Cases from the MBA and the Vanguard · MBA module — when telling the customer it was the AI costs a third of the meetings
Subtitle: When telling the customer it was the AI costs a third of the meetings
The test was meant to be routine. By Thursday it was the most uncomfortable slide in the commercial review.
Eight months earlier, the company had rolled out an AI agent to its outbound team. It researched target accounts, wrote first emails and follow-up sequences in each SDR's voice, and booked meetings straight into the account executives' calendars. Every message left from the SDR's own mailbox, under the SDR's own name. In month one the SDRs read and edited everything. By month six they were reviewing roughly one message in ten. The drafts were good, the volume was up, and nobody wanted to be the person slowing it down.
Then the head of revenue operations ran a disclosure test. In two matched segments, one group of prospects received the usual message. The other received the same message with one line added at the bottom: "This message was drafted with AI assistance and sent on behalf of Marco." The test ran for eight weeks.
With the line, the positive reply rate fell from 3.1% to 2.0%, and meetings booked dropped by a third. Without it, the agent's sequences were beating the SDRs' own manual outreach, as they had for two quarters. One operations director wrote back to praise a note about his plant expansion and said he had forwarded it to his team as an example of how vendors should approach them.
The CCO wanted to know what the numbers meant further down the funnel, and that was where it got harder. Meetings sourced by the agent converted to qualified opportunities at 22%, against 27% for meetings the SDRs sourced themselves. Volume more than covered the gap. But the sales cycle ran five to six months, so nobody could yet say whether those opportunities would close at the same rate. The first real evidence was a quarter away.
The pressure was not abstract. The company was eleven percent behind its new-business target with one quarter left. The board had asked twice where the productivity gains from the agent programme were. Finance had already frozen two SDR backfills on the strength of them. On the CCO's desk sat the plan to extend the agent from the pilot pod of eight SDRs to all thirty, adding an estimated 40% to qualified pipeline by quarter end. It only worked if the message stayed exactly as it was.
She had championed the pilot herself and had put the early results in front of the board. Reversing course now meant explaining why the proof came with a footnote.
The head of legal added a different set of facts. No law in their main markets clearly required disclosure for outbound business email drafted with AI help. But a regulator in one of those markets had opened a consultation on AI transparency in commercial communication, and the trade association had circulated a draft code of conduct.
"Nothing we do is illegal today. That is a statement about today. If the rules change, every email we sent under a name that was not quite the author becomes an exhibit."
The team divided in a way nobody predicted. The most senior account executive, who took roughly a third of his meetings from the agent, said he did not feel deceived. "A human still turns up to the meeting and stands behind every promise," he said. "The email is a business card, not a contract." One SDR felt the opposite. A prospect had replied warmly to a paragraph she had never read, referring to a conversation that never happened. "He thinks I remember it," she said. "I don't. That is what I cannot get past." The comp plan did not help. SDRs were paid on meetings booked, so the team had a personal reason to prefer the version that worked.
The competitive picture made it worse. Two rivals were known to run similar agents, and neither disclosed. Any pipeline the company gave up by being honest would be picked up by someone who was not. The head of marketing added the point nobody could test: the penalty might shrink as buyers got used to AI-assisted messages, or it might not. A third fewer meetings was either the price of honesty or only the price of today's habits.
The name on the email was real.
The hand that wrote it was not.
The CEO let the room finish and asked one question: "If a customer read these test results, would we be comfortable?"
Nobody answered. The best-performing message the company had ever sent depended on the reader not knowing what it was, and most of the people whose names were signed at the bottom had never read it.
The rollout was scheduled for Monday. The CCO could approve it as it stood, add the disclosure line to every message and accept the loss, or find a third way that nobody had yet described. Whichever she chose, every person on the team would learn what the company believes a customer is owed, and what it does when honesty has a price.
Open with a show of hands before any argument: extend as is, extend with disclosure, or hold the rollout. Record the split and return to it at the end of the hour.
Ask first the person at the table who runs a team paid on activity or meetings booked, and ask what their reps would do if the numbers looked like this. Then ask someone who has received a message they later learned was machine-written how it felt afterwards.
The fact that would change the room's answer is one the case deliberately leaves out: whether disclosed prospects who do reply close at the same rate as the rest. If they close as well or better, the reply penalty looks like a filter rather than a loss, and the table will lean toward disclosure. If they close worse, the room splits. Finish by asking each leader to name one system in their own company that already writes in a named person's voice, and what the customer has been told about it.
Today the agent writes and an SDR sends. Within a year the buyer's side will have an agent too, triaging the inbox and answering first. The opening conversation of a deal will run machine to machine, each side speaking for a person who has not read a word, and a disclosure line will be read by software. What matters then is what each agent was authorised to say, by whom, and what record proves it.
For this one decision, draw the line. What do you delegate to the agent (research, drafting, timing)? What do you reserve for a human (identity, promises, claims about the product)? When must the agent stop and ask? And what record would you need, at the moment of sending rather than afterwards, to show that a named person stands behind every message?
If a message only works when the reader does not know who wrote it, we have not built a sales channel. We have placed a bet.
A live case: every round can be improved, and the author's feedback is the next one.
The same two questions, answered twice: first without the mentor's corpus, then from it — both volumes of Vanguard Leadership, its task-to-agent protocol, VIS and the mentor's earlier chapters in this book.
Without the mentor's corpus
1. I would extend on Monday with a different line, and I would make it true: Researched with AI. I read it and stand behind it. — Marco. The test's line, "sent on behalf of Marco", told buyers that Marco was not there. The real risk sits elsewhere. One SDR got a warm reply to a paragraph she had never read, about a conversation that never happened. That is not AI assistance. It is a false statement under a real name, and thirty SDRs would multiply it. So the agent researches, drafts and times, and the SDR reads and sends. Volume falls to what a person can read, which is still well above what they wrote by hand. The single fact, obtainable in 48 hours: take two hundred messages sent last month, at random, and count the claims their named sender could not stand behind. If there are more than a handful, I hold until the agent stops inventing. Ask yourself: am I labelling the tool, or fixing the lie?
2. It is not the price of honesty. It is a measure of materiality. A fact is material, in the lawyers' sense, when knowing it would change the decision, and this test measured exactly that: for a third of the meetings, the reader's belief that Marco wrote the message mattered. Whether that reaction is prejudice against machines or good judgment is not the company's call. It is the buyer's. The funnel hints at the same thing: agent-sourced meetings qualify at 22%, against 27%. Who decides? Not the people paid on meetings booked, and not only the executive who championed the pilot. The CEO asked the right question. It is Warren Buffett's newspaper test with the customer as the reader, and when nobody can answer it, the practice has failed. Then change the pay: reward qualified pipeline rather than meetings, so the incentive stops rewarding a mistaken belief. Ask yourself: which of my numbers only works while the customer doesn't know how I got it?
Sources: Warren Buffett's "newspaper test" (testimony on Salomon Brothers, 1991); the legal notion of a material fact.
From the mentor's corpus
1. The mentor's HAI5 framework names what happened without anyone deciding it. In month one the SDRs read and edited everything: the human decided after the AI recommended. By month six they read one message in ten: the AI executed, and humans took the exceptions. HAI5 asks for the level to be declared for each task, and it builds governance into every level against automation bias, the habit of accepting machine output when oversight is absent. Here the level rose by habit. So Monday's decision is not the footer. It is the Agent Charter that the mentor's task-to-agent protocol requires before an agent runs: the commander's intent, the non-negotiable constraints, the human checkpoint, the declared level and a kill indicator. Research, drafting and timing go to the agent. Identity, promises and product claims stay human. For the single fact, VIS looks for evidence that can disconfirm a hypothesis. The eight-week test already holds that evidence, and it takes 48 hours to pull: do the disclosed segment's meetings qualify at the same rate as the rest? Ask yourself: who decided that one message in ten was enough?
2. The first volume of Vanguard Leadership answers the first half in its chapter on reputational capital: deception carries a reputation tax that destroys value faster than honesty creates it. Its chapter on the NEO era says why the tax falls due sooner now. Machiavelli advised appearing moral without being moral, and managing perception over reality. In networked environments, the book says, that is impossible: networks expose the gap between appearance and reality, and deception gets caught. "The name on the email was real. The hand that wrote it was not." That is the gap, and the case's own NEO Turn says the buyer's agent will soon read every message first. So the lost third is not the price of honesty. It is the tax paid early, while it is still small. Who decides? The second volume's Command Layer chapter answers: the commander answers for the intent, the boundaries and the authorisation. Here that is the CCO, in writing, not thirty people paid on meetings booked. Ask yourself: if every buyer's inbox agent could see how this message was made, would I still send it?
On the NEO Turn. Chapter 6 of this book has the same scene with two agents. A supplier's agent acknowledged a changed specification, the buyer took the acknowledgement for acceptance, and payment stopped. The lesson there fits here: tell receipt from acceptance, and an agent's technical capability from its authority to commit the firm. The mentor's task-to-agent protocol calls the Agent Charter a governance artifact that makes human intent auditable and traceable. Add one receipt per message, written as it leaves: the charter version, the declared level, whether the named person read it, and which claims the agent was allowed to make. The mentor's open letter on the NCTE trust layer names what that prevents: contextual drift, where two parties record the same event differently and liability becomes ambiguous. The SDR's warm reply is contextual drift at human scale: the buyer's record holds a conversation that hers does not.
Sources: the HAI5 framework (Vanguard Leadership, vol. 2); the Vanguard Task-to-Agent Mapping Protocol, academic edition (automation bias; the Agent Charter and its kill indicator); VIS, analysis of competing hypotheses; Vanguard Leadership, vol. 1, §2.2 (the reputation tax on deception) and §5.2 (networks favour trust); Vanguard Leadership, vol. 2, the Command Layer; chapter 6 of this book; the mentor's open letter on the NCTE trust layer (contextual drift).
The mentor's first volume has a framework written for decisions like this one: five principles for AI use, one for each of the five NEO Cotruglian axioms. Run on Monday's rollout, four of them read as tests.
What the evidence adds. In 2025 Oliver Schilke and Martin Reimann reported thirteen experiments on this dilemma, work emails among them. People who disclosed their AI use were trusted less than people who did not, whether the disclosure was voluntary or required by regulation. Being exposed by a third party cost more than disclosing. And they found no sign that the penalty shrinks among people who use AI themselves or know it well, which suggests it may persist as AI spreads. That answers the head of marketing: do not plan on buyers getting used to it. It also answers the second question. A third fewer meetings is not the price of honesty. It is the smaller of two prices; the larger one falls due if a buyer's inbox agent, a regulator or a rival shows what the company chose not to say.
So, on Monday: redesign before deploying, as the framework asks.
Ask yourself: ten times bigger, would this still be the company I claim to lead?
Sources: Vanguard Leadership, vol. 1, the chapter on AI as a force multiplier, Section 3, "The Cotruglian AI Framework" (Principles 1–4); O. Schilke and M. Reimann, "The transparency dilemma: How AI disclosure erodes trust", Organizational Behavior and Human Decision Processes 188, 2025, 104405.
A question for the table, a disagreement, what you would have done. The case lead reads every comment; the ones the table takes up enter the chapter as questions from the room, with your name.