Meridian agentic data-quality audits and model building: human-review governance

Short answer: Treat Meridian's new agentic capabilities as modeling assistance that accelerates data-quality review and model construction, not as autonomous authority to approve an MMM. Google says Meridian is gaining agentic capabilities to audit data quality, resolve errors and guide model building in real time. Keep the underlying dataset, transformations, priors, model specification, diagnostics and reviewer decisions versioned so every material recommendation can be inspected. The agent can help find problems; the analyst still owns whether the data and model are fit for a business decision.

What Google currently documents

Google's September 2026 measurement update says Meridian, its open-source Marketing Mix Model, is getting new agentic capabilities designed to help users audit data quality, resolve errors and guide model building in real time.

Google also says back-end updates are intended to make analyses run faster and more efficiently.

Those capabilities reduce modeling friction. They do not eliminate uncertainty, identification assumptions or the need to review model fit.

Step 1: freeze the input dataset

Before accepting agentic guidance, preserve a versioned modeling input.

Record:

• data version;
• time range;
• geography;
• media channels;
• spend/exposure fields;
• outcome variable;
• control variables;
• transformations;
• missing-data treatment;
• owner;
• extraction timestamp.

A recommendation cannot be reproduced if the underlying data keeps changing silently.

Step 2: classify data-quality findings

Use categories such as:

• missing periods;
• unexpected zeros;
• unit mismatch;
• currency mismatch;
• outlier;
• duplicate rows;
• geography mismatch;
• channel-definition drift;
• outcome break;
• incomplete control variable.

Do not accept a generic “data quality improved” statement without knowing what changed.

Step 3: preserve proposed fixes

For each error resolution, record:

• finding;
• proposed fix;
• affected rows/fields;
• analyst decision;
• applied transformation;
• rationale;
• before/after check;
• model impact.

Never overwrite source data to make the model run without preserving the original evidence.

Step 4: separate software error from business-data ambiguity

Some problems are technical, such as invalid types or missing values.

Others require business judgment, such as:

• whether two campaigns belong to one channel;
• whether a price promotion is a control;
• whether a market closure should be excluded;
• whether an outcome definition changed;
• whether a brand event needs a separate variable.

Route semantic ambiguity to the domain owner rather than letting automation guess.

Step 5: version model specifications

For every material model iteration, store:

• model/version ID;
• data version;
• channels;
• controls;
• priors/assumptions;
• transformations;
• calibration inputs;
• run date;
• reviewer;
• result status.

This lets analysts distinguish an improved model from simply a different specification.

Step 6: review agentic model-building guidance

When Meridian suggests a next step, classify it as:

• diagnostic suggestion;
• data transformation suggestion;
• model-specification suggestion;
• calibration suggestion;
• interpretation suggestion;
• unresolved question.

Require explicit human review for changes that materially affect estimated channel contribution or business recommendations.

Step 7: keep uncertainty visible

An MMM should not be reduced to one deterministic answer.

Preserve:

• uncertainty intervals where available;
• sensitivity to assumptions;
• model-fit diagnostics;
• holdout/calibration evidence;
• data limitations;
• unmodeled events;
• structural breaks.

Agentic speed should not make uncertainty disappear from the decision memo.

Step 8: define approval boundaries

Use approval states such as:

• DATA_REVIEW_REQUIRED;
• DATA_ACCEPTED;
• MODEL_SPEC_REVIEW_REQUIRED;
• MODEL_RUN_ACCEPTED;
• INTERPRETATION_REVIEW_REQUIRED;
• DECISION_READY;
• REJECTED.

The agent should not silently transition a model to DECISION_READY.

Step 9: reconcile with causal evidence

Where geo-experiments or other causal evidence exist, use them to challenge or calibrate the model rather than assuming the MMM is correct because it converged.

Record:

• experiment source;
• period;
• treatment/control definition;
• measured incrementality;
• calibration use;
• conflict with model estimate;
• resolution decision.

Step 10: maintain an audit trail

For every decision based on the model, retain:

• dataset/model version;
• key agentic suggestions;
• accepted/rejected changes;
• reviewer;
• final interpretation;
• business decision;
• later outcome/review.

This separates reproducible assistance from opaque automation.

Governance states

Use states such as:

• INPUT_FROZEN;
• DATA_QUALITY_ISSUE;
• FIX_PROPOSED;
• FIX_VALIDATED;
• MODEL_SPEC_REVIEW;
• MODEL_ACCEPTED;
• UNCERTAINTY_MATERIAL;
• HUMAN_APPROVAL_REQUIRED.

The governance rule

Meridian's agentic capabilities should be managed as accelerated modeling assistance inside a versioned human-review process.

Let the system surface issues and guide construction, but preserve the data, assumptions and approvals that determine what the model actually means. Faster model building is not a substitute for accountable model governance.

Sources reviewed

• https://blog.google/products/ads-commerce/data-strength-updates/