Most Agentforce wrong answers trace back to the data the agent retrieves, not the model that generates the text. Agentforce agents read your Salesforce records, ground their responses in what they find, and act on it — so when the underlying records are incomplete, inconsistent, stale, or duplicated, the agent produces inaccurate responses from accurate retrieval of bad data.
This is a troubleshooting guide. Each section maps a failure symptom you can observe to the data problem causing it, the DQS metric that diagnoses it, and the remediation step that fixes it. If you are planning a deployment rather than debugging one, start with the Agentforce data quality guide and the data readiness checklist.
Why Does Agentforce Give Wrong Answers at All?
An agent does not invent facts in a vacuum. It retrieves records that match a request, reads the fields it has access to, and assembles a response from that context. The model is consistent; the data underneath it is not.
What looks like a hallucination is often the agent faithfully reporting what your records say. A blank field gives the agent nothing to ground on. Four spellings of the same country give it four answers. A duplicate record gives it a second version of the truth. Fix the data and most of these failures disappear without touching the agent configuration.
The six failure modes below cover the data root causes you can measure and correct in Salesforce.
Why Are Agentforce Responses Vague or Generic?
Symptom. The agent answers a specific question with a generic reply. A user asks why a Case was escalated and gets “the issue was reviewed and addressed” instead of the actual reason.
Root cause. Incomplete records. When the agent retrieves a Case with an empty Description, Comments, or Resolution field, it has no context to ground a specific answer. It falls back to a vague summary because there is nothing concrete to read.
Diagnosis. Completeness Rate, broken down by field, shows how many records in the agent’s scope are missing the fields it needs for context. A high overall score hides the problem if the empty fields are the ones the agent reads. Check Completeness on Description, Notes, Comments, and Resolution specifically.
Fix. Raise Completeness on the fields the agent uses for responses. Backfill from a trusted source, make critical fields required at entry, and rerun the scan to confirm. See Completeness for the field-level diagnostic flow.
Why Does Agentforce Contradict Itself Across Conversations?
Symptom. The same question returns different answers in different sessions, or two users get conflicting facts about the same account.
Root cause. Inconsistent values and duplicate records. When the Country field holds “US”, “USA”, “United States”, and “U.S.A.”, the agent treats them as four distinct values and segments its answers by whichever it retrieved. When two Account records describe the same company with different data, the agent answers from whichever one it found first.
Diagnosis. Conformance Rate measures how many values match your canonical set, so a low rate exposes fragmented picklist and reference fields. Duplicate Rate measures how many records are duplicates of another, which is the second half of this failure. Run both across the objects the agent reads.
Fix. Standardize values to a controlled set and merge duplicates. Use Import from Field in the Definition Builder to discover the variants already in your data, define the canonical value, and normalize. Then merge duplicate records and configure matching rules to block new ones. See Consistency and Uniqueness.
Why Does Agentforce Return Outdated Information?
Symptom. The agent reports a closed deal as open, names a contact who left the company a year ago, or quotes a price from a superseded record.
Root cause. Stale records. The agent has no concept of “old.” It reads the most relevant record by content match, regardless of when the data was last touched. A record that has not been updated since the facts changed feeds the agent obsolete context.
Diagnosis. Timeliness Rate measures how many records fall within a freshness threshold you define on a date field such as LastModifiedDate or a business date. A declining Timeliness Rate on the agent’s objects means the agent is increasingly likely to surface outdated facts. See Timeliness.
Fix. Flag records past your freshness threshold and route them to re-engagement or archival workflows. Archive or exclude records that should no longer inform answers, so the agent stops retrieving them. Schedule the Timeliness scan to repeat so staleness stays visible.
Why Does Agentforce Surface Sensitive Data It Should Not?
Symptom. A generated response includes a Social Security Number, a credit card number, or another piece of PII that has no business appearing in agent output.
Root cause. PII sitting in free-text fields. Email-to-case, chat transcripts, and manual notes capture personal data into Description and Comments fields. When the agent retrieves one of those records, the PII enters the AI context and can appear in a response.
Diagnosis. PII Exposure Rate measures how many records in scope contain detectable PII across your text fields. Scan every text field the agent can read, not the fields named for PII, because the exposure hides in general-purpose comment and description fields. See PII Detection.
Fix. Identify findings with a PII scan, then mask, delete, or restrict the confirmed matches. Configure per-field pattern overrides for fields where a matching pattern is expected and legitimate, so you act on real exposure. Rerun the scan to validate, and review the PII compliance guide before deployment.
Why Can’t Agentforce Find the Right Record?
Symptom. The agent reports it cannot find a customer who exists, or retrieves the wrong record entirely. A user searches by email and the agent returns nothing.
Root cause. Invalid formats and unsearchable values. An email stored as “john [at] company dot com”, a phone number with inconsistent punctuation, or an ID in the wrong format does not match the lookup the agent runs. The record is there, but the agent cannot resolve it.
Diagnosis. Validity Rate measures how many values pass the format rules you define on structured fields such as email, phone, and identifiers. A low Validity Rate on a key lookup field predicts retrieval failures. See Validity.
Fix. Mass-correct malformed values on the fields the agent searches, then add a validation rule to enforce the format on new records. Focus on the fields that drive lookups and matching, where a format error breaks retrieval outright rather than just degrading a response.
Why Do Agentforce Answers Get Worse Over Time?
Symptom. The agent performed well at launch and degrades over weeks. Vague answers, contradictions, and stale facts reappear after a clean deployment.
Root cause. Data quality regression with no monitoring. Records you cleaned before launch decay as users enter new data, integrations write new records, and the agent itself updates fields. A one-time cleanup is a snapshot, and the snapshot ages out.
Diagnosis. No single metric catches regression. You catch it by trend. Schedule recurring scans across all six dimensions and watch each rate as a line over time, so a dip surfaces within days of a new integration or process change introducing bad data. See Understanding Results.
Fix. Configure scheduled, recurring scans on the objects in the agent’s scope and track metric trends in Insight Studio. Set a cadence by risk: weekly PII scans on high-volume text fields, monthly Completeness and Consistency, a quarterly full scan. Catch regression on the dashboard before it reaches an answer.
Symptom-to-Fix Reference
| Symptom | Root cause | DQS metric | Fix |
|---|---|---|---|
| Vague or generic responses | Incomplete records (empty Description, Notes) | Completeness Rate | Backfill context fields; require them at entry |
| Contradictory answers across conversations | Inconsistent values and duplicate records | Conformance Rate, Duplicate Rate | Standardize values; merge duplicates; add matching rules |
| Outdated information | Stale records | Timeliness Rate | Flag and archive records past a freshness threshold |
| Surfaces sensitive data | PII in free-text fields | PII Exposure Rate | Scan text fields; mask, delete, or restrict findings |
| Cannot find the right record | Invalid formats, unsearchable values | Validity Rate | Mass-correct formats; add validation rules on lookup fields |
| Answers degrade over time | Data quality regression, no monitoring | All six (trend) | Schedule recurring scans; track trends in Insight Studio |
How to Run This Diagnosis
Run a DQS scan across every object the agent reads and review the six metrics together. The scan runs entirely inside Salesforce, so no data leaves the org. Match each weak metric to its symptom in the table above, remediate in priority order, and schedule the scan to repeat so the fix holds. For the full preparation workflow, see the data cleanup guide for Agentforce.
Next Steps
- Agentforce Data Quality: the full preparation guide and timeline
- Agentforce Data Readiness Checklist: pre-deployment verification
- Salesforce Data Cleanup for Agentforce: the remediation workflow
- Agentforce Data Quality FAQ: common questions answered
- The Five Dimensions: overview of every data quality dimension
- AI Readiness Assessment: get your current readiness score