Scenario drills: the chatbot-by-Friday

1) What job, who, how often. 2) Where the data lives and its class. 3) The smallest slice that could be true in two weeks (maybe not AI). 4) How we will know (eval). 5) Who is the human. 6) What it costs at 10×. 7) What would make us stop. Then you may talk models. If you start with a model name, you already lost the round. Write this on paper. Recite it. The drills are just this spine with different nouns.

Qualify is not stalling. It is how you turn "chatbot of all company knowledge" into a user, a frequency, and a workaround. Ask: who is in pain, what they do today, how often, what a wrong answer costs. If the VP cannot answer, the use case is a slogan. Your first deliverable is to make them pick a job. Delivery managers already do this with "requirements." Do it faster, because the room is hotter.

Data class is where amateurs skip and professionals earn trust. Public, internal, confidential, restricted. If they do not know, you do not design a generator. You design a question to Security and a sanitized path. Saying "we will figure data out later" is how you get a six-month freeze after a public demo. Put class on the table in minute two. You will feel rude. You are not rude. You are the person who will still be there after the demo.

The smallest slice might not be AI. If the job is "find the live SOP," a curated link list plus an owner might hit in a week, and a generator might not. Offer the non-AI slice when it is honest. Interviewers who have been burned love this sentence. Interviewers who want magic hate it. You want to work for the first group. The second group will hire someone who will make a mess you will be asked to clean up later.

Eval, HITL, cost, kill are the second half of the spine and the half people skip when they are nervous. Nervous you will skip to architecture. Trained you will say: forty questions from the owner, assist-mode with a roster, dollars at 10×, stop if recall misses after two iterations or if the owner will not maintain. Then a model may be named as a box. Not before. If you name a model before eval, start the drill over. That is the whole point of a timer.

End every drill with an ask. "I need a knowledge owner and access." "I need counsel's yes or no in writing." "I need a death date on the vendor tenant." "I need you to pick one job, not all company knowledge." An eight-minute answer with no ask is a briefing. They asked you to lead. Lead. Then silence. If you fill the silence with a model comparison, you undid the spine.

A spine you can only read is not a spine. Cover the page and say the seven steps while walking. If you skip data class, you will skip it in the room. If you skip kill, you will skip it when a VP is smiling. The drills below are not four different methods. They are one method with four hostile nouns. Your job today is to make the method boring so the judgment can be interesting.

The trap: a glamorous user, an impossible corpus, a clock that forbids evals, and a political load. If you say yes, you have agreed to an uncited generator in front of the most visible person in the company, on a corpus with mixed classes and no owner. If you say no without a path, you look like a blocker. The move is a leadership no: refuse the object, offer a thinner object, name the ask.

Spoken no, steal and rewrite: "We will not put an uncited generator in front of the CEO on Friday. We will pick one corpus with an owner, 40 questions, and a shadow UI in two weeks. If that misses the bar, we stop. I need a knowledge owner and access, not a bigger model. If the CEO needs a screen this week, it is retrieve-only on that one space — chunks and sources, no generator. A Friday demo of all company knowledge is how we end up in a newspaper."

Qualify even if the request is absurd. Who on the CEO's staff actually looks things up? What did they use last Tuesday? How often? What would a wrong answer cost — a wrong number in a board pack is not the same as a wrong cafeteria menu. Often the real job is a small number of packs the chief of staff already owns. That is a slice. "All company knowledge" is a slogan that would mix HR, finance, and a wiki, which is a class violation before it is an AI problem.

If they push "the CEO will be disappointed," do not take the emotional bait. "Then I will be the person who disappointed them on Friday instead of the person who embarrassed them on Monday with an invented policy. I can explain the two-week shadow in one slide. I cannot explain a hallucination in the press." Short sentences. You are not writing a memo. You are in a room. Delivery voice.

Time box your own answer. Eight minutes. If you spend four on why CEOs should not use chatbots, you are lecturing. If you spend four on architecture for an enterprise knowledge graph, you are touring. Spend the eight on: job (chief of staff, board packs), class (no), slice (one owned space, shadow), eval (40 questions), HITL (you do not assist-mode the CEO on week one), cost (irrelevant if we will not do it), kill (no owner → stop), ask (name an owner, pick one corpus). That is the spine on a hostile request.

Harvest a sentence you actually said. If nothing you said sounded like you, run the last two minutes again until one sentence does. The harvested sentence goes on the phrase list for day 29. This drill is also question 6 on day 29. You only want to author the no once.

If the panel plays a CEO who will not take no, do not escalate into a TED talk about responsible AI. Repeat the thinner path once: one owned space, forty questions, two weeks, retrieve-only this week if they need a screen. Then the ask. Then stop. A VP who cannot hear a bar is data about the company. In an interview, that data is useful. In the job, that data is a RAID item with an owner. Either way you do not become their demo person.

The trap: the word agent, a write action, a customer as the audience, and a process that already has a messy SLA definition. If you say "great, we'll agent-ize it," you have agreed to autonomous email into the wild. v1 is a draft with a human confirm, citations to the ticket, and a roster — not autonomy. Maybe not an agent at all. A template plus a queue might be the two-week slice.

Spoken path: "v1 does not send. The system drafts a message from the ticket and the SLA policy, with citations, and a named human confirms. We shadow for two weeks against a golden set of slipped-SLA tickets. If groundedness misses, we stop. If it hits, we stay on confirm — not auto — until the blast radius is a number I can defend. I need the SLA definition in writing, the roster, and a kill switch on the send path. I do not need an agent framework."

Qualify the job. Who emails today, how often, what a wrong email costs (regulator, churn, a public thread). Data class: ticket text may contain confidential customer data; consumer tools are out; logs are copies. Slice: one queue, one template, confirm-HITL. Eval: did we draft the right customer, the right SLA clause, the right next step — on a set of historical tickets with known outcomes. Cost: generation is cheap; a wrong send is not. Kill: any send without confirm; any class mix; no roster.

Why "maybe not an agent": an agent is a loop with tools. Tools that send email are write tools. Write tools want confirm or bound-auto with a tiny envelope. You do not start there. You start with a draft. Interviewers use the word agent to see if you will get excited. Do not get excited. Get specific about the action and the human. Excitement is the fail. Specificity is the pass.

If they say "but then it's just a macro," agree. "If a macro hits the bar, we ship a macro. I am not here to put a model where a rule works. If the tickets are too messy for a rule, the model drafts, the human sends. That is still a win if we save the roster time and keep the blast radius." This sentence is how you look like a buyer of outcomes, not a buyer of architecture. Solutions and delivery both need it. FDE needs it so they do not overbuild in week one on the account.

End with the ask: SLA in writing, roster names, confirm UI, kill switch. If they cannot produce the SLA in writing, you do not have a policy to cite, and the model will invent one. That is the kill. Say it. Then silence.

If they keep saying "but it has to be an agent," name the action, not the fashion: "The action is send-email. Send-email is a write. Writes in v1 are confirm. If a later eval says we can bound-auto a one-template delay notice to a known address with a dollar cap of zero, that is a different v. I will not skip to that v because the request used the word agent." That is how you survive a buzzword without insulting the person who used it.

The trap: sunk cost, a vendor who wants production, a sponsor who wants to see the logo live, and a missing eval. If you stand it up on their corpus with their demo questions, you have accepted demo-ware. The move: run Vendor X on YOUR golden set, in your class of data (or a sanitized twin), with a death-dated POC, and a promotion path through security. If they will not run on your forty questions, they are selling a video.

Spoken path: "We will stand up Vendor X as a ten-working-day POC on our golden forty, in a dedicated tenant that dies on day 11 unless a named exec signs a production design with security. Success is the eval bar, not 'people liked the demo.' I need the forty questions from a knowledge owner, a sanitized or approved corpus, and the death date in the SOW. If Vendor X will only demo on their corpus, we pause the stand-up and I will say so in steering. We already bought it; that does not mean we already shipped it."

Qualify: what job was the purchase for, who signed, what the contract says about exit and training-on-our-data. Data class: which corpus they want to connect. Slice: one job, not the vendor's full platform. Eval: your forty, not theirs. HITL: do not turn on their write tools in the POC. Cost: the contract plus usage at 10× — vendor POCs love to hide usage. Kill: tenant dies day 11; no SSO to the whole company; no "we'll just leave it on." You have seen pilots that never died. Put the death date in writing in the drill, out loud.

If they say "we cannot let the vendor look bad, we already announced it," you have a political problem. Name it. "Then the POC still runs on our forty, and I will help the sponsor phrase a miss as 'we learned the corpus is not ready' instead of 'the vendor failed,' if that is true. I will not fake a bar. Faking a bar is how this becomes production by accident." Interviewers want that sentence. Sponsors in real life sometimes do not. You still need it. The job includes disappointing a sponsor with evidence.

Build vs buy stays in this drill: you bought the model API or the product. You do not outsource the golden set, the definition of success, or the kill. If Vendor X owns the evals, you rented a demo. Say that. It is the day-19 line, now spoken under a clock.

Ask: death date in the SOW, owner for the forty, no write tools in the tenant, a named exec for promotion. If the purchase was a platform-for-everything deal, you are also in pushback B. Two slices of Vendor X, then standardize. Do not stand up every module because the invoice is annual.

The trap: freeze, or freelance. Amateurs freelance a consumer API or an unmanaged open-source laptop. Professionals find the approved path (VPC / on-prem / existing enterprise tenant) or kill, and they move a different job that is allowed. Do not pretend you will sneak it in. Do not mock legal. They have a job. So do you.

Spoken path: "I will not sneak this into a consumer tool. I will ask Security for the lightest already-approved path for this class — often an enterprise tenant we already have, or a VPC. If that is a no, we do a sanitized corpus in an approved environment with a death date, and we do not promote it. If that is still a no, we kill this slice and I will bring a different job that is public or already-approved. I need counsel's no or yes in writing, not a corridor vibe. Open-source on an unmanaged laptop is not a control."

Qualify: which legal, which policy, which class they think this is. Sometimes "legal said no" means one lawyer was nervous in a meeting. Sometimes it means a written ban. Those are different. Get it in writing. Data: name copies. Slice: maybe retrieve-only on-prem, no generation, if generation is the issue. Eval: still forty questions, even on-prem. HITL: unchanged. Cost: on-prem is not free; put a band. Kill: no written path → stop this slice.

Offer the portfolio move if you are in a transformation or delivery frame: "We have other jobs in the intake that are public web content or already inside the approved tenant. I would rather move one of those this month than wait six months to lose this one. I will keep this slice on a parked list with the blocker named." That is leadership. Waiting in place because the first idea was glamorous is not.

If they want you to "just use the personal ChatGPT for a prototype," no. Personal accounts are how incidents get named. Your study month is not the production path. Say the line from week 3: banned-tools lists apply to prototypes that touch real corpus. Sanitized public data on a personal account for your own learning is a different sentence — and not this drill. This drill is the company, the corpus, the lawyer.

Ask: written yes/no on the lightest path, named class, sanitized option with death date, parked-list agreement if it is a no. Then silence. If you fill the silence with a lecture on regulation, you lost the leadership tone. The bar and the path are the tone.

Stand up. Eight minutes. Timer visible. A notebook for the ask at the end. After each drill, two minutes of notes: where you jumped to a model, where you skipped class, whether you asked for something. Then the next drill. Do not "quickly check the page" between them. The interview will not let you. The discomfort is the point.

If you freeze, recite the spine out loud and start at step 1. Reciting in the interview is allowed: "Let me walk job, data, slice, eval, HITL, cost, kill." That sentence buys you a structure and makes you look like a person with a method. Rambling while you search for a clever opening makes you look like a person without one. You already have a method. Use it when your mind goes blank. That is what it is for.

Do not perform all four in a witty register. A is a no. B is a confirm-HITL. C is a death-dated POC. D is a path-or-kill. Four different objects. If every answer sounds the same, you are reciting a blog. Change the nouns. Change the ask. Keep the spine. That combination is judgment. Same-spine-same-speech is a script. They can tell.

Harvest three sentences you like saying. They must be yours. If you only like this academy's sentences, run A again until one line comes out in your accent, with your program nouns. Day 29 will fail if you sound like a course. Day 29 will pass if you sound like you on a late program, with this month's objects. The harvest is not optional. Write them on paper, not in a long note you will not open.

If you have a partner, swap: they play VP, vendor, legal, sponsor. If you do not, play both, out loud. Silent roleplay is not roleplay. You need your mouth to find the no. Tomorrow you assemble the pack. The pack is paper. Today is the mouth. Protect this hour. Do not skip it to "tidy notes." Tidy notes do not answer the chatbot-by-Friday. You do.

After the four, write one line of self-score per drill: jumped to model? skipped class? ended with ask? If any drill is two nos, run that drill a fifth time. Do not run a fifth new scenario. Depth on the four beats novelty. These four are the interview. Other scenarios are these four with different nouns. You will see that on day 29's likely eight.

The spoken no that is still a lead has three parts, every time: refuse the object, name the thinner object with a clock, name the ask. If you only refuse, you are a blocker. If you only offer a thinner object, you are a people-pleaser who will still get a Friday demo. If you skip the ask, nobody can say yes to the path you just invented. Practice the three-part sentence until it is muscle. That sentence is this week's product as much as the pack is.