Salesforce - Priscilla Hong

When an Agentforce agent gives a wrong answer, the reflex is to blame the prompt. Teams spend weeks wordsmithing instructions while the real problem sits untouched in their data: duplicate customers, five spellings of the same account, knowledge articles from 2023, records nobody trusts.

Here's the uncomfortable truth of the agentic era: an AI agent is a mirror for your data quality. A human support rep eyeballs three conflicting records and quietly picks the plausible one. An agent retrieves all three, trusts them equally, and confidently gives you a wrong answer grounded in your own CRM. The scale of this is enormous — Salesforce ingested 112 trillion records into Data 360 in FY2026 alone, up 114% year over year. Records are flowing in; quality is what decides whether they help or poison your agents.

These are the five data problems that break agents most often — and what actually fixes them.

1. No unified customer profile

The symptom: the agent tells a customer their order is "in transit" and "not found" in the same conversation.

Why it breaks agents: the same person exists as three separate records from three source systems. Humans deduplicate instinctively; retrieval does not.

The fix: identity resolution and harmonization — ingesting sources into Data 360, mapping them to a standard model, and resolving identities into unified profiles. This is the single highest-leverage data investment you can make before deploying any agent.

2. Stale and duplicate knowledge

The symptom: the agent quotes a return policy that changed six months ago.

Why it breaks agents: grounding retrieves whatever is closest, not whatever is current. Old articles often match retrieval queries better than new ones.

The fix: a content audit before launch — archive or expiry-date anything outdated, and assign ownership for keeping grounding content current. Treat knowledge articles like code: versioned, reviewed, and retired.

3. Grounding by stuffing instead of retrieving

The symptom: the agent works in demos with two test records, then degrades badly with real volumes.

Why it breaks agents: teams paste context directly into the prompt template. That works for one tiny scenario and collapses at scale — the agent can't tell which context matters, and costs explode.

The fix: retrieval-based grounding on unified data, so the agent pulls only the relevant records per request. This is the core architectural idea behind Data 360's role in Agentforce — and a heavily tested concept on the Data 360 Consultant exam.

4. No data boundaries

The symptom: the agent happily summarizes a customer's payment disputes — to a different customer.

Why it breaks agents: data spaces and permission boundaries weren't configured, so the agent sees everything the org sees.

The fix: data spaces to segment what different agents and users can reach, plus the Einstein Trust Layer for masking PII before anything leaves your org. Governance isn't a launch-afterthought anymore — it's a launch requirement.

5. No way to catch drift

The symptom: the agent was accurate at launch; three months later, answers quietly got worse and nobody noticed.

Why it breaks agents: your data changed under the agent — new products, new policies, restructured records — and there's no test suite to catch it.

The fix: a golden-set evaluation suite (50+ known-good question/answer pairs) that you re-run on a schedule and after every significant data change. Admins who bring eval discipline to an organization become genuinely hard to replace.

The pre-launch checklist

Every one of these problems lives in the data layer, and every one is fixable before your agent ever talks to a customer. If you want to go deeper on ingestion, harmonization, identity resolution, and activations — the full grounding stack — the Data 360 Consultant practice tests are built around exactly these exam domains.

Bottom line: before you spend another hour rewriting prompts, go look at your data. Nine times out of ten, that's where the fix lives.