Salesforce data quality

A service agent goes live, a customer asks about returns, and the agent quotes — confidently, politely, with a citation — a policy that was retired eight months ago. Nothing malfunctioned. The retrieval worked. The knowledge article was simply stale, and the agent had no way to know.

This is the shape of most AI failures we see in Salesforce orgs. Not model problems. Data problems, surfaced at machine speed. And the scale of the underlying issue is well documented: in Salesforce’s own State of Data and Analytics research, data and analytics leaders estimated 26% of their data is untrustworthy, and in the second edition of that research, 84% said their data strategy needs a full reset before their AI ambitions can succeed. Those aren’t laggards. Those are the people closest to the data.

This guide covers the four quality dimensions that actually matter for AI, what the native tooling you already own can and can’t do, how to measure where you stand, a remediation sequence that won’t stall out, and exactly how each class of data failure shows up in AI features.

Why AI amplifies whatever data quality you already have

Dashboards are forgiving. A human reading a report mentally discounts the duplicate account, skips the contact with no email, and knows the “Industry” field stopped being maintained in 2022. Bad data in a report costs you accuracy; a person filters the rest.

AI removes the filter. Agentforce grounds its responses in your CRM records, knowledge articles, and Data 360 objects through retrieval — it answers from what’s there, not from what should be there. Einstein’s predictive features learn from your historical field values, so skewed history produces skewed scores. And Data 360’s identity resolution builds unified profiles by matching records across sources; Salesforce’s own AI data readiness framework lists uniqueness, recency, and correctness among the attributes that decide whether an agent helps or embarrasses you.

The difference from every previous Salesforce feature is that agents act. They send the reply, update the record, trigger the flow. A quality problem a dashboard could tolerate becomes a business action executed without a human pause. The industry numbers back the intuition: nearly nine in ten organisations running AI in production have seen inaccurate or misleading outputs, and Gartner has estimated poor data quality costs organisations an average of $12.9 million a year — before any of it was wired into autonomous agents.

AI is a multiplier. It multiplies clean data into compound value and dirty data into compound liability. It fixes nothing.

The four data quality dimensions that matter for AI

Salesforce’s data quality module on Trailhead breaks quality into characteristics like completeness, accuracy, duplication, and age — and cites research that inaccurate data costs the average company 12% of revenue. For AI work specifically, we find four dimensions carry almost all the weight. Duplicates cut across two of them, which is partly why they’re so damaging.

Dimension	What bad looks like	How it surfaces in AI	First thing to measure
Accuracy	Wrong emails, retired policies, “Calif” vs “CA”	Agents cite wrong facts with full confidence; actions target the wrong record	Bounce rates, failed-contact rates on key channels
Completeness	Blank Industry on 60% of accounts, contacts with no role	Thin grounding, generic summaries, predictions built on too few signals	% populated on the 10–15 fields your processes actually read
Freshness	Records untouched for 18 months, knowledge nobody reviewed	Agents answer from the org as it was, not as it is	% of records with no modification or activity in 12 months
Consistency	Duplicates, free-text picklist chaos, conflicting values across systems	Fragmented customer views; identity resolution merges or splits the wrong profiles	Duplicate record inflow; distinct-value counts on fields that should be standardised

Notice what’s not on the list: volume. Orgs often frame AI readiness as “do we have enough data?” In practice the binding constraint is nearly always whether the data you have is true, filled in, current, and shaped the same way twice.

The native tooling you already own — and where it tops out

Before buying anything, exhaust what’s in the platform. Admins already own three quality mechanisms, and each one has documented ceilings worth knowing before you bet your cleanup on them.

Duplicate rules and matching rules

Matching rules define what counts as “the same record”; duplicate rules decide what happens when a match is found — alert the user, block the save, or quietly report it. Salesforce ships standard rules for accounts, contacts, and leads, and the documented limits are five active duplicate rules per object, three matching rules per duplicate rule, and five active matching rules per object. Under the hood, match keys narrow each comparison to roughly the 100 most likely candidates before the full matching equation runs.

Those defaults are genuinely good, and most orgs should switch them on today. But three ceilings matter at scale:

Dirty orgs can be too dirty to protect. Salesforce blocks matching-rule activation when too many records share identical match key values — beyond a 2% threshold by default, raisable by Support to a hard limit of 5%. If your data is bad enough, the tool designed to prevent duplicates refuses to start until you’ve cleaned some up first.
Bulk paths bypass the experience. Per the duplicate rules considerations, rules don’t run at all on Quick Create, standard lead conversion, undelete, or records synced by Einstein Activity Capture — and for records arriving through import tools or the API, no alert is shown. A blocking rule turns into flat save errors in your integration’s job results; an alert-only rule has nobody to alert. Your nightly sync is exactly where duplicates enter, and it’s exactly where the guardrails are weakest.
They prevent; they don’t cure. Duplicate rules fire on create and edit. Every duplicate already in the org stays put. Native duplicate jobs can scan existing accounts, contacts, and leads — but only in Performance and Unlimited Editions — and merging remains a manual, three-records-at-a-time affair. For a backlog in the tens of thousands, you’ll be reaching for Data Loader gymnastics or a dedicated dedupe tool.

One more structural gap: leads and contacts are the same human beings, but rules operate per object. Preventing lead-to-contact duplication takes deliberate cross-object rules, and most orgs never set them up.

Validation rules and required fields, without form fatigue

Validation rules stop bad data at the door: a record that violates the formula doesn’t save. They’re the cheapest quality control on the platform, and also the easiest to overdo.

The failure mode is form fatigue. Every required field is a tax on every save, and users pay taxes creatively — “aaa” in the phone field, 01/01/1900 in the date, “Other” in the picklist. A required field that users routinely fill with garbage is worse than an optional one, because it looks complete to every report and every agent grounding on it. This is where most quality initiatives go wrong: they mistake enforcement for quality.

Our working rules for keeping enforcement honest:

Require a field only if something consumes it. A routing rule, a forecast report, an agent instruction — a named consumer. “Might be useful someday” is how orgs end up with layouts users resent.
One rule, one intent. A mega-rule enforcing three requirements gives one confusing error. Split them so users learn exactly what to fix.
Guard against the integration. Rules fire on saves from flows, imports, and APIs too. Scope rules with ISNEW() and ISCHANGED() so the 3 a.m. sync updating an unrelated field doesn’t die on a rule meant for salespeople.
Write errors for humans. “Discount above 20% needs manager approval” beats “Validation failed: VR_OPP_07.”
Prefer picklists to free text anywhere consistency matters — state and country picklists alone eliminate a whole category of “Calif” problems.

Measuring Salesforce data quality: completeness, staleness, ownership

You can’t remediate what you haven’t measured, and gut feel is reliably wrong in both directions. Three measurements, all buildable with native reporting, give you a working baseline in a day or two.

Field completeness. Pick the 10–15 fields per core object that your processes genuinely read — not every field, the load-bearing ones. Build a report per object filtered to blanks, or add a simple formula field that scores each record by how many key fields are populated and report on the average. The trend matters more than the snapshot: completeness that’s 62% and rising beats 71% and falling.

Staleness. Define “stale” per object — for contacts, no modification and no activity inside a rolling year is a reasonable start — and count it. A quick SOQL check makes the point before you build the report:

SELECT COUNT()
FROM Contact
WHERE LastModifiedDate < LAST_N_DAYS:365
  AND (LastActivityDate < LAST_N_DAYS:365 OR LastActivityDate = NULL)

In mature orgs this number is regularly a third or more of the database. That’s not automatically a crisis — but every one of those records is a candidate for an agent to retrieve and treat as current.

Duplicate inflow. Set your duplicate rules’ report option so matches land in duplicate record sets even when users save anyway, then build a custom report type on Duplicate Record Sets. Grouped by rule, it tells you both how fast duplicates are entering and which matching rules are earning their keep.

The last measurement isn’t a report — it’s a name. Every data domain needs an owner accountable for its numbers, because “everyone’s responsible for data quality” means nobody is. If you want a structured first pass across these checks, our org health scorecard walks through them.

A three-stage remediation sequence: stop the bleeding, clean in batches, monitor

The classic cleanup failure is starting with the backlog. Teams spend a quarter merging historical duplicates while the integration mints new ones nightly, and the metrics never move. Sequence the work the other way.

Stop the bleeding. Activate duplicate rules (alert plus report) on accounts, contacts, and leads. Add validation on the handful of critical fields — no more. Standardise the highest-chaos fields with picklists. Then audit the paths that bypass the UI: give integration and import flows their own matching pass, because that’s where volume enters. Nothing else works while the inflow continues.
Clean in batches. Work the backlog in waves — by owner, territory, or record age — so no team wakes up to a transformed database. Export a backup before any merge, and decide survivorship rules up front: which record wins, field by field, when values conflict. Triage stale records into enrich, archive, or delete; keeping everything “just in case” is a decision to ground your AI on it. Where native merging runs out of road, this is the one place third-party dedupe tools usually pay for themselves.
Monitor. Put completeness, staleness, and duplicate inflow on a dashboard someone actually reviews. Treat a spike in duplicate record sets like an incident, because it usually means an upstream process broke. Schedule a deeper quarterly pass. Quality is a control loop, not a project with an end date.

The stages compound. Prevention makes cleanup finite; cleanup makes monitoring meaningful; monitoring tells you when prevention has sprung a leak.

How data quality failures show up in Agentforce, Einstein, and Data 360

Each dimension fails in AI in its own recognisable way. Once you’ve seen the mapping, agent misbehaviour stops being mysterious.

Stale content becomes confident misinformation. Retrieval ranks by relevance, not by truth-as-of-today. An expired policy in a published knowledge article is, to the retriever, a perfectly good answer. Freshness has to be engineered — review cadences, expiry dates, archiving — because the agent can’t infer it.
Duplicates fragment or poison the customer view. Split a customer across three contact records and the agent grounds on whichever one it retrieves — a third of the history, no open case, wrong tier. Worse is the opposite error: merge two records that were never the same person, whether by hand or via identity resolution matching on shared junk values, and the agent now confidently discusses someone else’s account. Bad merges are far harder to undo than duplicates.
Incomplete fields become garbage summaries and skewed scores. Ask an agent to summarise an account where Industry, role, and last-touch fields are blank and you get fluent emptiness. Predictive scoring trained on records where fields were only filled for deals that closed learns the filling, not the winning.

None of this is exotic. It’s the same list from the production Agentforce lessons we keep relearning: grounding beats prompting, and grounding is only as good as the records behind it. When an agent misbehaves, our first question is no longer “what’s wrong with the topic?” It’s “what did it retrieve?”

The unglamorous multiplier

Nobody demos a merged duplicate. Data work has no launch moment, no screenshot for the board deck, and it will lose every prioritisation argument against a shiny agent pilot — right up until the pilot ships and starts quoting the wrong policy.

But the economics run the other way. The same Salesforce research that found widespread inaccurate AI outputs also found 86% of leaders saying AI results depend directly on how well data is governed and maintained, with organisations running formal data quality processes roughly twice as likely to report strong AI ROI. Every hour spent on deduplication, completeness, and freshness is inherited — free — by every AI feature you turn on afterwards. The reverse is also true: every quality problem you carry into an AI rollout gets executed, at machine speed, with your brand on it.

So run the sequence. Measure the four dimensions this week. Turn on the native guardrails and learn their ceilings. Clean in waves, monitor the inflow, and give the numbers an owner. It’s six weeks of unglamorous work that quietly decides whether the next three years of AI investment compounds or corrodes. If you’d rather not sequence it alone, that prioritisation is exactly what our data and AI strategy work exists for.

Understanding the basics

What are the main dimensions of data quality in Salesforce?

Salesforce’s Trailhead guidance breaks data quality into characteristics including accuracy, completeness, consistency, duplication, and age. For AI readiness, four dimensions carry most of the weight: accuracy (values are true), completeness (load-bearing fields are populated), freshness (records reflect current reality), and consistency (the same entity is represented one way, once). Duplicates damage both accuracy and consistency, which is why deduplication tends to deliver the most visible improvement.

How does duplicate management work in Salesforce?

Two components work together: matching rules define what counts as a duplicate, and duplicate rules decide what happens on a match — alert the user, block the save, or report it. Limits include five active duplicate rules and five active matching rules per object. Rules fire on record create and edit, but interactive alerts don’t appear for imports or API loads, and existing duplicates require duplicate jobs (Performance and Unlimited Editions) or third-party tools.

Why does AI make CRM data quality problems worse?

Humans reading reports instinctively discount bad records; AI features can’t. Agents ground their answers in whatever they retrieve, predictive models learn from whatever history exists, and identity resolution unifies profiles based on whatever values match. Each mechanism treats your data as true. Because agents also act — replying, updating, escalating — a data error stops being a reporting inaccuracy and becomes a business action performed at machine speed.

Staring at a decade of duplicates between you and an AI rollout? Talk to us — making orgs agent-ready is what we do.

Salesforce data quality: a practical guide to getting AI-ready