resources

Why your CRM data quality decides whether AI works

AI multiplies whatever it is given. Six CRM data quality checks, each a number you can pull from Salesforce or HubSpot in an afternoon, that decide the outcome before any model runs.

September 11, 2026

11 min read

THE MULTIPLIER

SIX CHECKS

FIXING IT ONCE

Why your CRM data quality decides whether AI works - cover image

TL;DR

Validity’s 2025 State of CRM Data Management, a survey of 602 CRM users, found 76% believe less than half of their CRM data is accurate and complete, and 37% have lost revenue directly because of poor data.

Salesforce’s own 2026 State of Data and Analytics (7,652 leaders) estimates 26% of enterprise data is untrustworthy and reports that 89% of organizations with AI in production have already seen it produce inaccurate outputs. Only 43% have formal data governance.

Six checks decide what an agent will do with a CRM: duplicate rate, completeness on the fields the use case reads, staleness, ownership, definitions, and lineage and permissions. Each is a number tied to a specific object, field and threshold that a Salesforce or HubSpot admin can produce in an afternoon.

Cleanups decay at the same rate the data does, roughly 30% a year by industry estimates. The durable fix is structural: duplicate rules that block, validation at stage transitions, an archive policy, an owner map and a twenty-minute monthly review.

The agent will act on the record you gave it

Nothing an AI system does inside a CRM is more sophisticated than the record it reads. Picture an agent asked to draft a renewal email. It opens the Account, finds the Contact flagged as primary, reads a Title last updated in 2023 and writes a confident, personalized note to someone who left the company fourteen months ago. A human rep would have paused. The agent has no reason to; the record said what it said.

That is the whole argument. CRM data quality, meaning the completeness, accuracy, consistency and timeliness of the records in Account, Contact, Lead, Opportunity and Case, decides more about an AI rollout than the model behind it. Clean, owned, connected data gets compounded. Three versions of the customer and a pipeline nobody has touched since the last quarterly review get multiplied too, faster and with more confidence than any human would bring to the job.

Nobody trusts the data, including the vendors

The most striking evidence comes from the companies that sell the systems. Salesforce’s State of Sales, sixth edition, surveyed 5,500 sales professionals in 2024 and found only 35% completely trust the accuracy of their organization’s data; the same study puts 70% of a rep’s time on non-selling work, much of it moving and correcting information. Salesforce’s first State of Data and Analytics, in November 2023, found 57% of data leaders and just 43% of line-of-business leaders completely trust data accuracy. By the second edition, in January 2026, 84% of 7,652 leaders said their data strategy needs a complete overhaul for AI, 19% of data was reported siloed or inaccessible, and 26% was estimated untrustworthy.

Validity’s 2025 State of CRM Data Management, from 602 CRM users, is blunter. 76% say less than half of their CRM data is accurate and complete, and 37% have lost revenue directly because of it. The 2024 edition found 31% putting the cost of bad data at 20% or more of annual revenue, and 67% of CRM admins worried about readiness for AI and machine learning. Outside the CRM world the trend runs the wrong way. Precisely and Drexel’s LeBow College found 67% of organizations do not fully trust the data they use for decisions in 2025, up from 55% in 2023, with 64% naming data quality their top challenge. IBM’s 2025 study of 1,700 chief data officers found only 26% confident their data can support new AI-enabled revenue streams.

The cost was measurable before any of this. Gartner’s estimate of poor data quality’s cost to the average organization is $12.9 million a year, and IBM, citing Forrester, reports that more than a quarter of organizations lose over $5 million annually to it, with 7% losing $25 million or more. What AI changes is the transmission. A wrong field used to be caught, sometimes, by a person who noticed. Now it becomes an action.

Why AI makes it worse, not only faster

Three mechanisms. The first is silent inheritance: every model and every agent reads from the same records, so a defect in Contact Email or Opportunity CloseDate reaches every downstream decision at once. The second is confidence. An agent does not hedge on a stale record; it writes as if the record were true. Salesforce’s 2026 study found 89% of organizations with AI in production have already seen inaccurate outputs, and 49% say the company occasionally or frequently draws incorrect conclusions from its data. The third is decay. Industry estimates put B2B contact data decay at 2–3% a month through job changes, mergers and reorganizations, around 30% a year. A CRM nobody maintains is a third wrong within twelve months, and an agent reading it does not know which third.

Fragmentation compounds all three. MuleSoft’s 2026 Connectivity Benchmark counts 957 applications in the average enterprise, 27% of them integrated; 96% of IT leaders say agent success depends on data integration, and half of the agents already deployed run in silos. The six checks below are each a number, tied to a specific object and field. If nobody in the company can produce the number, that is the first finding.

Diagram: What we measure before any AI touches a CRM

Six checks, and how to measure each one

Run these on the objects the AI use case will actually touch. For a sales agent that is Account, Contact and Opportunity; for a service agent, Case and Contact; in HubSpot, Companies, Contacts, Deals and Tickets. Measuring everything is how the exercise turns into a quarter-long project that nobody finishes.

1. Duplicate rate

What to measure: the share of Account and Contact records that exist more than once. In Salesforce, activate Matching Rules and Duplicate Rules on Account, Contact and Lead in report-only mode, let them run, then report on Duplicate Record Sets; records in sets divided by total active records is the rate. A faster first pass is a Contacts report grouped by Email with a row count above one, and an Accounts report grouped by Website. In HubSpot, the Data Quality Command Center lists suspected duplicate Contacts and Companies; export the list rather than merging from the screen.

Threshold: our working line is 3% on Account and 5% on Contact. Above that, an agent will update the wrong record often enough that reps stop trusting anything it touched, and merging becomes the first workstream rather than a cleanup task.

2. Completeness on the fields that matter

Not every field. Pick the eight to twelve the process and the AI use case depend on, and measure fill rate on those. For a pipeline use case: Opportunity Amount, CloseDate, StageName, NextStep and the primary Contact role; Account Industry, Website and Owner; Contact Email and Title. In Salesforce, one report per object with a filter of “field equals blank,” grouped by Owner Role, gives fill rate per team, which is where the pattern hides. In HubSpot, property insights inside the Data Quality Command Center show fill rate per property directly.

Threshold: 90% on any field the workflow reads or writes. Below 70%, the field is decorative. Either enforce it with a validation rule at the stage where it becomes required, or retire it so the agent stops reading noise.

3. Staleness

What to measure: the distribution of last-touch dates on records the business calls active. Salesforce keeps two useful fields, LastModifiedDate and LastActivityDate, and the second is the honest one, because integrations touch LastModifiedDate constantly. Build a Contacts and Accounts report with a bucket field on Last Activity: under 90 days, 90 to 365, over a year. Add an Opportunities report filtered to open stages with CloseDate before today; every row is a deal the forecast is counting that the rep has given up on. In HubSpot, the equivalents are Last activity date and Last modified date on Contacts and Companies, plus Deals with a close date in the past and a stage that is not closed.

Threshold: if more than 30% of “active” Contacts have no activity in a year, the list is an archive, and an agent should not prospect from it. Open Opportunities past their close date should be near zero at the start of any AI work. Each one is a false signal the agent will act on.

4. Ownership

An OwnerId on a record is a routing field, not accountability. The check is whether every object, integration and dashboard the use case depends on has a named person answerable for its quality. Two measurements make it concrete. First, records owned by deactivated users: an Accounts report filtered on Owner Active equals false, and the same for Contacts and open Opportunities. Second, an owner map, a sheet with one row per object, integration and dashboard, and a name beside each. If the name in every row is the admin, the answer is no one.

Threshold: zero active records owned by inactive users, and no blank rows in the owner map. HubSpot shows the same problem as Contacts and Companies with Contact owner unassigned, which is where imported lists go to be forgotten.

5. Definitions

What to measure: whether lead, opportunity, customer, churn and revenue each have one written definition used by every report and every model. The test takes an hour. Ask sales, marketing and finance for last quarter’s pipeline number, separately, and compare. Then open the Opportunity StageName picklist and ask three sales managers what “Proposal” means; in HubSpot, do the same with Lifecycle stage and Deal stage. The definitions problem usually reveals itself as a picklist with the same value spelled three ways, or a stage that one team skips.

Demand Gen Report’s 2026 survey found only 50% of B2B companies have a unified sales and marketing database, which is another way of saying half of them have two definitions of a customer. dbt Labs’ 2024 State of Analytics Engineering found 57% of practitioners naming poor data quality their top problem, up from 41% in 2022; a large share of what they call quality is definitions that disagree. Threshold: one written definition per term, with a version date, and every dashboard filter that implements it saved as a named report.

6. Lineage and permissions

Two governance checks that decide whether anyone can defend what the AI did. Lineage: take the largest number on the last board deck and trace it to a saved report and its filters. If the trail ends at a spreadsheet, there is no lineage. insightsoftware’s 2024 survey of 500 finance professionals found 75% spend five to six hours or more a week recreating reports, and 82% name poor integration as the top reporting obstacle. That time is the lineage being rebuilt by hand, every week.

Permissions: for the user the agent will run as, list object- and field-level access from its permission sets and compare it with a human in the same role. Anything the agent can write that the human cannot is a finding, and Modify All Data is an automatic no. In HubSpot, the check is the user’s permission set and team assignment. Salesforce’s 2026 study found only 43% of organizations have formal data governance; this is the minimum version of it. Threshold: every field the agent writes has a named approver, and the board number traces to a report ID.

The four numbers below come from the vendors’ own surveys. We put them on the first slide of every data review, because they settle the question of whether the problem is real before anyone has to argue it.

Diagram: Nobody trusts the data - including the vendors’ own surveys.

What we actually see

A composite from the first two weeks of engagements, and an observation rather than a study; the figures are typical, not averaged. Duplicate rate on Contact sits above the 5% line more often than under it, and the company usually has Duplicate Rules configured in alert mode, which reps learned to click through years ago. Fill rate on Opportunity Amount and CloseDate is high, because validation rules enforce them. Fill rate on NextStep and the primary Contact role is low, because nothing does. Around a third of Contacts have no activity in a year. The owner-map exercise produces the same result nearly everywhere: two or three integrations nobody in the room can name an owner for, one of which is still writing to Lead records.

The definitions check is where the room goes quiet. Marketing’s lifecycle stages in HubSpot and sales’ stages in Salesforce were designed in different years by different people, and the sync between them maps MQL to a stage nobody uses. None of this is unusual, and none of it is visible from the dashboard, which is precisely why the dashboard looks fine.

Fix it once, then keep it fixed

A one-time cleanup decays at the same rate the data did. Six months after a deduplication project the duplicate rate is climbing again, because the import path that created the duplicates was never closed. The durable fix is structural, and most of it is configuration rather than software.

Switch Duplicate Rules on Account, Contact and Lead from alert to block for manual entry, and route the integration user’s inserts through a matching rule as well. In HubSpot, enforce deduplication on email for Contacts and company domain for Companies at import.
Move completeness into the process. Validation rules on Opportunity that require NextStep and a primary Contact role at the stage where the forecast starts counting the deal, not at Closed Won.
Write an archive policy. Contacts with no activity in eighteen months leave active lists and the agent’s scope; open Opportunities more than 60 days past CloseDate get closed or re-dated by their owner, monthly.
Publish the owner map, one name per object, integration and dashboard, and reassign every record owned by an inactive user this week.
Put the five definitions in one governed document with a version date, and make every dashboard filter that implements them a saved, named report.
Review the six numbers monthly, on one page, with the owners in the room. Once the reports exist the review takes twenty minutes.

Do that and the six checks stop being a project and become a dashboard, one that says, before any AI rollout, whether the foundation will hold. It is the least glamorous AI work there is, and every model the company deploys in the next five years reads from it. Build it once, on purpose, and each deployment gets cheaper and safer. Skip it, and each one starts by rediscovering the same defects.

Frequently asked questions

Why is CRM data quality important for AI?

Because an AI agent acts on the record it reads, without the pause a human would take. Duplicates send updates to the wrong Account, stale Contacts get confident outreach, and disagreeing definitions produce a forecast nobody can defend. Salesforce’s 2026 State of Data and Analytics found 89% of organizations with AI in production have already seen inaccurate outputs, and data is the usual cause.

How do you measure CRM data quality?

Six numbers on the objects the use case touches: duplicate rate on Account and Contact, fill rate on the eight to twelve fields the workflow reads, the share of active records with no activity in a year, records owned by inactive users, one written definition per key term, and whether a board number traces to a saved report. A Salesforce or HubSpot admin can produce all six in an afternoon.

What causes bad CRM data?

Decay does most of the damage: B2B contacts change jobs at roughly 2–3% a month by industry estimates, so a list untouched for a year is about a third wrong. The rest is process: imports without matching rules, integrations writing with no owner, free-text fields where a picklist belongs, and validation rules that arrive at Closed Won instead of at the stage that matters.

How often should CRM data be cleaned?

Continuously, through configuration, with a monthly review. Duplicate rules set to block, validation at stage transitions and an archive policy prevent most defects from entering. A twenty-minute monthly review of the six numbers catches the rest. A one-off annual cleanup decays at the same rate as the data and is usually back where it started within a year.

How does data quality affect AI outputs?

Directly and silently. AI inherits every defect in the fields it reads, and an agent does not flag doubt on a stale record; it acts. Validity’s 2025 survey found 76% of CRM users believe less than half of their CRM data is accurate and complete, which is the input most agents are being given today. Fixing the six checks first is cheaper than debugging an agent later.

What is an acceptable duplicate rate in a CRM?

Our working thresholds are 3% on Account and 5% on Contact before any agent reads or writes those objects. Above those lines, merge first: agents update the wrong record often enough that reps stop trusting the system. Measure it with Duplicate Record Sets in Salesforce or the Data Quality Command Center in HubSpot, and keep it there with duplicate rules set to block on manual entry and import.

NEXT STEP

See where your organization stands before the next AI purchase.

A short conversation is enough to tell whether the AI Readiness Health Check, 36 checks on your Salesforce or HubSpot org, is the right first step. No decks, no obligations.

book a strategy session

Read the 36 checks →

WRITTEN BY

Tal Oryon

MORE NOTES

Perspective

The world is changing. Is your organization part of it?

In one year the platforms your company runs on rebuilt their pricing, their products and their partnerships around agents. A note to executives on the one decision that is still theirs.

AI transformation

An AI transformation roadmap for a 300–2,000-person company (12 months)

Four quarters, one workflow at a time. The deliverables, the owners and the failure mode of each quarter, written for mid-market companies that cannot afford an enterprise timeline.

ACTIVEAPPS