Key takeaways
- AI assistants in Salesforce work from your CRM data, so duplicates, gaps and stale records lead to wrong answers.
- Measure four things first: duplicates, completeness of key fields, staleness and consistency.
- Review permissions. An AI feature should never surface data a user isn't allowed to see.
- Start with one narrow use case on your cleanest data, then expand.
AI features are now part of most Salesforce conversations. Agentforce agents, generated email drafts, opportunity summaries and predictive scoring all have one thing in common: they work from the data in your org.
That makes data quality a business question, not only an admin's hobby. If a customer has three duplicate accounts, an AI summary may describe the wrong one. If half the opportunities have no next step, an agent can't tell a rep what to do next.
This checklist helps you find out whether your data is ready, and what to fix first.
Step 1: pick the use case
Data quality is always relative to a purpose. Before you measure anything, write down the first AI use case you are considering, for example:
- summarising an account before a sales call
- drafting follow-up emails after meetings
- answering customer questions from case history and knowledge articles
- prioritising leads
Then list the objects and fields that use case depends on. You only need to measure those.
Step 2: measure four dimensions
| Dimension | What to measure | Simple target to start |
|---|---|---|
| Duplicates | Share of accounts, contacts and leads with a likely duplicate | Under 5% |
| Completeness | Share of records with key fields filled | Over 85% on fields the use case needs |
| Freshness | Records not updated or opportunities past their close date | Under 10% stale |
| Consistency | Free-text values where picklists should be, mixed formats | Standard values on all key fields |
Build these as reports and put them on one data health dashboard. You'll use it to track progress.
Step 3: fix the biggest gaps
Duplicates. Activate matching and duplicate rules on accounts, contacts and leads. Merge existing duplicates, starting with accounts that have open opportunities.
Completeness. Make key fields required at the right moment, for example next step when a deal reaches a later stage, rather than at creation. Use dynamic forms so the right fields appear when they are needed.
Freshness. Run a scheduled Flow that flags opportunities with past close dates and asks owners to update or close them.
Consistency. Replace free-text fields with picklists where the values are predictable, and clean existing values with a one-off import.
Step 4: review permissions and sharing
AI features generally respect the running user's access, but they make it much easier to find information. Before turning anything on:
- check the organisation-wide defaults and sharing rules on the objects involved
- review field-level security on sensitive fields, such as margins, personal data or HR information
- confirm that inactive users and integration users have only the access they need
If a user shouldn't see a field today, an AI summary should not show it either.
Step 5: check your processes and knowledge
AI agents that answer questions need source material. If your support team doesn't maintain knowledge articles, or your sales process lives only in people's heads, write the basics down first. A short, current article is more useful than a long, outdated one.
Step 6: start small and measure
Choose one use case, one team and one month. Compare the results with the baseline you measured in step 2. Ask users where the AI was wrong and trace each error back to its data cause. Then expand.
The good news
The work that makes data ready for AI also makes your reports more accurate and your reps faster today. Even if you postpone AI features, a cleaner org pays off immediately.
If you'd like a measured baseline and a clean-up plan, the data quality service includes a data audit, duplicate rules and a data health dashboard you can keep using.
Frequently asked questions
Do we need Data Cloud to use AI in Salesforce?
It depends on the feature and your edition, and Salesforce's packaging changes often. Check the current requirements for the specific feature with Salesforce or your account executive. The data quality work in this checklist is useful whichever option you choose.
How clean does our data need to be?
It does not need to be perfect. It needs to be good enough on the specific objects and fields the AI use case relies on. Start by measuring those, set targets and fix the biggest gaps first.