Everyone stress-tests their Agentforce agent's prompts. Almost nobody stress-tests the data underneath it. That's backwards — because when an agent goes wrong in production, the prompt is usually fine. The data isn't.
Here's the failure that keeps us up at night, and it has nothing to do with the model:
A rep asks the agent, "What's Valley Health System worth to us?" There are three "Valley Health System" accounts in the org — a duplicate nobody cleaned up. The agent picks one, sums its opportunities, and answers with total confidence: "$140,000." The real number, across all three records, is $410,000. The agent wasn't wrong because it's a bad model. It was wrong because the data told it a lie, fluently.
That's the thing about an AI agent grounded in bad CRM data: the failure doesn't look like a bug. It looks like a fluent, plausible, incorrect sentence — the kind your customer believes and your rep repeats.
Why you can't catch this the way you catch normal bugs
A traditional bug throws an error. You see a red stack trace, you fix it. A grounding failure produces a confident answer that happens to be false. There's no error, no log line, no signal — just a customer who got quoted the wrong price, or emailed a contact who opted out, or was told a closed deal is still open.
And you can't find these by scrolling through records. The problems live in the relationships between records — a duplicate here, a stale timestamp there, a contradiction between two fields — exactly the connections a human would have to open each record to see, and exactly the connections an agent flattens into one confident paragraph.
So the only real defense is to measure your data's readiness before the agent goes live, not debug it after. After we kept hitting this on real Agentforce projects — the agent was fine, the data wasn't — we built GroundTruth to do exactly that. It scores your org across six dimensions. Here's the checklist, and you can run it as a mental audit today.
The six checks that catch a confidently-wrong agent
1. Completeness — are the fields the agent reads actually filled in? An empty Amount, a blank Industry, a missing Owner: to a human, "unknown." To an agent, often a confident zero or a guess. If your agent quotes numbers, every blank required field is a wrong answer waiting to happen.
2. Freshness — is the record current, or a fossil? An opportunity nobody has touched in 90 days still reads as "open" to an agent. It will happily tell a customer a deal is live that everyone internally knows is dead. Stale records are the most common source of "the agent said something that isn't true anymore."
3. Consistency — do the fields contradict each other? A Won opportunity with a null Amount. A closed case marked "new." When two fields disagree, the agent picks one and narrates it as fact — and it's a coin flip which one. Contradictions are where agents sound the most authoritative and are the most wrong.
4. Uniqueness — are there duplicates? This is the Valley Health System problem. Duplicates split a customer's history across records, so the agent sees a fraction of the truth and reports it as the whole. Deduplication isn't hygiene here — it's the difference between "$140K" and "$410K."
5. Integrity — are the values structurally valid? Free text where a picklist should be. An "Industry" of "helthcare." A status the schema doesn't recognize. The agent can't reason over a value it can't categorize, so it either ignores it or invents a meaning.
6. Exposure — is there compliance risk hiding in the data? A contact who opted out of email but is still marked contactable. A PII field visible org-wide. An agent acting on that data doesn't just give a wrong answer — it creates a breach. This is the check that turns a data-quality problem into a legal one.
Score it, don't guess it
You can eyeball a few of these. You can't eyeball all six across every record — that's the whole point. The reason grounding failures slip through is that no human is looking at all six dimensions on all your data at once.
That's the job we gave GroundTruth: it scans your org on-platform (nothing leaves Salesforce), rolls the six dimensions into a single Readiness Score, and — this is the part that matters for go-live — gives you a portable Benchmark score you can hold up as a gate. Below the line, you're not ready. It shows you the exact records failing each check, lets you fix, merge, or dismiss them in place, and can even sit behind your agents at runtime so an agent declines to answer on data that isn't ready.
The uncomfortable truth is that most orgs are closer to "confidently wrong" than they think, and they won't know until an agent says something false to a customer. The six checks above are how you find out first.
If you want to see where your own org lands, we'll run a free readiness assessment with you — the score, the six dimensions, and exactly what's between you and a safe go-live.