Marketing AI Has a Trust Problem

Marketing AI moved beyond chatbots a while ago. Models have shifted from answering questions to powering systems that pursue objectives, query databases, run analyses and take action.

Agents can reallocate budgets, pause campaigns and trigger workflows. The distance between “the model said something wrong” and “the model did something wrong” is collapsing.

The critical question is what data the model uses. A model working from bad data does not simply produce a bad answer, it produces a well-argued bad answer.

Agency amplifies data quality. It never substitutes for it.

The context AI needs is not in the warehouse

Give the modern data stack its due: it solved problems that were impossible a decade ago. Storage is effectively infinite, compute elastic, pipelines industrial.

What it has not solved is semantics, or meaning. A warehouse can tell an agent that a table called campaign_perf_v3_final contains a column called "spend". It cannot say whether that spend is gross or net, local currency or USD or which of four conversion tables finance trusts.

A human analyst resolves that ambiguity through institutional knowledge. A uniformed agent cannot.

It sees how the data is organised, but not what it means, where it came from or whether it can be trusted. It cannot tell the table behind the board deck from the abandoned experiment beside it.

Marketing data is the hardest version of this problem

If enterprise data has a semantics problem, marketing data has a semantics crisis, for structural reasons no pipeline hygiene fixes.

Identity is fragmented by design. The same customer appears as a cookie, device ID, hashed email, CRM record and loyalty number across systems never built to communicate with each other. With no universal join key, identity must be resolved, probabilistically and defensibly, before customer-level questions can be answered.

The platforms grade their own homework. Each reports performance on its own definitions and attribution window, with a structural incentive to claim credit. Summing platform-reported conversions routinely yields more than the business recorded. The raw data is not merely messy. Pieces of it are adversarial.

Taxonomy is entropy. Naming conventions decay as teams, agencies and regions touch an account, and channel definitions drift: "paid social" can mean different things in different quarters. The warehouse stores every inconsistency and enforces none.

The journey is the unit of meaning, and it exists in no table. The question a CMO actually asks – "What sequence of touchpoints moved this customer, and what did each contribute?" – requires paths reconstructed across fragmented identities, duplicated exposures and conflicting timestamps. No platform exports it and no warehouse infers it. It has to be built.

Point the most capable model in the world at this sprawl and you get confident wrong answers.

Agentic workflows compound the risk. An agent interprets the question, chooses the data, resolves the metrics, runs the analysis and synthesises an answer, each step depending on the one before.

Thanks to Lusser's law, we know that connecting 10 independent steps together that each have 95% reliability will result in just 60% reliability for the entire end-to-end process.

Agents do not average out data errors. They multiply them, then act on the result.

Definitions alone cannot fix the data

Semantic layers let organisations define metrics once and hand agents consistent definitions. For finance data, where entities are stable and definitions contractual, that may be enough.

For marketing, it solves only the last mile. A semantic layer can tell an agent what ROAS means, but it cannot complete the underlying record or flag when the data is unreliable.

Definitions govern meaning, they do not create the resolved, reconciled entities meaning depends on.

AI needs a system of record for marketing performance

Marketing needs a deeper, persistent layer where that work is done once and trusted everywhere.

A System of Record for Marketing Performance provides a shared, traceable foundation for measurement, simulation, optimisation and AI.

As agents gain autonomy, that lineage is what makes their decisions defensible and trustworthy.