The foundation: engage.re
AI and Data 30 July 2026 11 min read

Why Does AI Get Your Own Business Data Wrong?

Your AI tool answers questions about the world accurately. It then gives you a wrong figure about your own records, and it gives that figure confidently. Your database stores the structure of your data, and it does not store the meaning. The model must therefore invent the business rules that nobody wrote down. This article gives the measurements, and it explains what closes the gap.

Connect with us about your data foundation · · Replies within 24 hours

21.3%
Accuracy of a leading model on real enterprise schemas, from 91.2%
~0%
Accuracy of GPT-4o on a benchmark built from real data warehouses
+17 to +23
Percentage points gained when the model receives the meaning of the data

Why does AI get your own business data wrong?

Your schema holds a table and a column. It does not hold the conventions that your organisation applies to them. The model reads the column name, and it then guesses the rule. A guess returns a number rather than an error, so nobody sees a warning.

One word shows the problem. A council uses the word "resident". That word can mean a person on the electoral roll, or a person in a council property, or a person with an open case. Each meaning gives a different count, and each count is defensible. A care home group uses the same word for a person in a bed, and it must also decide whether a respite stay counts.

The model therefore does two jobs at once. It translates your question into a query, and it invents the rules that your schema never recorded. It does the first job well. It cannot do the second job at all.

What do the benchmarks show?

Models score above 85% on academic datasets. Their accuracy then falls on real enterprise data, and the fall is large.

Benchmark Data used Result
Spider Academic schemas Above 85% execution accuracy
Spider 2.0 Real enterprise schemas, hundreds to thousands of columns A leading model falls from 91.2% to 21.3%
BEAVER Actual enterprise data warehouses GPT-4o reaches close to 0% end to end
LogicCat Questions that need external knowledge The best models reach 33.2% on complex queries

The second row uses the same model on the same task. Accuracy falls by seventy points when the schema becomes a real one. Spider 2.0 exists to measure that fall.

A real schema carries more columns, more history and more local convention than an academic one. The model has more to guess about, so it guesses more often.

What happens when the model receives the meaning?

Accuracy rises sharply, and it rises by a similar amount for every model. A 2026 benchmark tested this directly. Three frontier models answered 99 business questions on a public retail dataset, twice each.

The first run received the database schema alone. The second run added a document of 4 KB. That document described the measures, the conventions and the rules for ambiguous terms. It added no data and no code.

Model Schema only Schema and meaning Change
Claude Opus 4.750.5%67.7%+17.2 points
Claude Sonnet 4.646.5%68.7%+22.2 points
GPT-5.445.5%68.7%+23.2 points

Every gain is statistically significant, at p values of 0.0015 or lower. The same effect appears in other studies. One added sentence of external knowledge lifted GPT-4 by 20 points on the BIRD benchmark. An ontology took an insurance schema from 16.7% to 54.2%. A pharmacovigilance study reported a gain of 70 points.

Does a larger model correct the error?

No. The benchmark compared the three models against each other as well as against themselves. With the meaning document, the three models became statistically indistinguishable. Without it, they also became indistinguishable. Every comparison across the two groups was significant.

The conclusion in the researchers' words. "the most consequential architectural decision is not which frontier model to use but whether the system is grounded in authoritative business semantics at all." From Semantic Layers for Reliable LLM-Powered Data Analytics, 2026.

A team that records what its data means can select a model on cost and speed. A team that has not recorded it cannot recover the accuracy with a larger model.

What is a silent hallucination?

A silent hallucination is a wrong answer that arrives without an error. The model writes valid query code against the wrong column, or it applies the wrong aggregation. The query runs and returns a number. The researchers who named this failure call it more dangerous than a visible error, because a visible error stops the work.

Detection frameworks find 69% to 83% of these cases after the event. Between one in six and one in three therefore passes into a report or a board paper.

What the industry calls the answer

The answer has several names. A semantic layer, an ontology, a context layer, a metrics layer, or AI-ready data. Each name describes a written account of what your data means, supplied to the model with the data.

A written account works, and the figures above are the evidence for it. It also carries three weaknesses. A person writes it. A person maintains it. The data can change without it, because the document describes the data and does not govern the data.

What ESRE Media offers

We build your system on engage.re. The platform holds the meaning of every record type in one shared dictionary, and it enforces that meaning as a condition of storing data.

Your system therefore arrives with four properties on its first day.

  • Your vocabulary is declared and enforced. Your application registers every record type, field and relationship before it writes anything. The platform refuses a write outside that declaration. Your data cannot drift away from its meaning.
  • Your definitions survive a rename. Each dictionary entry carries a permanent identifier. You can change "Client" to "Resident" across your whole organisation, and no report and no screen breaks.
  • Your screens follow your definitions. Your application asks the dictionary for the fields of a record type, and it draws the answer. You add a field to a concept, and the screens that show that concept follow.
  • Your meaning matches everybody else's. Identifiers are the same across every system on the platform, with no synchronisation. Your system and a partner's system mean the same thing by employment, and neither of you holds a meeting about it.

The platform holds everything in five database tables, and those five tables never change. A new record type adds a dictionary entry instead of a table. engage.re has run in production since December 2025, and Sense Future built it.

A document against a dictionary

Property A written semantic layer An enforced dictionary
Who writes itA person, for each warehouseEach application, once, on registration
Who keeps it currentA person, from memoryThe server, on every write
Effect of a renameSomebody edits the documentNone. The identifier does not change.
CoverageOne warehouse and one teamEvery application on the platform
Agreement between two organisationsA mapping exerciseThe same identifier, with no negotiation
Cost of the tenth systemA tenth documentLower. It inherits nine vocabularies.

The same problem at two sizes

A council runs separate systems for housing, revenues, social care and waste. Each system holds its own idea of a person. A question that crosses two of them needs a mapping exercise, and a question that crosses four needs a project. On a shared dictionary the same concept carries one identifier in all four systems, and the question needs no project.

A care home group runs a records system, a rota tool and a family portal. Each one holds a resident. The registered manager wants a single occupancy figure, and today somebody reconciles three sources by hand. On a shared dictionary one record of a resident exists, and the figure computes from it.

Our guides to charity data management and care home software cover the smaller version of this problem. The care home software and charity software pages set out the sector detail.

What we do not claim

  • A shared dictionary does not make a model correct. The benchmark reached 68.7% with the meaning document, and not 100%. Shared meaning removes one large class of error.
  • These figures come from a 99-question benchmark on a public retail dataset. Your own error profile differs.
  • Written semantic layers work. The evidence in this article is the evidence for them. They need human upkeep, and a declared vocabulary does not.

What to do next

  1. Choose one number that two of your systems disagree about, and write down the definition behind each figure.
  2. Ask who recorded those definitions. If nobody recorded them, your AI tool invents them for you.
  3. Ask your AI tool ten questions whose answers you already know. Count the wrong answers that arrived without an error.
  4. Record the meaning of your data before you buy a larger model.
  5. Then decide where that meaning lives: in a document that a person maintains, or in a declaration that your server enforces.

Common questions

Why does AI get my own business data wrong?

Your database stores the structure of your data, and it does not store the meaning. The model reads your column names, and it cannot read the conventions your organisation applies to them. It therefore invents the business rules that nobody recorded. On real enterprise schemas, one leading model falls from 91.2% accuracy to 21.3%.

Will a larger AI model fix wrong answers about my data?

No. A 2026 benchmark tested three frontier models with and without a written description of the data. With the description, the three models became statistically indistinguishable from each other. Without it, they also became indistinguishable. The description accounted for nearly all the variance in accuracy, and the choice of model accounted for almost none.

What is a silent hallucination?

A silent hallucination is a wrong answer that arrives without an error. The model writes valid query code against the wrong column, or it applies the wrong aggregation. The query runs and returns a number, and nothing warns anybody. Detection frameworks find 69% to 83% of these cases after the event.

What is a semantic layer?

A semantic layer is a written account of what your data means, supplied to a model with the data. It records the measures, the conventions and the rules for ambiguous terms. An ontology, a context layer, a metrics layer and AI-ready data describe the same idea.

How does a shared dictionary differ from a semantic layer document?

A document describes the data, and a person maintains it. A shared dictionary defines the data, and the server enforces it. An application registers its vocabulary before it writes anything, and the platform refuses any write outside that declaration. The dictionary also covers every application instead of one warehouse.

What does ESRE Media build?

We build your system on engage.re, which holds the meaning of every record type in one shared dictionary. Your vocabulary is declared and enforced, your definitions survive a rename, your screens follow your definitions, and your meaning matches every other system on the platform.

Connect with us about your data foundation · · Replies within 24 hours

Sources and further reading