Before an AI feature uses a CRM record, your team should be able to explain who the record represents, whether its key information is current and which person owns the next action. An AI setting cannot resolve an ambiguous company association or decide which duplicate contact your business considers authoritative. Start with the data rules for the specific task you want AI to support.
TLDR: Define the required fields, identity checks, associations, ownership and freshness rules for one AI use case. Test an approved sample, route failures to named reviewers and verify the permitted data access. Passing a data check is a prerequisite for testing AI behavior, not proof that an AI output is correct.
Product documentation checked August 30, 2026. The sample contains 20 entirely fictional records. Its results are arithmetic examples, not a benchmark, customer audit or estimate of your account's quality.
We recommend choosing a narrow use case before reviewing the entire CRM. For example, your team might want a person to review an AI-assisted summary before following up with an existing business contact. List the information needed for that task and the actions that remain under human control.
If your wider review concerns how AI fits into a business process, our introduction to forward-deployed AI provides related context. This checklist focuses on the input data, not on evaluating an agent's output or approving autonomous actions.
Write an acceptance rule for each required input. “The record is complete” is vague. “The contact has an accountable owner and the company association has been reviewed for this use case” can be checked.
The following are our recommended rule categories. The required fields and thresholds must reflect your actual use case. They are not HubSpot's universal AI requirements.
| Rule category | Question to answer | Example of an acceptance rule |
|---|---|---|
| Completeness | Is the information required for this task present? | Every record in the pilot has the approved contact method required by the task |
| Identity | Are there unresolved candidate duplicates? | Ambiguous identities remain outside the pilot until reviewed |
| Normalization | Do values follow the agreed vocabulary? | Territory uses an approved value, with unknown values sent to review |
| Associations | Is the intended company or deal relationship clear? | The relationship used by the task has been checked against an approved reference |
| Ownership | Who can resolve a question about this record? | Each pilot record has an accountable operational owner |
| Freshness | Is the relevant information current enough for the task? | A reviewer has checked the required information within the team's approved interval |
Keep a reason beside each rule. If a field is required only because it has always been in the spreadsheet, ask whether it affects the task. This prevents the pilot from becoming an unrelated property cleanup project.
HubSpot supports validation rules for certain property types, but its documentation identifies exceptions, including workflows, chatflows, meeting scheduling submissions and legacy forms. Updated forms enforce applicable rules. Test the actual entry paths your process uses rather than assuming a property rule covers every update. HubSpot: property validation rules.
HubSpot's data quality tools provide views of issues such as duplicates and formatting, with access and feature eligibility requirements. The current overview documentation is marked beta. Treat the available tool as an input to your review and verify the features present in your account. HubSpot: data quality tools.
We recommend testing one valid and one invalid example through each approved data-entry route. Record whether the system rejects, accepts or changes the value. If a route bypasses a rule, decide whether a separate validation step or exception review is needed before that data reaches the AI use case.
Do not bulk rewrite records merely to improve a dashboard count. Preserve the original finding, investigate its cause and have the responsible owner approve a scoped correction.
Assume a fictional team is preparing a human-reviewed follow-up summary. For this example, it requires an approved contact email, an owner, a reviewed company association, an allowed territory and a recent information review. The sample includes no real people, company data or customer results.
| Finding in the fictional sample | Record references | Records affected |
|---|---|---|
| Missing accountable owner | QA-03 | 1 |
| Company association needs review | QA-06 | 1 |
| Territory value is outside the approved vocabulary | QA-08 | 1 |
| Unresolved candidate duplicate pair | QA-11 and QA-12 | 2 |
| Required contact email is missing | QA-15 | 1 |
| Information review is outside the example's chosen interval | QA-18 | 1 |
| No failures under these example rules | Remaining sample records | 13 |
The issue groups in this fictional sample do not overlap. Seven of 20 records fail at least one rule, leaving 13 that pass the stated checks. That is 65% of this sample, not a statistical estimate of a real database or a suggested launch threshold.
For a real review, count both issue occurrences and unique affected records. A single contact can fail several rules, so adding issue totals may overstate the number of contacts involved. Also record how the sample was chosen; a convenient sample does not establish the condition of every record in the account.
We recommend holding a pilot record when an unresolved issue could cause the task to use the wrong identity, context or owner. A missing optional descriptive field may have a different consequence. The acceptance decision should reflect that consequence rather than a single overall quality percentage.
For each failed rule, create an exception with the record reference, observed problem, responsible reviewer, proposed action and retest evidence. Avoid filling missing values with plausible guesses. If a reviewer cannot establish the correct value, keep the uncertainty visible and exclude the record where the task requires certainty.
In the fictional example, the two candidate duplicates remain on hold until someone resolves their identities. The example does not authorize a merge, deletion or automatic choice of a preferred contact.
HubSpot's AI settings include separate access choices for CRM data, customer conversations and files. Managing those settings requires Super Admin permissions. Review the current settings and the intended data sources with the authorized account owner before enabling a use case. HubSpot: manage AI settings.
We recommend documenting which information the task needs, why it needs it and who approves that access. Keep secrets and unnecessary personal information out of prompts, sample files and troubleshooting notes. Ask the responsible security or privacy reviewer to assess any sensitive-data requirement before proceeding.
A clean record is not automatically appropriate for every AI feature. Data quality, permitted access and acceptable use are separate approval questions. This article does not make a compliance finding for your account.
Before the pilot moves forward, we recommend a short acceptance record containing the use case, chosen population, required rules, unresolved exceptions and approved data access. Include the version of the dataset and the person who accepted it.
Then test the AI behavior separately. A data-quality pass does not establish that an output is accurate, appropriate or safe to act on. Keep the human review and action permissions defined for the use case until its own acceptance work is complete.
For a broader review of the portal, use our HubSpot audit checklist. Keep this pilot's record-level findings distinct from a general account score.
HubSpot's current Data Quality overview includes duplicate issues, formatting issues, unused-workflow recommendations and property insights. It is reached through Data Management > Data Quality, with access controlled by permissions and some features requiring additional subscriptions. Select a date range before comparing the displayed counts. HubSpot data quality tools.
If you know this area as the data quality command center, focus on the current tools and the records behind each recommendation. HubSpot also documents that a formatting correction overwritten later may not be flagged again. A quiet dashboard therefore does not prove the entire database meets your rules.
Use property insights to investigate duplicate properties or unused properties before changing the AI input set. An unused field and an outdated value are different problems. Check the properties' meaning, dependencies and owners; do not automatically delete a property or merge records just to reduce a displayed issue count.
A duplicate-record review asks whether two records represent the same person or company. Formatting issues ask whether the value follows your agreed convention. Outdated data asks whether a previously valid value is still reliable. Keep those decisions separate when reviewing individual contact and company records.
Consider a fictional sales rep preparing an account summary. Company properties might identify the intended account correctly while the contact is associated with an old employer. The AI input can still mislead the reader. Check contact relationships, company details and the relevant deals together before treating that record as ready.
When you review workflows, look for a process that may recreate the bad data after cleanup. An old import, integration or automated action can be part of the investigation. Record the suspected creation path and verify it with an approved test before changing active automation.
Record the population, date range, failed rule, responsible owner and correction evidence. Use consistent column headers in exports so another person can compare the same cohort later. Keep new records and previously reviewed records distinguishable in your analysis.
Sales, marketing and operations should agree which exceptions matter for the intended AI task. A missing optional field may be acceptable; an unresolved identity or ownership error may not be. This gives your data management work a business decision to support without pretending that all data quality issues carry equal risk.
We recommend meeting the specific rules for the chosen pilot and holding records with unresolved risks. A bounded, reviewed population is more useful than an undefined demand to clean everything first.
No. A quality percentage describes only the rules and population measured. Test the AI output and its permitted actions separately before relying on it.
Have the responsible owner resolve the identity question first. Do not treat a duplicate suggestion as approval to merge or delete a record. Exclude unresolved identities when the use case requires a reliable match.
Assign each issue to the person who can establish the correct business value or relationship. An administrator can implement an approved correction, but should not have to invent the answer.
Our HubSpot operations services can help you define the input rules, review paths and responsibilities for a scoped AI pilot.
When you talk with Selworthy, bring the proposed task and the fields it depends on. That gives the review a clear purpose and a practical acceptance standard.