Part III — Keeping the book
Chapter 15
Let AI listen and draft, never confirm
On July 19, 2026 I ruled for my team on how our book of value should fill:
"the value CAN be created at a scale and speed no human can keep up with. there must be some kind of dynamic listening in place but not 'always on'. humans and agents need an opportunity to explicitly confirm value BUT most of the [time] we will need to monitor and record IMPLICIT value recognition and realization found in conversation transcripts."
Those are my words as the ruling recorded them; the one word in brackets is mine. Value lands in conversations that nobody will ever listen to twice. An AI can listen to all of them. That's the opportunity.
It's also where the danger is. A machine that can hear everything can also write up everything, in fluent, reasonable sentences, and a book full of fluent, reasonable sentences can say a great deal of value landed when nobody it landed for ever said so.
So the line in this chapter is simple. Let the AI listen. Let it find the person's own words and draft the entry. Never let it be the one who says the value landed.
What this is
There are four jobs in getting a moment of value into the book, and an AI is good at three of them:
- Listening. Reading the conversations you already record, for the moment someone says that something landed for them.
- Finding the words. Copying that person's sentence exactly, with who said it and where in the recording.
- Drafting. Writing the entry, open and plainly marked as not yet confirmed.
- Confirming. This one isn't the AI's. The confirmation is the receiver's own words or their own act. For value that lands inside your team, it's whoever your rules name. Never the doer, and never the machine that drafted the entry.
Our vocabulary calls the first two jobs Value Witnessing: "Recording value by pointing to the moment it actually happened — witnessed, timestamped, cited — rather than asserting value without pointing to where it landed. The citation is the thing a nice-sounding claim cannot fake." Until the receiver's word is on it, an entry the AI drafted is still only something we believe landed, and it stays honestly marked that way.
Underneath all of it sits one of the Five Core Beliefs, AI-Human Partnership over Replacement: "AI should multiply human capability, not replace it." The Value Creation Protocol teaches the beliefs. Here it means something narrow. The AI multiplies how much gets heard. The person who received the value still decides whether it happened.
You need recorded conversations and an AI that reads them. You don't need our software.
How to tell where you are
- You record more conversations than anyone has time to listen back to.
- Your AI meeting summaries say the client was happy, and there's no sentence from the client behind it.
- Your AI writes up what your team delivered, in your team's words, and that write-up is what gets filed as the value.
- People say something landed for them, and it gets heard and never written anywhere.
- An entry an AI drafted names the AI, or the work it did, as the thing that received the value.
What to do
1. Let the listening ride the reading you already do. Whenever an AI processes a recorded conversation, for a summary or notes or anything else, have the same pass listen for this moment. No separate watcher running all the time. My own constraint was "not 'always on'", and it's the cheap way as well: the conversation is already open, the people in it are already known, and the place in the recording is already there to cite.
2. Teach it what counts. It counts when the person who received the value says, in their own words, that something landed: a result they got, something they can now do, time or effort they got back (in their words, not yours), a decision that's now made, a worry that's gone, a thing that now works for them.
It doesn't count when it's your own claim that you helped, a promise or a plan, politeness, someone saying they understood or were surprised, or a milestone you announced. Our listening rule says of politeness: "It costs nothing to say and means nothing on the record."
A promise or a plan isn't a moment of value, but when it names what one person is expected to receive, it can be an expectation. The AI writes that on an entry at Expected, the first stage, only from words someone actually said, the person doing the work or the person it's for, cited to where they said it. It never infers one. Our rulebook draws the line in one sentence: "Nobody writes an expectation and attributes it to someone who did not state it." If what was said names no single person, it's an intention, and it belongs in a CARE record (Chapter 2), not on an entry.
3. Copy the words exactly, with who said them and where. Every word, in order, with the speaker and the place in the recording. Our rule puts it bluntly: "A cleaned-up quote is a fabrication, not a tidy quote." If the transcript is garbled and the AI would have to guess what was meant, the moment is held out. The AI says the transcript is unclear and doesn't fill it in.
4. Hold the near-misses out, and say why. A sentence that's close but not clean goes in a list of its own, with its quote, where it was said, and the reason it was held. Then a person can judge it later without listening to the whole conversation again.
5. Draft the entry open. Waiting on the receiver's word, and visibly marked that way. A drafted entry claims only that a moment was heard and who it was about.
6. Keep the confirming hand off the machine. Two rules from our rulebook carry this. "The counterparty's own words, cited to a real timestamped moment, count as their confirmation — whoever found them." And: "An agent's own say-so, with no counterparty language behind it, is still banned." The AI can be the one who finds the words. It is never the one whose word confirms. In our rulebook's words, "The confirming word is the beneficiary's own."
7. Never let an AI be the receiver of its own work. If an AI agent on your team gains something it can use from somebody else's work, that's real value and it can go in the book. The human who owns that agent confirms it. When my team asked me in September 2026 whose word confirms value an agent, a team or a product received, my pick was: "The human who owns it."
8. Check what was heard and never written. Look at what's been staged for the book and not yet entered. A moment the AI heard that never reaches the book is a moment you lost, and nobody will notice unless you look.
How you know it worked
The receiver's sentence was found, not composed. You can go to the place the AI cited, in the recording, and hear the person say it. And it is that sentence, cited to its moment, that confirms.
The best sign is when the receiver's correction reaches the record before the entry does. I have one, and it's mine. On October 16, 2026, an AI agent on my team prepared a reading of the day that found something I could now do: have an AI set up the LinkedIn events instead of doing them myself. On the call I'd said I wasn't saving any time at all. I was still watching the work. The day's reading records that I "explicitly corrected the interpretation: no time saved yet", and the entry says what I said. It records a new ability, and it records that no time was saved.
Where it goes wrong
The AI's say-so becomes the confirmation. The AI's own sentence about what someone found valuable gets filed as if that person had said it. Check every confirmed entry for one thing: whose words are these?
The AI as its own receiver. Of 39 entries that had passed our own audit, 37 named the agent that did the work as the one that received the value. Our rulebook's verdict on that: "A record that names its own author as its beneficiary is a claim, not a record." Chapter 14 covers what we changed.
The smoothed quote. This is the one no software catches. Our own staging tool says so in its code: it "refuses on a MISSING quote and cannot refuse on a smoothed one." A tidied sentence reads better than the real one, and nothing downstream can tell them apart. The only check is a person reading the quote against the recording.
The summary that says more than the conversation did. In October, a machine-generated meeting summary marked a problem as resolved. The conversation itself didn't say that. Summaries are written to sound finished. Go to the words.
Courtesy taken for confirmation. Our listening rule lists "Marvelous." as politeness that doesn't count. Decide where your own line between courtesy and confirmation sits, and write it down, so that two people reading the same note don't draw it in two different places.
Heard, and never written. A moment that's been heard and staged and never entered doesn't fail loudly. It just doesn't arrive, and step 8 is how you notice.
My own hand on someone else's value. When an AI prepares the day and I rule, my yes can confirm value I received myself, or value one of our AI agents received where I'm the human who owns it. For anyone outside the team, the rule is that I never confirm; their own words do. The slip to watch for is a day's pass that records the owner's yes as the confirmation on an entry about someone outside the team.
Read further
- The Value Creation Protocol, Chapter 2, "The Five Core Beliefs", for AI-Human Partnership over Replacement in full, and Chapter 7, "The AI Replacement Trap", on what happens when AI is used to replace people rather than multiply them.
- The Value Era, Chapter 16, "Making 'it happened' cheap to say", on the thank-you, and why the cheapest confirmation can't be the confirmation on its own.
- The AI-Native Shift, Chapter 15, "Counting What Counts", on asking your records for the story rather than the score.
- In this book: Chapter 6, on hearing the moment yourself; Chapter 8, on whose word counts; and Chapter 14, on correcting what gets through.