← Research library

Check your data before you deploy marketing agents

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Marketing teams are moving agents from demos into workflows that touch real customer records. At that point the limiting factor is rarely the model. It is the data the agent reads and writes.

An agent working from bad data does not make one mistake. It makes the same mistake quickly, confidently, and at scale. A segmentation error that a person might catch once becomes a thousand personalized emails with the wrong company name, and each one looks deliberate to the person who receives it.

That is why we treat data readiness as the first step of any agent project rather than a cleanup task afterward. Below is the six-part checklist we use before connecting an agent to marketing data, followed by how to score it. Run it per workflow, not per company, because the data one agent needs can be in good shape while another agent's data is not.

1. Ownership

Unowned data degrades, and agents tend to expose that degradation quickly. Before an agent reads a source, answer three questions about who stands behind it:

  • Is there a named owner for each data source the agent will use?
  • Does that owner know an agent will be reading from it, and possibly writing to it?
  • Is there a route for reporting data problems the agent discovers?

The third question is easy to skip. Agents notice malformed records, contradictions, and missing values as a side effect of their work. If there is nowhere to send those observations, they are lost, and the agent either guesses or stalls. A simple queue that the owner reviews weekly turns the agent into a data quality sensor.

2. Definitions

Agents interpret fields literally. If "lifecycle stage" means different things in your CRM and your marketing automation platform, the agent will not notice the difference, and it will act on whichever meaning it finds first. Check three things:

  • Are key fields such as "qualified lead", "active customer", and "industry" defined in writing?
  • Do different systems use the same definitions for fields with the same name?
  • Are the definitions available to the agent in plain language?

Written definitions are also the first step toward a formal agreement between data producers and agents. We describe that agreement in the marketing data contract.

3. Completeness and accuracy

You do not need perfect data. You need to know where the gaps are so the agent can be told how to handle them. Three questions establish that:

  • What share of records have the fields the agent needs?
  • When was a sample of records last checked against reality?
  • Are there known problem areas, such as outdated job titles or duplicate accounts?

Duplicates deserve special attention because they cause visible failures: the same person receives two versions of a message, or an opt-out on one record is missed on another. Identity resolution in the age of agents covers how to reduce that risk.

4. Freshness

Personalization based on a job change from two years ago is worse than no personalization at all. Freshness problems are hard to see because stale data looks exactly like current data. Ask:

  • How often does each source update, and is that schedule written down?
  • Will the agent ever act on data older than is safe for its task?
  • Can the agent see when a record was last updated?

The last check is cheap to fix and widely useful. If the agent can see a last-updated date, you can instruct it to treat older values with caution, for example by leaving them out of a personalized opening line.

5. Consent and permissions

Consent is the section where a gap does the most harm, because the failure lands on a person who asked not to be contacted. Check that:

  • Consent status is stored and reliable for every contact the agent might message.
  • Records the agent must never use, such as opted-out contacts, are identified.
  • Records subject to regional rules are identified, and those rules are written down.
  • The agent's access enforces these restrictions in code, not only in its instructions.

Instructions can be misread or overridden by other content the agent sees. A filter applied before data reaches the agent cannot. We discuss the design in more depth in consent-aware personalization, and the legal specifics for your regions are worth confirming with counsel.

6. Write safety

Read access exposes data to an agent. Write access lets the agent change it, which is a different level of risk. If the agent will write, confirm:

  • Which fields it can change, listed explicitly.
  • That every change is logged with the agent's identity and reason.
  • That changes can be reversed easily, and that reversal has been tested.

Start with read access only, and grant write access field by field, with logs. Connecting agents to your CRM walks through that sequence for the system where it matters most.

Scoring the checklist

For each section, rate readiness as ready, partly ready, or not ready. The scoring rule we recommend is simple. Any "not ready" in consent or write safety blocks deployment. Gaps elsewhere can often be managed by instructing the agent to flag missing or stale data rather than guess.

Consider a hypothetical lifecycle agent that will send renewal reminders. Ownership, completeness, freshness, and write safety are ready. Definitions are partly ready, because "renewal date" lives in two systems. Consent is not ready, because opt-outs from a recent webinar import were never synced.

That single gap blocks launch, even though four of six sections look healthy.

The fix in this example is narrow. Sync the opt-outs, add a pre-send filter, and pick one system as the source of truth for renewal dates. The agent can then launch with an instruction to flag records where the two dates disagree.

What to measure after launch

Readiness is not a one-time state. Data that passed the checklist in spring can drift by autumn as fields are added, imports run, and owners change roles. We recommend re-running the checklist each quarter. In between, track two signals: the number of data problems agents report, and the share of runs that stop because required data was missing or stale.

A rising count of agent-reported problems is often a good sign early on, because it means the reporting route works. The question we are studying is what a healthy steady state looks like, and how quickly owners should be expected to clear what agents find. If you are still choosing which workflow to start with, data readiness is a strong input to the impact and risk matrix.

How to apply this: run this checklist with your data or marketing operations lead before connecting any agent to customer data. Fix consent and write safety first, and tell the agent how to handle the gaps you find elsewhere.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.