A September 2025 paper by Adam Tauman Kalai and co-authors describes hallucinations as plausible but incorrect statements a language model produces instead of admitting uncertainty, and argues that training and evaluation reward guessing over saying "I don't know" (Why Language Models Hallucinate, arXiv:2509.04664). Sales teams now meet that behaviour from two sides: the AE who would rather guess than pause, and the AI note-taker that turns a hedged maybe into a firm commitment in the follow-up summary. What makes these mistakes expensive in a demo is the paper trail.

Which invented answers cost the most

The useful measure is how hard a claim is to unwind once it has left the call, and the worst kind sounds like a contract term. Say an AE is asked whether customer data can stay in EU regions, hasn't checked, and answers "yes, everything is hosted in Frankfurt." The note-taker writes it into the CRM as a commitment, and procurement later quotes it back in a redline. Correcting it means admitting the spoken answer was wrong just as legal starts reading, which is about the worst moment to do it. Data residency, uptime percentages and audit scope all carry that kind of weight, and so does anything else a lawyer could paste into an agreement. The follow-up is usually whether backups stay in-region too, covered in cloud vs on-prem buyer questions.

An invented feature fails differently. A native Salesforce sync that doesn't exist, or a SCIM endpoint that hasn't shipped, gets caught the moment an IT evaluator tries to configure it. It kills fewer deals than a contractual yes. What it burns is the technical champion who staked their reputation on your answers.

Commercial slips usually sit quietly until the order form. "SSO is included on Pro" when it's Enterprise-only belongs here, as does a per-seat number nobody approved. So do the competitive lines reps reach for under pressure, like a G2 score misremembered upward or "we're the only vendor that does this." A follow-up email can retract most of these, and the deal often survives, though finance and the buying committee will check the AE's next claim a little harder.

Then there are the errors the AI summary introduces by itself. The TofuEval benchmark, published at NAACL 2024, found that large language models produce significant amounts of factual error when summarising dialogues, across model sizes (TofuEval, ACL Anthology). On a sales call that can mean the bot credits the AE with a promise the SE floated hypothetically, or rewrites "we could look at that next quarter" as "vendor committed to ship X." Accuracy figures that note-taker vendors publish tend to describe transcription. Check whether yours says anything about action items or commitment extraction, because that's the part that lands in the account record.

When the buyer forwards the transcript

When that email arrives, usually a security analyst or procurement lead pasting a line from the summary and asking you to confirm it, go to the recording or the raw transcript before you look at the AI summary again. Find the exact exchange and note who spoke. An AE's firm yes and an SE's hedged "we can" are different situations, and a product manager's roadmap aside is different again. That difference is often the whole case.

Then check the claim against what your company already has in writing, such as the trust center, the current docs or the order form language. When the spoken answer matches, reply with the link and leave it there. When it doesn't, say so in one sentence: what was said on the call was incorrect, and here is the accurate position with the page that documents it. Buyers tend to forgive a corrected fact. A second invented fact stacked on the first is much harder to come back from.

Before the demo

Build a short "true for this account" sheet with two columns. One holds claims that are safe to state for this plan and deployment; the other holds claims that must be deferred. Put the SSO tier on it, along with hosting regions, the uptime commitment and any beta features the buyer has already asked about. If your demo environment uses feature flags, note which ones are on for this tenant. A demo can show a flagged feature the buyer's plan doesn't include.

If an AI assistant will sit on the call or draft the follow-up, give it the same sheet as context and tell it not to invent roadmap items. Models graded only on accuracy learn to guess, as Kalai et al. argue, so a model with gaps in its context will tend to fill them. Keep the sheet short enough that an AE will actually open it.

When a question lands that isn't on the safe side of the sheet, say so on the call: "I don't want to guess on residency for backups. I'll send you the written answer." Buyers who care about compliance expect that pause.

After the call, read the AI summary against the recording before anything syncs to HubSpot or Salesforce. Look for invented action items, wrong plan names and firm answers on security or hosting topics, and correct them in the note-taker or the CRM record itself. A fix posted in a side Slack thread never reaches the CRM.