.

Regulators now expect a dictionary of the business

The Basel Committee's principles for risk data aggregation, known as BCBS 239, expect banks to produce accurate, consistent risk numbers across the whole group ⁴. In May 2024 the European Central Bank published a guide on how it expects banks to meet them ⁴. The ECB reported that in its review of 25 significant banks, none had fully followed the principles, including the global systemically important ones, and it judged progress since then "generally insufficient" ⁴.

The guide asks for a dictionary of the main business definitions and a metadata repository, uniform definitions with clear ownership, validation rules, and lineage, meaning a record of where each figure came from and how it was transformed, down to individual data attributes ⁴. Outside banking, that list is a fair description of a shared model.

The tools have converged on the same idea

The World Wide Web Consortium's Web Ontology Language, OWL 2, became a W3C Recommendation in 2009, with a second edition in December 2012 ⁵. It describes classes of things, their properties and the individual instances, and it is exchanged as RDF, the W3C's standard format for linked data ⁵.

dbt Labs' semantic layer lets data teams define a metric such as revenue once and have every downstream tool use that definition; when the definition changes, it changes everywhere it is used ⁶. Palantir describes its ontology as an operational layer that maps existing data sources into objects, properties and links, and adds action types that capture decisions and write them back to existing systems ⁷. A model that also defines allowed actions is what an integration or agent needs before it changes anything.

In September 2025, Snowflake, Salesforce, dbt Labs, BlackRock and RelationalAI launched the Open Semantic Interchange, with a dozen partners, to create a vendor-neutral format for semantic definitions ⁸. In January 2026 dbt Labs reported the first version of the specification was published under the Apache 2.0 license, covering datasets, metrics, dimensions and relationships, with Databricks also named as a partner ⁹.

Gartner placed the rising importance of semantic modeling among its top data and analytics trends for 2025 ¹⁰. According to TechTarget's coverage, Gartner predicted that organizations prioritizing semantic modeling would improve the accuracy of their AI tools by 80% and cut the cost of building and maintaining them by 60% ¹⁰. Treat those as analyst predictions without published measurement behind them.

AI raises the cost of disagreement

Spider 2.0, a benchmark of 632 real enterprise database tasks presented at ICLR 2025, found that an agent built on OpenAI's o1-preview model solved 17.0% of tasks, against 91.2% on the older, simpler Spider 1.0 benchmark ¹¹. Enterprise schemas are large and cryptic, and the business meaning lives nowhere in them, so an AI tool reading raw tables guesses what "active" or "shipped" means.

Two Cases

Citigroup, 2024. In 2020 the Office of the Comptroller of the Currency fined Citi $400 million for shortcomings in risk management, data governance and internal controls, and the Federal Reserve ordered it to fix its data quality practices ¹². On July 10, 2024, the Federal Reserve fined Citigroup $60.6 million for insufficient progress remediating its data quality management problems and for lacking controls to manage the risk those problems created ¹³. The OCC added about $75 million, for a combined total near $136 million ¹²¹³. Four years of remediation had not made the bank's data consistent. Fixing definitions after decades of separate systems takes years, so the work should start before a regulator sets the deadline.

WM Technology, 2024. WM Technology, the parent of the Weedmaps cannabis marketplace, reported "monthly active users" as a headline operating metric, and its SEC filings described those users as people who had engaged with its site ¹⁴. On September 24, 2024, the SEC charged the company and its former CEO and CFO with negligent misrepresentations about the metric for May 2021 to May 2022, a period that included the company's listing through a SPAC merger ¹⁴. According to the SEC, a large and growing share of the counted users had been shown the site through pop-under advertisements on other websites, and most of them never clicked or engaged ¹⁴. Reported users grew while engagement was flat or falling ¹⁴. The company agreed to a $1.5 million penalty, and the two executives paid $175,000 each and accepted three-year bars from serving as public company officers, without admitting or denying the allegations ¹⁴.

The written definition said one thing and the counting rule in the systems did another, and nobody reconciled the two before the number went to investors. In a shared model, "active user" carries one definition with an owner, and the mapping states exactly which traffic counts, so a change in what gets counted becomes visible as a change to the definition.

Why the Usual Approach Fails

Definitions are written where the work is funded. A dashboard project needs "revenue," so the analyst writes a rule. An integration project needs "order," so the developer picks a status code. An AI pilot needs "customer," so someone writes it into a prompt. Each rule is reasonable, and each lives where only its builder can see it. After a few years a mid-size company has dozens of definitions of its five most important nouns.

Consolidating systems leaves the problem in place. Reports, integrations and AI tools still sit outside the new ERP, and each still needs its own logic for what its fields mean.

Large enterprise data models failed in a different way. IT teams modeled the whole business before delivering anything, the business never signed off, and the model aged before it was used.

Tool vendors each ship their own semantic layer. Buying three tools gives the company three places to define "customer" that do not talk to each other, the problem the Open Semantic Interchange was formed to address ⁸.

Nobody is accountable for meaning. The definition of an active customer or a completed order usually belongs to no one, so disputes get settled meeting by meeting and never written down ³.

The Framework: One Model, Four Parts

flowchart LR
    A[Source systems<br/>ERP, CRM,<br/>maintenance, billing] --> B[Mappings<br/>fields and rules<br/>per system]
    B --> C[Shared model<br/>objects, relationships,<br/>actions, owners]
    C --> D[Reports and<br/>dashboards]
    C --> E[Integrations<br/>between systems]
    C --> F[AI agents<br/>read and act<br/>through the model]

Five rules make the model work in practice.

The business owns the definitions. Each core object has a named executive owner who settles disputes about its meaning. The COO is often the right sponsor, because operations crosses every department that argues about definitions.

Start with the objects that carry money. Most companies can name the five to ten objects behind most of their revenue, cost and risk: customer, order, product, asset, site, contract, employee. Model those first and stop.

New work must use the model. Every new report, integration and AI agent reads core objects through the shared model, and the architecture review rejects anything that defines them privately.

Actions are part of the model. Agents and integrations change things, so the allowed actions, the roles that may take them, and the checks that apply belong in the same place as the definitions ⁷.

Keep the model portable. Store definitions in an open format, such as OWL and RDF or the Open Semantic Interchange specification, so the company can change vendors without rewriting it ⁵⁹.

Costs, Risks, and Tradeoffs

The model costs executive time more than software. Settling what "customer" means requires a decision from someone with authority over the departments that disagree, and that decision has political cost.

How to Act

Questions to Ask Your Team

1. How many different definitions of "customer" are in use across our reports, integrations and AI tools today?
2. Who has the authority to decide what an active customer or a completed order means, and has anyone used it?
3. Which number did we last argue about in an executive meeting, and how long did it take to settle?
4. Which actions can our integrations and AI agents take on customer, order or asset records, and where are those rules recorded?
5. If we changed our reporting or AI vendor next year, would our definitions come with us?
6. What would a regulator or auditor find if they asked for our dictionary of business definitions and the lineage of our top ten figures?

Closing

Every company already pays for its definitions. It pays in each project that rewrites them, in each meeting that reconciles them, and in each AI tool that guesses at them. Fund one shared model of the business instead: a small set of core objects, their relationships and allowed actions, mapped to the source systems, owned by the executives accountable for the numbers, and used by every report, integration and agent that comes after.