Featured image: AI-generated conceptual illustration. Sources checked 26 October 2026.
A CRM hygiene agent should be judged by the mistakes it prevents, especially the irreversible ones. Finding two similar names is easy. Establishing that two records represent the same person, deciding which information should survive, and understanding what a merge will trigger require a more careful process.
This creates a useful but bounded service: inspect records, identify data-quality problems, prepare evidence-backed corrections, and apply only authorized changes. A one-time cleanup can lead to recurring monitoring, provided the ongoing service catches meaningful issues rather than producing an endless stream of low-value alerts.
Separate defects from identity decisions
Some problems are straightforward: an invalid date format, a missing required field, or a telephone number containing letters. Others are semantic: an outdated job title, two possible employers, or a contact whose email domain has changed. Identity merges belong in a higher-risk category because they can combine histories, permissions, and communications that should remain separate.
Create distinct queues for formatting fixes, missing information, stale information, and possible duplicates. Give each queue its own evidence standard. Do not use one confidence score to justify every action. A model that is excellent at formatting addresses may still be unreliable at identifying people with common names.
HubSpot's contact API guide shows operations for managing contact records. Because this is a legacy reference, check current API documentation and migration requirements before building. Technical access to a merge operation does not establish that a merge is correct.
A match needs contradictory evidence too
Suppose two records share the name Jordan Lee and the same company. One contains a finance email alias and the other a personal business address. They might be one person, two colleagues, or a shared mailbox mistakenly stored as a person. The agent should show the supporting and conflicting evidence rather than treating similarity as proof.
Strong identifiers can help, but their meaning must be understood. An exact email match may indicate a duplicate, yet shared inboxes and recycled addresses complicate identity. A telephone number can belong to a switchboard. A customer ID may refer to an organization rather than an individual. Document what each identifier actually represents in the customer's systems.
Use rules to identify candidate groups, then require review for ambiguous merges. Present the source of every field, the most recent verification date, and the consequences of the proposed action. If the reviewer cannot resolve the identity safely, leave the records separate and mark the issue for investigation.
Preserve history before changing state
A cleanup job should begin with an export or another tested restoration method. A backup that cannot be restored is not an operational safeguard. Understand whether the CRM supports undoing a merge and which associated records, activities, subscriptions, and permissions might move with it.
Build a change plan before execution. Each row should identify the record, proposed operation, old value, new value, evidence, approver, and intended batch. Apply a small batch first, inspect the results, and pause if unexpected downstream behavior appears. Large-scale cleanup should not be the first time the team learns that a field change triggers marketing outreach.
After each approved mutation, verify the resulting state. Evaluation guidance from Anthropic supports checking the actual outcome rather than an agent's claim. For data hygiene, count the correct records changed once, not the number of tool calls that returned without errors.
Make monitoring quieter over time
Recurring monitoring should focus on new defects and unresolved high-priority cases. Repeatedly notifying the customer about the same accepted exception creates alert fatigue. Record dispositions such as confirmed duplicate, legitimate separate records, insufficient evidence, and deferred investigation. Preserve the reason and when it should be reconsidered.
Look for upstream causes. If a form routinely creates contacts without company identifiers, fixing the form may be more valuable than cleaning the same defect every month. If a sales team uses inconsistent field definitions, a training and schema change may outperform a more capable model. A managed service should reduce the source of the work, even when that reduces its own processing volume.
Avoid enriching records with speculative personal information. A hygiene service should improve the accuracy of authorized business data, not build intrusive profiles. Use only permitted sources and respect deletion, access, and communication preferences throughout the pipeline.
Price by risk and responsibility
A sensible project price accounts for record count, number of systems, field complexity, review effort, and rollback requirements. A database with 20,000 well-structured contacts may be easier than 2,000 records with tangled histories and unclear ownership. Charge for the agreed cleanup scope, not a promise to produce a particular number of merges.
For illustration, a US$3,000 cleanup project requiring 20 hours of analysis and review at a loaded US$80 hourly cost leaves US$1,400 before software and overhead. If the customer cannot supply a reliable export or an authorized reviewer, the effort can grow substantially. State prerequisites and change-order rules before beginning.
A recurring subscription can cover monitoring, a capped correction queue, and a monthly root-cause report. Do not price destructive operations as an incentive to maximize volume. A service that earns more by merging more records has a conflict with the customer's need for caution.
Measure precision before coverage
Report the proportion of proposed duplicate groups that reviewers confirm, the number of incorrect merges, restoration success, review minutes, and downstream contact errors. Coverage matters, but an aggressive detector that overwhelms staff with false matches can be less useful than a conservative one.
Stop automation when the system cannot distinguish people reliably or when restoration is untested. Continue with a read-only audit if that remains useful. The commercially durable promise is cleaner, more trustworthy records with an explainable change history. It is never that every similar-looking contact can safely become one.
Sources and research note
HubSpot contacts guide. Demystifying evals for AI agents. Research checked October 26, 2026. Product behavior and prices can change. Business examples and proposed prices in this article are illustrative unless explicitly attributed.
Leave a Reply