Where we start
Most AI projects that fail, fail in the first meeting, and they fail quietly. The business person names a tool because "AI" is the word she has for "make this problem go away", the developer hears a build, and both walk out believing they agreed. The running case for all six sessions is Harbourline, an invented chain of 38 pharmacy and clinic outlets across Singapore and Malaysia, and its operations director Adeline Tan. She has never written a prompt and does not need to. She has a request, three area managers who live in WhatsApp, and a board that keeps asking what she is doing with AI. You have 45 minutes with her.
★ What you walk out with tonight. A one-page door sheet for "can AI answer the WhatsApps", two open questions behind each door written in the past or present tense, a plain teach-back line attached to each one, and a mark on the current-way row, the door an AI developer opens first and widest. Session 3 puts that sheet in front of Adeline and counts what it gets out of her.
Two tunes, and the word AI
5 min liveIn 1990 a Stanford psychology student, Elizabeth Newton, split people into tappers and listeners. Tappers picked a well-known song and tapped its rhythm on a table; listeners had to name it. As reported by Chip and Dan Heath, tappers guessed the listeners would get about half the songs. Listeners got 3 of 120. The tapper cannot stop hearing the tune, so the taps sound obvious; the listener hears knocking. Every discovery meeting has a tapper on each side of the table, and the word AI is the loudest knocking there is.
LiveHer tune, his tune, and why neither side can hear the other's3 min▶
Newton's tappers were not bad at tapping. They were unable to un-know the song while they tapped, so every knock arrived with the melody attached, and they judged the listener's chances from inside that experience. Economists had named this a year earlier: Camerer, Loewenstein and Weber called it the curse of knowledge, the finding that better-informed people cannot fully set their information aside when predicting what less-informed people will understand.
In an AI discovery meeting there are two tunes. Adeline's is the operational worry. When she says "can AI answer the WhatsApps", she means "get my managers back into the outlets": head office expects each area manager to visit two outlets a day, they manage one because of the phone, and that is what she is measured on. The developer's tune is the machinery: a model, a prompt, an integration with a messaging platform, a way of reading the thread history. Each hears their own tune behind every word the other says. Both leave sure they agreed.
The gap is not closed by her learning what a prompt is, or by you learning pharmacy. It is closed by questions that make the tune audible: about the workflow as it happens today, which steps in it a person must keep, and what she would see if it worked, asked in words that need no translation, with the reason for asking said out loud. That is the rest of tonight.
Self-studyWhy the word AI widens the gap instead of naming it2 min▶
Most requests arrive as a solution: a dashboard, a report, a model. "AI" is a solution with an unusual property: it sounds like it could be anything, so each side fills it with whatever they already had in mind. Adeline has two reasons for saying it. She wants her managers in outlets, and the board wants the word AI on a slide. Those are different outcomes, and only one of them is a test you could pass.
The developer's version of the trap is to answer the word with a build plan. "Should we fine-tune, or go retrieval-augmented over the thread history?" is a good question between two developers and knocking to a person who runs pharmacies. On the bench in session 3, opening with five technical questions like it is the lowest-scoring way of running the meeting: 11 points, against 54 for the same 45 minutes walked door by door. The words are not wrong. They are the wrong tune for the room.
The bot she already switched off. A constructed case, but the shape is one every AI developer meets. Last year Harbourline's POS vendor switched on a chatbot, it answered a customer's prescription question wrongly, and Adeline switched it off within a week. She will not mention this unless you ask whether she has tried anything that answers on its own, and it is the single most important fact about what she will trust. An assistant that replies to suppliers by itself is the same thing in her mind, and she is right to think so: one yes to a substitute drug nobody checked is not a customer-service slip, it is a wrong product on a pharmacy shelf.
The current-way door, opened widest
6 min liveEvery discovery meeting walks the same six doors: outcome, decision, current way, data state, constraints, success and partnership. They are stated in full, with the question behind each and where each move comes from, in the corridor, session 1 of the data analyst sibling; read it once and come back. For an AI developer one door matters more than the rest. Whatever you build will replace or assist a step a person does today, so the facts that decide whether it is safe sit behind door 3: the workflow as it happens, how often it goes wrong, which replies are lookups and which are decisions, and who checks the output.
LiveWatch one thread before you design anything2 min▶
The first question an AI developer asks is not about data or models. It is "show me one thread from yesterday, start to finish. What did you do at each message?" Adeline's answer, drawn above, is the spec nobody wrote. A supplier sends a confirmation with items and prices. The manager checks it against what the outlet asked for, which lives in a different WhatsApp group. The manager replies OK or queries. About forty threads a day per manager, and about four in five end in a plain OK.
Watching one thread tells you three things no dataset would. The work is a match between two messages, not a reply to one, so the real project may be the second message. The work happens on a phone, so the design lives there. And most of it is routine, which is the part an assistant could safely draft, while the rest is where the person has to stay.
LiveLookups and decisions: the split every assistant rests on2 min▶
Split every step of the thread into one of two kinds. A lookup is a step where the right answer already exists somewhere: does this price match the price list, is this the item the outlet asked for, when is the delivery. A decision is a step where a person chooses and owns the choice. At Harbourline one kind of reply is a decision: accepting a substitute product or a price change. Only the area manager may say yes to it. Delivery times just need to reach the outlet; nobody decides anything about them.
The line matters because the cost of a wrong answer is not the same on each side. Pharmacy products need a substitution record by law, and a wrong drug on a shelf is not a customer-service issue. So you ask how often it goes wrong today, and her answer, once she trusts you a little, is that about one confirmation in twenty has a wrong price or a substitute in it. The managers catch most of them. That catch rate is the bar: any assistant has to do at least as well, and the catching is the part a person keeps.
You do not draw the line yourself. You ask her where it is, because it is her line to draw and her name on the record.
Self-studyCheck the non-ML answer before the ML one2 min▶
Google's introduction to machine learning problem framing makes a point that most AI projects skip: before building a model, check whether a simpler, non-ML solution would solve the problem, and define what success looks like before anything is built. It reads as caution. In a discovery meeting it is a question you owe the person across the table.
Harbourline has at least two non-ML answers sitting in the current-way door. Nobody keeps a record of what was decided; the thread is the record, and when an area manager left in March her phone went with her. A shared decision log fixes that with no model at all. And the supplier contracts renew in three months, so a standard confirmation format could be written into them. You only find either by watching the thread first. Session 5 turns both into candidate projects, beside the assistant that only drafts.
Why current way first, and what it is worth. Session 3 counts it. On the bench, the Six Doors walked in printed order score 54; the same doors tuned for an AI developer, current way first and widest, score 55. One point, and that is the honest number: her request is already about the current way, so the printed order asks the lookups-versus-decisions question by minute 12 and hears the one-in-twenty error rate by minute 16 anyway. The order is a habit worth having. What moves the score by forty points is the kind of question, which is Part 3.
The record that walked out of the door. Still the constructed case. In March one of Harbourline's area managers left, and her phone went with her. Every substitute she had accepted, every price she had queried, every OK she had sent, lived in WhatsApp threads on a personal number. Nobody had asked where a decision lives once a thread is done, because the thread felt like the record. An assistant built without that question would have added a faster way of making decisions to a system with no way of keeping them. The first useful output here is a log, before it is a single answer.
Why I am asking, in one sentence
4 min liveEvery open question in the bank comes with a second sentence: a plain statement of why you are asking it. "The worst case sets how much checking the assistant needs." Said out loud, it does three things at once. It makes the question feel like a reason rather than an interrogation, it earns trust, and it teaches the business person one true thing about how AI work is decided. By the end of the meeting Adeline knows why an assistant should draft and a person should decide, and nobody showed her a slide.
LiveThree teach-back lines an AI developer says in almost every meeting2 min▶
A teach-back line has three properties. It is plain: no word she would have to ask about, so "checking" not "evaluation", "draft" not "generate". It is true: a real principle of the craft, not a pleasantry. And it is one sentence, or two short ones, because a paragraph is a lecture and she stops answering.
Three from the bank, word for word, that earn their place in almost any AI meeting:
- Before "show me one thread from yesterday, start to finish": "The workflow as it happens today is the spec. I cannot automate a step I have not watched a person do."
- Before "which replies need a person to decide, and which are just looking something up?": "I split every workflow into lookups and decisions. Assistants do lookups, people keep decisions, and the line between them is yours to draw."
- Before "what is the worst thing that happens if a reply is wrong?": "The worst case sets how much checking the assistant needs. A wrong delivery time and a wrong drug are not the same mistake."
Read them in order and they add up to one idea she can repeat to her board: the assistant drafts, a person decides. Session 4 builds twelve of these and picks the one concept per meeting worth ninety seconds.
Self-studyFive kinds of question, and only one of them earns trust2 min▶
Session 3 sorts every question in the bank into five kinds, and the simulator scores them differently. Open questions about the past or present uncover facts and earn a point of trust. Closed yes-or-no questions are cheap and get you one fact at most. Vague questions ("what would you like AI to do for you?") cost minutes and get a generic back. Jargon costs twice: she has to ask what you mean, and she is a little less sure you understand her shop. Leading questions are the dangerous kind: "so the assistant should reply to the suppliers automatically?" gets a polite yes, and a belief that is wrong goes into your notes and later into your build.
The trust point is not decoration. Eight of the 24 facts on Adeline's sheet only come out once some trust exists, the one-in-twenty error rate among them, and the most sensitive one needs a great deal of it: the only IT person, who holds the login to the product list the assistant would need, is leaving in January, and she has not told the team.
The same question, with and without its second sentence. Constructed, and familiar. Asked bare, "how often is a confirmation wrong?" sounds to an operations director like an audit of her managers, and the honest answer is the one she is least likely to give a stranger. Asked with "the error rate today is the bar any assistant has to beat, and how you catch errors is the part the assistant has to keep", it sounds like what it is: the developer protecting her managers' judgement, not questioning it.
Your door sheet for "can AI answer the WhatsApps"
★ 22 min · everyone writesOne page, six rows. Use Adeline's request, or a real "can AI do X" from your own inbox, and build the sheet you will carry into the meeting. Every line is written before the meeting; the meeting is for listening. The row that gets the most ink is current way.
Write the request as it was received, in her words. Not your translation. If it arrived as "can AI answer the WhatsApps", write that, not "LLM triage for supplier messaging". You will want this sentence back at the end of the meeting to see how far the real request was from it.
Two open questions per door, in the past or present tense. "Show me", "walk me through", "tell me about the last time". Nothing with "would you". Nothing that can be answered with yes. Write them so a person who has never written a prompt can answer every one. Spend the most care on door 3.
Attach a teach-back line to each question. Plain, true, one sentence or two short ones. Read each pair aloud once; if the second sentence has a word she would have to ask about, such as model, prompt or eval, change the word.
Mark the door your role opens first. For an AI developer it is door 3, current way: your output replaces or assists a step a person does today, so the workflow, its error rate, the lookup-decision split and who checks the output are the facts you cannot build safely without. Session 3 measures what that choice is worth, and is honest that on this sheet it is worth one point over the printed order.
Rehearse the opener against Adeline. Her first line is on the page. Say your first question out loud, with its teach-back. If it took more than twenty seconds, it is two questions; split it.
Keep the sheet after the meeting. Session 2 walks the thread you asked to see end to end and maps where its data lives, session 5 turns her answers into three candidate projects, and session 6 turns the filled-in sheet into the one-page brief you send her within 24 hours. The same page carries all of it.
Before session 2
30 min- Finish the sheet for "can AI answer the WhatsApps" or your own request: twelve questions, twelve teach-back lines, one mark on current way. Bring it; session 3 puts it on the bench.
- Find the tune. For your request, write the one sentence you think the business person would say if you asked "if this went away, what would be different?" Then write the sentence you fear they would say. Both are guesses; the meeting replaces them.
- Tag a workflow you know. Take one process an assistant was proposed for, at work or anywhere, and write each step on its own line. Tag every step lookup or decision, and next to each decision write the name of the person who owns it today.
- Find the last attempt. Ask one colleague whether anything that "answers on its own" was ever tried where you work, and what happened. Write down what it taught them to distrust. Session 2 starts from the thread; that answer tells you how it will be read.
Three questions
Clear all three and session 1 gets stamped in your passport.
1 · Adeline asks "can AI answer the supplier WhatsApps?" Why does an AI developer treat that as a question about a workflow rather than a model?
Two tunes: she hears two outlet visits a day behind the word; the developer hears a model and a prompt. Neither notices. "Show me one thread from yesterday" is the question that makes her tune audible, and it is why door 3 opens widest.
2 · In a supplier thread, what is the difference between a lookup and a decision, and why does it matter?
The line is hers to draw, which is why you ask "which replies need a person to decide?" instead of assuming. The leading version, "so the assistant should reply automatically?", records exactly the belief that produces a substitute drug nobody checked.
3 · About one confirmation in twenty has a wrong price or a substitute, and the managers catch most of them. Why ask for that number before proposing anything?
"Mistakes are rare, so accuracy is not critical?" is one of the five leading questions on the bench, and it records a false belief. One in twenty is rare and not cheap: a missed one is a wrong invoice at best and a wrong drug on a shelf at worst.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Session 1 cheat sheet · pin this
You asked to see one thread. Session 2 watches it end to end: the supplier's confirmation, the outlet's original order typed into another group in English, Malay and Chinese with product nicknames, the manager's match, the OK. Then it asks where each half lives, what is recorded and what is not, and draws the data-state map for an assistant: what it could read, what it would need a dictionary for, and what it must never see.