You've implemented AI across the org. Now everything you publish sounds like everyone else's.
Idiograph is a supervised agent — supervised for you, not by you — that finds the true stories hidden in your calls, transcripts, and notes. Stories no one else can claim.
That's why the brand guide didn't work. It changes tone. It doesn't touch any of this.
It's a catalog of true, verified claims to draft from and check against.
Idiograph is the supervised agent that builds it. It pulls the rich, specific detail buried in your calls, transcripts, decks, and notes — the detail your sales and marketing teams need to actually prove your value proposition — traceable back to the exact call, transcript, or document it came from.
Nothing hallucinated makes it into the library by accident. A pulled moment has to align with your value proposition, clear the Idiograph Credibility Standard, and carry a verifiable chain of evidence back to its source — before it's committed to your catalog as a kernel — a single verified claim: one moment, one source, checked before it's stored.
Before it reads a word of your archive, it's set up to know what actually matters to your business and who you're trying to reach.
Calls, transcripts, decks, notes — checked against the Idiograph Credibility Standard before anything counts as a story.
Dates, figures, and claims cross-referenced against the public record before anything is called true.
Each verified story ranked high, medium, or low on how likely it is to actually shift belief — not sorted by where it might be used.
A named person on your side clears it for use. A deliberate, separate decision, every time — not a formality.
Every approved story lands in a queryable catalog your content and sales teams draw from. Provable claims only.
What you get at the end of the process
A catalog of true tales, not tall tales.
You get a delivered catalog of verified claims — ready for sales, marketing, and comms to draw from immediately. Ongoing supervision keeps new stories flowing in, if you want it.
Case study · One real client archive, verifiable claims in minutes
XPRIZE had three decades of innovator stories, mission-defining moments, and institutional lore — scattered across transcripts, individual memories, and folders no one had opened in years. None of it structured. None of it governed. None of it activatable when leadership needed to make the case to a donor, a partner, or a board.
We built a structured story catalog: a queryable record that treats story as a living, governed asset. Every kernel passes through their messaging playbook and a proprietary extraction framework. Every kernel is traceable to a verified moment. Every kernel can be assembled into a fundraising memo, an analyst briefing, or an executive keynote without a writer starting from a blank page.
A founder's notebook on why harvesting your truest stories will boost your brand's credibility — the thesis as it sharpens, the objections worth taking seriously, and the conversations that keep changing how I describe this.
A 1932 memory experiment and a 2023 replication on GPT-3 find the same pattern: unfamiliar, specific detail doesn't survive retelling. It gets replaced by something conventional, and the account still reads as complete — so nothing invites a second look.
Ask a model something it can't fully answer and it doesn't hedge — it gives you a number. Specificity is the texture of evidence, so when the material underneath is thin, the model gets more specific, not more careful.
Ask a generative model a question it can't fully answer and watch what it does. It doesn't hedge. It gives you a number. That's the model doing exactly what it was built to do: produce text that sounds like it knows something.
A 2020 Google study found hallucinated claims in over 70% of AI-written summaries — and the invented ones weren't garbled memories of something real. They were facts appearing from nowhere, and when checked against reality, over 90% of them were wrong. Specificity is the texture of evidence. So when the material underneath is thin, the model doesn't get more careful. It gets more specific, because specific is what the training rewarded. A voice guide changes how the invention sounds. It can't talk a model out of inventing — the pull isn't a tone problem, it's the objective function.
Responding to "The Role of Organizational Memory in Scaling Enterprise AI" (Forbes Tech Council, Mar 2026). The diagnosis is correct: AI doesn't fail because the model is weak, but because it doesn't know how this company works. Then the article prescribes cleaner documentation — and quietly destroys the asset it's trying to protect.
Responding to "The Role of Organizational Memory in Scaling Enterprise AI," Forbes Tech Council, March 2026.
A recent article in Forbes, The Role Of Organizational Memory In Scaling Enterprise AI, asserts that to successfully scale enterprise AI, companies must treat their internal organizational memory (historical data, documentation, and operational context) as a strategic asset.
The article, in short: enterprises keep chasing AI success by buying better models and bigger datasets, but that's not where the problem is. AI struggles in a real company because it doesn't know how that company actually works — the decisions, the history, the lessons that live in documents nobody opens again. The author argues organizations should treat their institutional memory as a strategic asset: document decisions better, curate what matters instead of hoarding everything, review internal content regularly, and identify the key people who hold operational knowledge. Do that, he says, and the AI gets a cleaner signal, employees start to trust its output, and adoption spreads. His close: "AI doesn't replace organizational memory; it relies on it."
Here's my response.
The article gets the diagnosis right and the fix wrong.
The diagnosis is correct. AI doesn't fail in the enterprise because the model is weak. It fails because the model doesn't know how this particular company actually operates. Better models don't fix that. More data doesn't fix that. The missing piece is the organization's own memory — what it learned, what it tried, what happened. The author's line "AI doesn't replace organizational memory; it relies on it" is exactly right. I'd sign it.
Where I split from him is what he thinks organizational memory is, and what he thinks you should do about it.
He treats memory as a documentation problem. Write decisions down better. Curate the good stuff. Review your internal content on a schedule. Clean it up so the AI gets a clear signal. That sounds reasonable, and it's wrong in a way that matters.
The knowledge that's actually valuable is not the cleaned-up version. It's the raw version. It's the debrief someone sent at 11pm before anyone decided what the official story was. It's the transcript of two people who understood the problem actually working it out. It's the field note from the call, written before it got sanded down for the deck. That material is dense with the specifics of what really happened — and the specifics are the point, because the specifics are the one thing AI can't manufacture.
The moment you "curate" that, you strip out the very thing that made it worth keeping. Curation removes the particulars. It turns the debrief into a bullet point. You end up preserving the press-release version of your memory, which is the version that was already sanitized. So the standard advice — document more, curate, review — quietly destroys the asset it's trying to protect.
There's a second thing missing from his piece, and it's the more important one. He gets as far as "ground the AI in real operating context" and stops right before the actual answer: provenance. Context isn't enough. What makes a claim credible is that you can trace it back to a specific person who was there, who said it, and who can be held to it. That chain — this happened, this person saw it, here's the record — is what an AI can't fake. It can imitate the tone of something real. It cannot have been in the room. It cannot be accountable for what it says. That's the line that doesn't move, and it's the line everything should be built on.
So: right that memory is the constraint. Wrong that the fix is cleaner documentation. The fix is capturing the raw, specific, traceable evidence your organization is currently throwing away — and keeping the fingerprints on it, not polishing them off.
When generative AI makes polished corporate competence free and infinitely replicable, what's left? The argument: credibility now requires a structural shift — from manufacturing content to documenting reality. The differentiator is provenance. Claims tethered to a verifiable record that predates the claim itself.
Which is a harder problem. In real interaction, we've got built-in signals for what's credible. AI has learned those signals cold. What it can't produce is the residue of human presence.
In real interaction, we've got built-in signals for what's credible. We've evolved to pick up and communicate credibility through cues in tone, hedging, and social proof references. AI has learned those signals cold.
What it can't produce is the residue of human presence. Historian Carlo Ginzburg called it dark proof: genuine knowledge surfaces through involuntary residue — things that couldn't have been staged because no one knew to stage them at the time. All those little idiosyncrasies are the proof of life that gets smoothed over by AI-generated content.
A sociologist warned a room full of AI researchers that the real disruption isn't superintelligence — it's "Artificial Good-Enough Intelligence," cheap and fast enough to make every signal we use to infer authenticity structurally unreliable. What remains, when the cover letter and the case study can be perfectly faked, is the thing historians have always relied on: provenance. A chain of custody back to a moment that had witnesses.
A story is what you say. A history is a story that survived a fact-check. The distinction sounds academic until you watch a board respond to a claim that has provenance attached versus one that doesn't.
Most organizations have a strong story and almost no history. They have the line — the founding myth, the impact stat, the inflection year — but they cannot produce the artifact that proves the line. The board meeting that decided it. The customer who first articulated the need. The internal memo nobody filed.
I keep watching this play out in pitch rooms. The slide says "we serve over 40,000 students annually." The room nods. Then someone asks which student, in which classroom, last Tuesday. And the deck goes quiet, because the story was never built on a history.
This is the argument I'm still sharpening: belief moves on the verifiable. Not on the well-told. The well-told gets you in the room. The verifiable gets you the wire transfer.
It is not. A CMS stores artifacts. Idiograph stores verified kernels of organizational history with provenance attached — so any agent, writer, or system that touches your name can ground its output in a moment that actually happened.
The conversation, lightly edited:
CMO: So this is basically a content management system with extra steps.
Me: A CMS asks "where did we put the asset?" Idiograph asks "what actually happened, who witnessed it, and where is the residue?" The first is a filing problem. The second is an epistemology problem.
CMO: Give me the version I can repeat to my CEO.
Me: Your CMS stores what you said about yourself. Idiograph stores what is true about yourself, and lets every downstream system — writers, agents, decks, donor portals — draw from the same verified source. One is a folder. The other is a foundation.