Data-Density Rule for AI Assistance
Use AI for dense evidence synthesis, not unsupported expert imitation
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 91%
Evaluate an AI task according to the density and relevance of its source data. When evidence for the exact question is sparse—such as reconstructing one expert's position from a handful of posts—the model may combine general category knowledge with the person's identity and confidently invent a plausible view. When the dataset is large and directly relevant, such as millions of sales recordings, extensive CRM records, or a substantial competitive-document library, AI becomes more useful for summarizing information and detecting patterns. Before relying on an output, inventory the available evidence, confirm that it addresses the question, and require traceability to sources. Use weakly grounded generations only as drafts to challenge. Strategic positioning should still be interpreted by a cross-functional group that combines the distinct customer knowledge held by sales, marketing, and product.
Origin
Extracted from Marketing Against The Grain during April Dunford's discussion of ChatGPT hallucinations, CRM analysis, competitive-document summarization, and The Jolt Effect's analysis of 2.5 million sales recordings.
Core principles
- 01AI performance depends on the quantity and quality of relevant evidence.
- 02Sparse source material encourages plausible but unsupported completion.
- 03Large, disparate datasets are well suited to pattern finding and summarization.
- 04AI output should be treated as a starting point when source grounding is weak.
- 05A human cross-functional team remains necessary for strategic positioning judgments.
How to run it
- 1
Specify the Exact Task
Define the question, output, and decision the model is expected to support. Avoid broad requests to reproduce an expert's entire method or worldview.
Pro tip Break a large strategic task into evidence retrieval, synthesis, and judgment.
Watch out A vague prompt makes it difficult to assess whether the model has relevant data.
- 2
Inventory the Evidence
Identify the available call recordings, CRM fields, documents, transcripts, and source publications. Check whether they directly cover the requested subject.
Pro tip Count relevant examples rather than total documents or tokens.
Watch out A large dataset can still be sparse for a narrow question.
- 3
Classify Data Density
Treat the task as sparse when only a few relevant statements or cases exist, and dense when many directly comparable examples are available. Adjust expectations accordingly.
Pro tip Ask whether a human analyst could infer the answer reliably from the same evidence.
Watch out Do not confuse the model's broad general knowledge with evidence about a particular company or expert.
- 4
Choose an Appropriate AI Role
For dense data, use AI to query, summarize, compare, or detect patterns. For sparse data, limit it to brainstorming or drafting and do not treat the result as authoritative.
Pro tip AI is especially useful for reducing large document sets into material a team can review.
Watch out Expert imitation from sparse data often produces convincing fabrication.
- 5
Require Source Grounding
Trace important findings to calls, records, or documents and inspect the supporting context. Discard conclusions that cannot be substantiated.
Pro tip Ask for source identifiers, quotations, or record references with every key claim.
Watch out Fluent language and confident tone are not evidence of accuracy.
- 6
Interpret Cross-Functionally
Bring the grounded findings to sales, marketing, and product so each function can test them against its own customer context. Use the discussion to make positioning or pitch decisions.
Pro tip Pay special attention when two functions see the same pattern in different settings.
Watch out Do not automate away the collaboration that produces strategic insight.
In the wild
A user asks ChatGPT what April Dunford believes about positioning statements. Although Dunford has published several pieces opposing them, the model combines generic positioning advice with her reputation and claims that she considers a positioning statement essential.
→ The fluent response is wrong because the relevant evidence is sparse and the model fills the gap with a plausible average.
Researchers behind The Jolt Effect analyze two and a half million sales recordings with AI and compare patterns associated with successful and unsuccessful calls. The scale reveals counterintuitive findings about talk ratios and fear-based messaging that small anecdotal samples could miss.
→ Dense, directly relevant data supports pattern discovery that can challenge accepted sales wisdom.
A company has extensive competitive documents and customer PDFs. AI summarizes the corpus into a digestible evidence pack, after which a cross-functional group evaluates the findings and decides what they mean for positioning.
→ AI reduces synthesis effort without replacing human strategic judgment.
Common mistakes
Mistaking Fluency for Grounding
A model can produce a coherent answer by averaging general information even when it lacks evidence about the specific expert, company, or market.
Ignoring CRM Data Quality
AI cannot reliably extract win and loss insights when competitive fields, outcomes, and call records are incomplete or inconsistent.
Automating the Cross-Functional Judgment
Summarization can assist positioning work, but it cannot replace the distinct contextual knowledge contributed by product, marketing, and sales.
Is it for you?
Best for
It is best for marketers and sales teams deciding whether to use AI for CRM analysis, call analysis, competitive synthesis, or pitch preparation.
Not ideal for
It is not ideal as justification for delegating positioning decisions to a model that lacks complete customer, product, and market evidence.
From the transcript
“Where people wish it did a better job, and sometimes it can trick you into thinking it's a good job, is where there's actually kind…”
“Like again, like these things don't work with sparse data, but you got a lot of data, two and a half million sales pitches, that's…”
“There's magic that happens when cross-functional teams get together and say, I see this with customers. Oh, that's funny. I see this with customers.”
From the episode
The 3-Step Framework To Win Every Sales Pitch ft. April Dunford
April Dunford