How do I stop AI hallucinating about a deal?
Ground every claim in the actual record. Pass the timeline rather than a memory of it. Constrain the output to fields. Require citations to source lines. Route anything external through review. Hallucination drops with grounding; it never reaches zero.
The short answer
- Ground every claim an AI system makes about a deal in the actual record. Pass the real timeline, notes and messages into the call, rather than relying on the model's own memory or a vague description of what the deal is about.
- Constrain the output to specific fields where possible: last contact date, next step, stated objection. Not open-ended prose. A structured output against a schema gives the model less room to invent plausible-sounding narrative detail.
- Require citations back to the specific source line for any claim. A summary that says a meeting happened on a specific date should point to the note or message that establishes it. A reader can then verify the claim instead of trusting the tone.
- These techniques reduce hallucination a lot but do not eliminate it. Route anything seen outside the company, or acted on with real consequences, through a person, however well-grounded the drafting process is.
Why grounding works, and why it is not a full fix
A language model without access to the actual record is, in effect, being asked to remember details it never saw. It will produce something plausible rather than admit it does not know. Grounding fixes the most common version of this. Instead of asking "what happened in this deal", the system passes in the deal's actual timeline, notes and message history as part of the call. The model is summarizing real text rather than generating from a vague impression.
This closes the most common failure but not all of it. A model can still misread ambiguous phrasing in the real record. It can conflate two similar events that both appear in the timeline. It can draw a confident conclusion from an implication that was not stated. Grounding reduces how often a claim is entirely fabricated. It does not guarantee every claim drawn from real text is interpreted correctly.
Techniques and their effect
| Technique | Effect | Cost |
|---|---|---|
| Ground in the actual timeline | Removes fabrication from thin air; the model works from real text | More context per call, which increases cost somewhat |
| Constrain output to fields | Reduces room for invented narrative detail | Less flexible output; not suited to open-ended writing tasks |
| Require citations to source lines | Lets a reader verify a claim instead of trusting it on tone | Slightly more complex output format to build and check |
| Route external claims through review | Catches whatever grounding and citations miss before it is seen outside the company | A person's time on every reviewed item |
What this looks like for a deal summary specifically
A grounded, cited deal summary reads differently from an ungrounded one, even when both sound equally confident. The grounded version can point to exactly which note or message supports each claim it makes. If a claim looks off, checking it takes one click, not a full manual review of the deal's history. The ungrounded version offers nothing to check against. Its confidence is the only signal available, and confidence is not evidence.
Disclosure: SalesCrew is our product, and its record summaries and lead scores are built on this pattern. They cite the specific evidence, timeline entries, thread messages, knowledge-base sources, behind each claim, rather than producing unsourced prose. That does not make the summaries immune to error. That is why they remain something a person reviews rather than a silent, unattended write to the record.
A summary that invents a meeting date will be believed
Questions
- Can hallucination ever reach zero?
- No known technique gets a language model's output to zero risk of an invented detail. Grounding and citations reduce the rate a lot. The honest position is that review stays necessary for anything consequential, not that grounding solves the problem entirely.
- What does 'constrain the output to fields' mean in practice?
- Instead of asking an open-ended question like 'summarize this deal', ask for specific fields: last contact date, next step, stated objection. A model producing structured fields against a schema has less room to invent plausible-sounding narrative detail than one writing free-form prose.
- Why does a summary that invents a meeting date get believed?
- Because it reads exactly like every other correctly generated summary next to it. There is no visual or stylistic difference between a grounded claim and an invented one. That is why citations to the specific source line matter. They let a reader check the claim instead of trusting the tone.