← Research library

Identity resolution matters more once agents act on your records

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.

As agents take on more outreach, research, and personalization, they act directly on customer records rather than on reports a person reads first. That makes the quality of those records a frontline concern, and identity is where the quality problems hide.

Identity resolution is the work of deciding which records refer to the same person or company. The same buyer might appear as a form submission with a personal email, a CRM contact with a work email, an event attendee with a misspelled name, and an anonymous website visitor. Stitching these together has always been hard. Agents raise the stakes, because they act on whatever picture of the customer the records give them.

This post covers why agents make identity more important, the failure modes to watch for, the principles we recommend, and where agents can help with the work itself.

Why agents make identity more important

A human marketer looking at two partial records might notice they seem to be the same person and act accordingly. An agent usually takes records at face value. It sees two contacts, so it treats them as two people.

Speed and volume compound the problem. A person working through a list makes a handful of identity mistakes, and many get caught in passing. An agent working through the same list in minutes makes every mistake the data allows, and it makes them consistently.

Agents also combine data in new ways. A research agent may pull a job title from one record, a recent support ticket from another, and a webinar attendance from a third. If those records were linked incorrectly, the resulting brief describes a person who does not exist.

Four failure modes to watch for

Poor identity resolution shows up in agent work in four recognizable ways. Each one points to a different weakness in the underlying data:

  • Duplicate outreach: the same person receives two versions of a message because they exist as two contacts.
  • Wrong personalization: details from one person's record are applied to another because records were merged incorrectly.
  • Missed context: an agent treats an existing customer as a new prospect because their records are not linked.
  • Consent errors: an opt-out recorded on one record is not applied to a duplicate.

Consent errors are the most serious of these, and they are a strong reason to fix identity before scaling agent outreach. A missed link that causes a duplicate email is an annoyance. A missed link that ignores an opt-out can breach the trust the person placed in you and may create legal exposure, which is worth reviewing with counsel.

Principles for agent-ready identity

The goal is a set of rules that keeps agent mistakes small when the identity graph is imperfect, which it always will be. Five principles do most of that work.

Be conservative with merges. A false merge, which combines two different people, is usually worse than a missed merge. It mixes personal details, purchase history, and consent status in ways that are hard to untangle. Require strong evidence, such as a verified email match, before merging records automatically.

Propagate consent across linked records. If any linked record has opted out, treat the person as opted out everywhere. This rule should be enforced in the data layer, not left to agent instructions. Consent-aware personalization covers how to build it in.

Expose confidence. When records are linked with less certainty, mark them as such. Instruct agents not to rely on details from low-confidence links for personalization, while still using them for suppression.

Keep account and person identity separate. Company-level matching and person-level matching have different rules and different risks. Agents doing account research, as in agentic ABM, need clean account records even when contact data is messy.

Let agents report problems. Agents often notice likely duplicates or inconsistent data. Give them a way to flag these for review rather than ignore them or quietly fix them.

A worked example: one buyer, three records

Consider a hypothetical buyer who appears three times. One record is a webinar registration with a personal email. Another is a CRM contact with a work email and a job title. The third is a newsletter subscriber with the same personal email, who opted out of promotional email last month.

The webinar record and the newsletter record share an email address, so a conservative rule links them with high confidence. The opt-out now applies to both. The CRM contact shares a name and company domain with the webinar record but no email, so the link is marked low confidence.

A lifecycle agent preparing a promotional sequence now has clear rules. It suppresses both high-confidence records. For the low-confidence CRM contact, the safe default is to apply the opt-out until a person confirms the match. An extra suppressed contact costs little.

If the agent drafts anything for that contact, it does not reuse details from the webinar record.

Where agents can help

Agents can assist with identity work itself, under human supervision. They are well suited to proposing likely matches with their reasoning, flagging records with conflicting information, and summarizing data quality issues for an owner to review.

Final merge decisions, especially for high-value accounts, should rest on rules that are well tested or on people who can review the evidence. An agent's proposed match with a written reason is a useful input to that decision. It should not be the decision.

This division of labor also produces useful evaluation data. Each proposal that a reviewer accepts or rejects becomes a labeled example, which tells you over time whether the agent's matching judgment is good enough to automate more of it.

What to measure

Identity quality is easy to neglect because its failures are scattered. We recommend tracking a few signals each month so the problem stays visible:

  • Duplicate rate: suspected duplicates per thousand active contacts, from agent flags and periodic checks.
  • Consent propagation: linked records where the opt-out status disagrees, which should be zero.
  • Merge reversals: merges that had to be undone, a direct measure of false merges.
  • Personalization corrections: reviewer edits caused by details from the wrong person or account.

Treat a nonzero consent propagation count as a defect to fix that week. The other measures are trends to watch.

Identity rules belong in the same written agreements as other critical data, as described in the marketing data contract. This work feeds the data layer shown in our signal data flow diagram. The question we are studying next is how identity confidence should travel with a record into agent context. An agent should be able to weigh a fact by how sure we are about whose fact it is.

How to apply this: check whether opt-outs propagate across duplicate and linked records in your systems. If they do not, fix that in the data layer before any agent sends messages at scale.

Related research

Data & Infrastructure

Connecting agents to your CRM without breaking it

The CRM is the system of record for customer relationships. Giving agents access safely takes a deliberate plan for permissions, writes, logging, and reversal.

5 min read

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.