Our own research engine taught us a hard lesson about AI for sales prospecting. Before it could consult a source, it said it had found evidence of budget authority for 100% of the contacts it researched. Give it tools to check, and that figure dropped to 84%. For these contacts, that 16-percentage-point gap in claimed budget-authority evidence is roughly one in six. The comparison covers this budget-authority check; it does not measure accuracy across every field or establish a universal hallucination rate. It had invented the gap. A better model isn't the remedy for an AI that can make unsupported claims. You need a system that prevents the AI doing the research from judging its own results.
The engine was behaving as designed: it supplied an answer for every contact. We hadn't made "I don't know" an acceptable response. That omission, rather than a broken AI, was the problem.
Why do AI hallucinations happen in sales prospecting?
Most prospecting tools expect the AI to answer every question and never explicitly allow it to decline. That expectation helps explain the hallucinations. In its 2025 announcement of research with Georgia Tech, OpenAI explained that models graded only on accuracy are "encouraged to guess rather than say 'I don't know.'"
Our engine kept demonstrating the problem. It assigned a contact "medium" budget authority using the job title we had supplied as its evidence and our supplied profile link as its source. When we asked about the substance of a profile, it reported that the "About section appears empty" even though it had never opened the profile.
The pattern extends beyond our engine. Walters and Wilder examined 636 citations ChatGPT generated for literature reviews. Fabrications accounted for 55% of the citations from GPT-3.5 and 18% of those from GPT-4.
Why does fabricated AI output look better than honest output?
A fabricated result can fill every field, which makes it appear more complete. Our faulty engine populated 100% of the budget-authority fields; the version that checked honestly populated 84%. Once it could look for activity, its activity findings dropped from 95% to 65%. For the remaining contacts, there was no activity to find.
A completeness score would have led us to choose the engine that invented answers. AI creates that trap in go-to-market work: confidence looks like accuracy, and a dashboard can't distinguish the two. These percentages describe fields our engine filled. They are not controlled estimates of overall accuracy or financial return.
How do you make AI for sales prospecting trustworthy?
Growth Orbit uses four layers to make AI for sales prospecting trustworthy, all built around a single rule: the AI producing the work cannot grade it. Use the questions behind these layers to evaluate any AI system you plan to build or purchase.
1. Ground the work before it starts
Limit research to your market. The AI should investigate only businesses included in your clean TAM.
Make evidence a condition of scoring. Code automatically assigns zero weight to a finding that lacks a source link. The AI may disregard a prompt, but it can't bypass that code-enforced rule.
Allow uncertainty as a third response. Alongside yes and no, accept "couldn't establish." Our tool now states the distinction explicitly: "This is undetermined, NOT evidence of a thin or low-substance profile."
Keep observations and interpretations separate. Record what was found and where it came from without blending it into an explanation of what it might signify.
2. Make something independent check it
Give each researcher an AI fact-checker. The practical example below shows how we use that pairing.
Reject invented figures. If commentary from the AI includes a number absent from the verified source, a check fails the entire build. Interpretation is permitted; invention isn't.
Review the reviewers, too. After cheaper AI models audited our software, a stronger independent model checked their conclusions again. It rejected two of their claims and discovered a defect every one of them had missed.
3. Prove it before you accept it
Design checks that are capable of failing. We use this rule: "A check that cannot fail is a task that cannot be verified."
Establish performance on answers you already know. Before trusting the system with new figures, we had it reproduce 333 audited figures drawn from public financial filings.
Verify writes by reading the result back. An automated push of ours failed without an alert for about 12 days, even as our ledger reported 100% completion. That experience established our distinction between completing a task and accepting its result.
4. Roll it out deliberately
Evaluate each task using your own data. A lower-priced model that had passed our account-research validation cost a third as much for contact research, yet lost 15 points on the most important signal. We switched back and spent about $49 more for that run.
Start with a small pilot and withhold unfinished work. For one delivery, we released 75 verified contacts while holding back 182 contacts that weren't ready.
How does AI fact-checking work in practice?
An AI fact-checker is a separate AI charged with testing the first AI's claims against the sources. Anything it can't establish gets rejected. Growth Orbit uses this approach for a trade publisher's weekly market-intelligence brief: each AI researcher works with an AI auditor. The auditor opens the sources again, checks every quotation word for word, and tests whether the source supports the headline, summary and date. Its standing instruction reads "Default REJECTED when unsure."
| In one weekly run | Items |
|---|---|
| Confirmed by the auditor | 321 |
| Rejected | 14 |
| Unverifiable | 19 |
Roughly one item in ten was kept from readers because the fact-checker couldn't substantiate it. Publication is reserved for confirmed items. Rejected items remain on file with the reason for rejection available to review.
Does all this checking slow AI down too much?
The checks hold up material that shouldn't be released. That's their purpose. The practices described here use AI tools available today, and the checks are mostly automated. Their cost is small compared with acting on fabricated findings, using those findings in outreach, or presenting them to the board.
Scale makes verification more necessary. Gartner forecasts ten AI agents for every seller by 2028, while fewer than 40% of sellers will report a productivity improvement from those agents. Expanding AI without expanding checks increases the volume of untested claims.
Frequently asked questions
How do you stop AI hallucinations in sales prospecting?
Let the AI answer "I don't know," and use code to require a source for each finding. Before release, a separate auditor should check those sources again. These protections work together; none provides enough protection by itself.
Should you trust an AI tool that never says "I don't know"?
No. An AI research tool that answers everything will sometimes be guessing. Ask its vendor how frequently it responds "couldn't establish," and how the vendor tested error rates on data comparable to yours.
Is a more expensive AI model always more accurate?
Price alone doesn't establish accuracy in either direction. Compare models on your data for each individual task. Success with one research assignment doesn't guarantee accurate results on another, regardless of the model's reputation.
Growth Orbit was founded by Steve Schilling. Read Is Your CRM Data Ready for AI in 2026? for the consequences of letting AI read CRM records, and 5 Questions Every Sales Metric Must Answer for the questions to put to every number.