The myth: the data is too dirty for AI
The usual explanation for a stalled AI project is poor data. Gartner predicted in February 2025 that, through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. Its July 2024 survey of 1,203 data management leaders found 63% of organisations lacked, or were unsure they had, the right data management practices for AI.
Most leaders read that as a cleaning problem: duplicates, missing fields, late feeds. RAND's 2024 study, built on interviews with 65 experienced data scientists and engineers, puts a different cause first. Stakeholders misunderstand or miscommunicate the problem, and models are deployed that have been optimised for the wrong metrics. Missing data comes second.
The reality: clean enough, and still not agreed
Two terms need defining. Data is clean when each record is accurate, complete and current. Data is agreed when everyone who relies on a figure means the same thing by it. I design data platforms for leaders who have to trust what they read, and the first condition is met far more often than the second.
Consider four ordinary words. Revenue can be booked, billed or recognised. A customer can be a legal entity, a billing account or a parent group. Margin can sit before or after rebates and allocated overheads, and headcount may or may not include contractors.
Each version is defensible, and each has a department behind it. The consequence is what I call a number nobody would sign: a figure that appears in a pack every month, yet no executive would put a name beside it if asked to certify it. People cope with such numbers because they know which colleague to ring. A model has no colleague to ring.
Lineage answers a question about people
Lineage is usually described as a technical map of where data came from. Its more useful meaning is the answer to one question: who would defend this figure in front of the auditor? If no name comes back, the trail is incomplete, however tidy the pipeline.
Regulators already treat meaning as part of the duty. In Australia, where I work, APRA's guide CPG 235 on managing data risk expects banks, insurers and superannuation funds to understand data lineage, and counts consistency of definition as a dimension of data quality. The EU AI Act goes further for high-risk systems: Article 10 requires governance of the assumptions about what the data are supposed to measure and represent.
In Saudi Arabia, the National Data Management Office's standards set 77 controls across 15 domains for government entities, audited every year. In all three markets, an enterprise that cannot say what a figure means is carrying a compliance gap as well as an analytical one.
Why an agent multiplies the disagreement
A dashboard built on an ambiguous metric starts an argument in a meeting. An AI agent built on the same metric starts acting. Tell an agent to protect margin and it will reprice, approve discounts and reorder stock against whichever margin it found first.
By the time finance notices that the agent's margin excludes rebates, thousands of decisions have been taken on it, each one internally consistent. Disagreement that used to surface once a month now compounds at machine speed. The agent did nothing wrong. It was given a word the company had never defined.
The dictionary is an executive document
The technical remedy is a semantic layer: one governed place where each business term is defined once, with its calculation, and from which every report, model and agent must draw. I think of it as the company's dictionary. Building it is well within a data team's reach.
The data team cannot write the entries. Deciding whether revenue means booked or recognised changes bonuses, targets and how a division looks to the board. That is why the first AI project in any company is really a negotiation about meaning, and why it belongs to the executive team. Delegating it produces a dictionary nobody senior has read.
The meeting where nobody asks whose number it is
Picture the executive meeting once that negotiation is done. The margin figure on screen carries a small line beneath it: the definition in one sentence, the owner's name, and the date it was last reconciled to the ledger. The agent that adjusted prices overnight used that same entry, and its log cites it.
The chief executive does not ask whose number it is. The discussion begins with what to do about it, and the time once spent reconciling versions goes to the decision. When a definition has to change, the owner proposes it and the committee records it. Every model and agent picks up the new meaning the same day.
This is what makes wider delegation to AI safe. An enterprise can let software act on a figure only when a named person would sign it.
Four words to settle this quarter
Put revenue, customer, margin and headcount on the next executive agenda. For each, ask every member to write down the definition they use, then agree one and name the person who will sign for it.
An AI opportunity evaluation with NectarGlobal starts from those signed definitions and shows which AI uses they can already support.