What it means for AI-generated sales content to be grounded

An AI agent drafts an outbound email in four seconds, and it names a price, claims a competitor gap, and asserts an integration the product may or may not ship. The copy reads beautifully and the rep sends it. Nobody in the revenue org — not the VP of GTM, not the positioning owner, not Counsel — can say which of those three claims was ever verified, or where any of them came from.

That gap has a name now that AI writes the content. The reflex is to edit the email, but the email is the output of a process, and editing one output fixes nothing about the next four-second draft.

The Commercial Truth manifesto argues that marketing is the last commercial function without infrastructure — engineering inherited version control, finance inherited the ledger, and the claims a company makes about itself never got a substrate (Source: Assay brand canon, §10 Truth Graph manifesto). “Grounded” is the property that substrate gives back to AI-generated content.

The concept, named plainly

Grounded sales content is content in which every factual claim traces to a recorded source, and the confidence each claim carries is bounded by how that source was obtained. The opposite is ungrounded AI: copy where nothing is invented, except that nothing is sourced either, so a verified price and a confident guess look identical on the page (Source: Assay brand canon, §PIL-B1 Grounded).

The word doing the work is grounded, not correct. A grounded claim can still be wrong, but it can always tell you who said it, when, and on what authority — and that is the precondition for catching the error before it ships. Assay’s brand canon states the standard directly: nothing is invented, nothing hallucinated, nothing assumed; we never say something we cannot source (Source: Assay brand canon, §PIL-B1 Grounded).

This is why grounding is a property of the claim, not a tone of voice. You cannot make content grounded by instructing a model to sound careful. You make it grounded by attaching a provenance to every assertion underneath it.

Why this matters now

Grounding was a tidiness preference when humans wrote every sentence by hand. It became load-bearing because the substrate underneath GTM changed.

AI agents now generate sales content at machine speed — outbound, battlecards, proposal sections, follow-ups — each grounding from whatever document it was last pointed at. When that source is unversioned prose, every agent inherits the same ambiguity and amplifies it into live conversations faster than any human can chase (the truth gap). Five tools end up asserting five versions of the same company (5 AI tools, 5 versions of your company).

A document cannot tell an agent whether a claim was human-verified or machine-guessed, so the agent treats both as equally true (the confident wrong answer). The whole point of grounding is to make that distinction travel with the fact, so the content an agent emits carries its evidence rather than burying it.

The primitives grounding introduces

Grounding is not a vibe; it rests on a small set of structural primitives the document model has never carried.

Source type. Every claim records how it arrived, drawn from a fixed vocabulary of six categories (Source: Assay source-type vocabulary). The set runs from FIRST_PARTY_VERIFIED — a human entered it and a known reviewer checked it — through SECOND_PARTY (a customer or partner provided it), THIRD_PARTY (public web or news), AI_EXTRACTED (a model pulled it from a known source), down to AI_IMPORTED, a model inference with no direct source (Source: Assay source-type vocabulary).

A confidence ceiling per source type. Each source type carries an explicit cap — the maximum score a claim from that origin is ever allowed to reach (Source: Assay source-type vocabulary). The values are concrete, not rhetorical:

Source: Assay source-type vocabulary.

The rule that makes it load-bearing. Confidence is computed, then bounded: final_confidence = min(source_ceiling, computed_score) (Source: Assay confidence-scoring canon). A model reporting 0.95 certainty on an inference it could not source still caps at 0.50 — explicit human verification is required to raise it (Source: Assay source-type vocabulary). The number that ships is governed by how the claim was obtained, not by how certain the last system to touch it sounded.

A confidence score that is not a binary flag. The score is a numeric value from 0.00 to 1.00 derived from source quality, recency, corroboration, and verification status — and it is a trust score over the evidence chain, not a probability that the claim is semantically right (Source: Assay confidence-scoring canon). It decays over time without re-verification, rises when a corroborating source is added, and falls when a contradicting one appears (Source: Assay confidence-scoring canon).

A worked example

Take the outbound email from the top — the price, the competitor gap, the integration claim. Run each through the primitives and the difference is immediate.

The price was approved by the deal desk and entered by a known reviewer, so it is FIRST_PARTY_VERIFIED with a ceiling of 1.00 (Source: Assay source-type vocabulary). The competitor gap was scraped from a public review site, so it is THIRD_PARTY and caps at 0.60 regardless of how confident the model is (Source: Assay source-type vocabulary). The integration claim is an AI inference from three old decks with no direct source — AI_IMPORTED, capped at 0.50, the example the canon gives verbatim: “We probably support Slack because we mention it three times in the docs” (Source: Assay source-type vocabulary).

The grounded version of that email is not the one that sounds the most polished. It is the one where the 1.00 price ships as stated, the 0.60 competitor line is hedged or held, and the 0.50 integration claim is blocked until a human verifies it — because an AI_IMPORTED reconciliation requires explicit verification to cross into customer-facing copy (Source: Assay source-type vocabulary). On the page the three claims looked identical; underneath, only one earned the right to be asserted flat.

What grounding changes is categorical. The exact share of AI-generated claims a typical revenue org could assign a source type to today is not yet measured publicly; what the primitives guarantee is that the share is knowable — a query, not a guess (your AI made 50,000 claims last month).

Closes / opens

Closes the §E pillar-substrate cluster on grounding: AI-generated sales content is grounded when every claim carries a source type and a confidence ceiling, and the score that ships is min(source_ceiling, computed_score) — so a confident guess can never impersonate a verified fact (source-type taxonomy for marketing claims).

Opens the operator question that follows: of the claims your AI agents made last week, what share could you assign a source type to right now, and what share are confident assertions no one has ever checked? That is the confidence-ceiling question, and it is where grounding turns from a definition into a daily control.

The methodology Assay is developing for the Commercial Truth Index scores vendors partly on exactly this — whether the content they publish carries a source-type distribution at all, and whether its confidence numbers respect a ceiling. A confidence score that ignores its own provenance is decoration; one capped by how the claim was obtained is evidence.

This essay is grounded in Assay’s brand canon (§PIL-B1), source-type vocabulary, and confidence-scoring canon. Methodology for the Commercial Truth Index is in development.