AI output can sound equally confident whether it is quoting your notes, summarizing a current source, connecting two clues, or filling a gap. In sales, those are not equivalent. A fabricated stakeholder motive, customer result, product capability, or deadline can damage a live deal even when the rest of the draft is strong.
The safest working rule is not merely “fact-check everything.” It is to make the evidence status visible before review. Require any research, account analysis, transcript summary, qualification score, forecast, proposal, or recommendation to label what it knows and how it knows it.
A fact has traceable support: an exact line from a call transcript, a current CRM field with an owner and date, an approved product document, a primary public source, or a verified customer-evidence record. A citation is a locator, not a guarantee. Open it, confirm the claim, check the date, and preserve the source beside the output.
Facts can still conflict. Do not average contradictions into a confident summary. Show both statements, their sources and dates, and identify what must be resolved.
The most damaging research error is rarely a wrong citation. It is a confident answer where the sources were silent. Require the model to write Not found in the sources reviewed instead of completing the sentence, and keep two outcomes apart: not found (the sources are silent, so ask) and contradicted (a source says otherwise, so resolve it).
Three rules keep account research honest:
Search-enabled tools do not remove this risk. OpenAI’s own help page for ChatGPT search warns that “Search results and citations can be incomplete, outdated, or incorrect” (OpenAI Help Center).
Good inference helps a seller prepare. Bad inference impersonates inside knowledge. “The company announced two acquisitions; integration capacity may become a priority” is a transparent inference. “The CIO is urgently buying an integration platform” is an invented fact unless the evidence says so.
Require every inference to include: the supporting facts, an alternative explanation, the confidence level, and the question that would confirm or disprove it. This turns inference into a discovery aid rather than a claim to repeat.
For consequential work, ask for a table with these columns:
Any fact without a source moves to unverified. Any recommendation without a traceable basis moves to unsupported. Any unknown that matters to the deal becomes a next-step question with an owner.
Verify numbers and calculations independently; competitor and product claims against current authoritative material; customer stories against the approved evidence library and usage restrictions; quotations against the original passage; and dates, owners, pricing, commitments, security controls, legal terms, roadmap items, and buyer motivations against the system or person responsible for them.
Separate the output into supported facts, reasonable inferences, assumptions, unknowns, and recommendations. Cite the exact source for every fact. For each inference, name its supporting facts and one alternative explanation. Do not convert an unknown into a plausible answer. Flag every claim that requires product, finance, legal, security, customer-reference, or seller approval before use.
This instruction improves visibility; it does not replace review. For the numerical-claim boundary, see Never use AI-generated numbers without verification. For who makes the final decision, see Human-in-the-Loop.
For account research, add these lines to the instruction:
Date every sourced fact and state the recency window you used. If the sources do not establish something, write “Not found in the sources reviewed” instead of estimating it. Do not estimate private-company financials, budgets, or buying intent. End with a “What not to conclude” list of statements the evidence does not support.
The goal is not to eliminate inference. Selling requires judgment. The goal is to prevent judgment from being mislabeled as evidence.
Three library prompts already build this discipline in: Build a Comprehensive Account Brief with Deep Research for a source-grounded account brief, Prepare Evidence-Based Pre-Meeting Sales Briefing for dated, labeled meeting preparation, and Stress-test an AI-drafted outreach message before sending to catch unverified personalization before a message goes out. To turn labeled evidence into an executive argument, use the source-grounded executive POV guide.