Portfolio: the delivery artifacts

Rows: job, users, non-goals, data class, eval bar, HITL, SLO, monthly $ / 10×, top risks, sequence, ask. If you cannot fit it, you do not own it yet. Cut until you can. Steering committees do not read your charter. They read a table, or they read nothing and decide on the demo they saw in the corridor. Your job is to put a table in front of them that can carry a decision.

Every row needs four cells: Owner (a named human), Bar (a number), Evidence (where the number comes from), Date. If a row has no owner it is a wish. If a bar has no number it is a vibe. If evidence is "the vendor said," it is not evidence. If the date is "Q3," it is not a date. You already know this from every RAID log that ever failed a program. Do not go soft because the object is a model. Models do not exempt you from owners.

Job row: the job-to-be-done in one line, with frequency. "Atlas PM cites the live SOP-14, ~3×/week, instead of pinging a senior." Users: 40 named seats, not "the business." Non-goals: the same refusals as yesterday, written so a CFO can see them. Data class: internal, not restricted; HR out. Eval bar: recall@5 ≥ 0.7 and groundedness ≥ 0.9 on the owner's 40 before assist-mode. HITL: assist for v1, named roster, hours in the plan. These are not new ideas. They are the charter, compressed until a skeptic can attack them.

Cost row is where delivery leads look adult and enthusiasts look like children. Monthly dollars at current volume and at 10×, plus the daily cap. If you do not have the number from week 3, compute a band tonight: tokens in (question + top-k chunks) × tokens out × price × questions per month. Show the band. A band with an owner is better than a blank. "We will optimize later" is how invoices become incidents. Put the 10× number next to the eval bar so the room cannot fund quality without seeing spend.

Top risks: five, not fifteen. Stale corpus, ACL gap, injection in a page, cost overrun at 10×, owner who will not maintain. Each with a mitigation that is a person plus a control, not a slogan. Sequence: shadow two weeks, assist if bars hold, no write tools, death date on any vendor POC. Ask: the last line of the table and the last line of the script. One ask. "I need a knowledge owner named this week and access to the Atlas space, not a bigger model." If you ask for five things, you will get a working group. You know what working groups do to programs.

Reuse a template your company already reads. If they use a one-page PID, this is a PID. If they use a steering slide with a RAID on the right, this is that slide with different rows. Do not show up with an "AI canvas" you downloaded. Novelty in template language is how skeptics stop listening. Novelty in the rows — eval, HITL, $ at 10×, kill — is how they start. You are not here to look like a futurist. You are here to look like the person who will still be around when the demo is over.

If someone said "you have five minutes," the pack is: slice in one sentence, eval bar, monthly $ at 10×, top risk, the ask. Everything else is an appendix. Time it: 45 seconds slice, 60 seconds eval, 60 seconds cost, 60 seconds top risk and mitigation, 45 seconds sequence, 30 seconds ask, 20 seconds silence. Silence is part of the script. Delivery leads who fill five minutes without an ask have run a briefing, not a decision.

Slice in one sentence: "Atlas program managers, live SOP from a named wiki space, citations or refuse, no write tools, 40 users." If you need a paragraph, you do not own the slice. Eval in one minute: who wrote the questions, what you score, the bar, what happens if you miss. Cost in one minute: now, 10×, the cap, who gets the alarm. Top risk: pick one for the room (usually stale corpus or legal path), mitigation, residual. Sequence: shadow, then assist, death date. Ask: owner and access, or a no.

This is a different rhythm from yesterday. Yesterday you were the builder walking a design. Today you are the PM/delivery lead walking a workstream. You will not draw six boxes unless they ask. You will not say "chunking." You will say "the knowledge owner writes forty questions; we do not put this in front of users until recall clears the bar; cost at ten times volume is X; the kill is Y; I need Z." Same use case. Different altitude. If you cannot switch altitude, you cannot sit in both rooms, and several of the five seats require both rooms.

Script a closer that is an ask, not a vision. Bad: "excited to partner to unlock value across the enterprise." Good: "I need Maya Chen named as knowledge owner this week, access to the Atlas space, and agreement that the POC tenant dies on day 11 unless you sign a production path with security. If that is not available, we should not start." The second closer is how a steering committee actually decides. It is also how an interview panel decides you have done this before.

Anticipate the CFO and the ops director as two different animals. CFO attacks cost, death date, and whether this becomes a platform invoice. Ops director attacks roster, HITL hours, and whether v1 creates more work for the seniors it was supposed to save. Answer both in the table before they ask, then let them attack. A pack that only answers the technologist is a demo in disguise. You have sat in those meetings. Do not recreate them from the other side of the table.

Practice with a slightly hostile voice in your head. "Why wouldn't we just buy Vendor X?" is tomorrow's pushback C, but it will show up in steering too. "Why isn't this in the Copilot license we already pay for?" is a real question. Answer: maybe it is, if Copilot can hit the eval bar on this corpus with ACLs; we will run the golden set there first; if it hits, we do not build. That answer is how you look like a buyer, not a builder looking for a toy. Put it in the sequence row: "measure the already-paid tool on our forty questions before we fund a slice."

You will get this from a VP, from a vendor, and from a hiring-panel actor playing a VP. The request is always the same: skip evals, skip data class, put something on a screen by Friday, "we will harden later." Later never comes. A demo without evals trains the org that fluent is finished. You have seen this with BI dashboards that became "the number" because they were on a wall. Do not do it with a generator that can invent an SOP.

Spoken answer, steal and rewrite: "We will not put an uncited generator in front of users as if it were the SOP. We will run a two-week shadow: the model answers in a side panel, the program manager does the job the old way, we score against forty questions the knowledge owner writes. You will see traces, not a magic box. If we miss the bar, we stop. If we hit it, assist-mode is the next conversation. I need a knowledge owner and access, not a bigger model, and not a Friday demo to the CEO."

Why this is leadership rather than blocking: you offered a thinner path with a clock, an artifact the VP can look at (traces, a table), and a kill. Blocking is "no, AI is dangerous." Leadership is "no to that object, yes to this object, here is the ask." Interviewers are scoring the difference. So are real VPs. If your answer is only no, you sound like InfoSec theatre. If your answer is only yes, you sound like the last person who landed them with a POC they cannot turn off.

If they insist on a screen: give them retrieve-only. Show the chunks, do not generate. That is a demo that cannot invent. It also teaches the room whether the corpus is even findable. Often it is not. That finding is worth more than a fluent paragraph. Script the line: "If you need something on a screen this week, it is search over the named space with sources, no generator. That is honest. A generator on a dirty corpus is a lie we would then have to unwind in steering."

If the interviewer is playing the VP as a bully, do not match energy. Repeat the bar, the clock, and the ask. Then stop. Delivery managers who argue for three minutes lose the room even when they are right. Delivery managers who put a bar on the table and go quiet win more than they expect. You already know this from change-control fights. Use that muscle. Do not invent a thought-leadership speech about responsible AI. The bar is the responsible-AI speech.

Write your version tonight in the voice you use when a program is late, not in the voice you use on LinkedIn. Record it. If you hear "leverage" or "journey," cut. If you hear a named owner, a number, and a death date, keep. This answer will also feed day 27 drill A (CEO chatbot by Friday) and day 29 question 6. You only want to author it once, well.

You will get this from enterprise architecture, from a vendor who sells a platform, and from a hiring-panel architect. The request is always the same: stand up a multi-tenant RAG platform, a prompt catalog, an eval harness, a gateway, an agent runtime, and a center of excellence — then, later, a use case. A platform with no slice is a catalog of slideware. You have seen this with integration layers that predated any integration. Do not fund it first.

Spoken answer, steal and rewrite: "We will fund two thin slices with evals, HITL, and a cost envelope. If they share a pattern — same ingest, same eval harness, same gateway — we extract a platform on a measured pattern and staff it. We will not staff a platform to hunt for a pattern. I will keep a list of what we are repeating so the second slice is cheaper than the first. That list is the start of a platform. A slide with boxes is not."

The mature sequence is: slice one in a contained tenant, slice two in a second team, then standardize what actually repeated. What usually repeats: a gateway to an approved model endpoint, a place to store golden sets, a logging pattern, an ACL tag convention. What rarely repeats on the first try: a universal ontology, an agent framework, a company-wide prompt library. Anyone selling you those in month one is selling a SOW, not a need. You may buy a gateway you already have. You may not pause the slice for a year-long platform program.

If they say "otherwise we will have twelve snowflakes," agree with the fear and disagree with the order. "We will have snowflakes if we never extract. We will have a platform nobody uses if we extract first. I have run both failures. The control is two slices with a written 'what repeated' list and a date to review extraction — not a freeze on delivery until architecture is happy." That sentence is how you keep architecture in the room without handing them the keys to the calendar.

If your primary is Transformation, this pushback is your home game. The 12-POCs-no-production case is this sentence in a larger coat: freeze new POCs, pick two, put evals on them, kill the rest, extract only what repeats. If your primary is FDE, still learn this answer; the customer will have an architect. If your primary is Solutions, this is how you stop your own company from overselling a platform the customer's first slice does not need.

Do not mock platform work. You will need it. Mocking it is how engineers stop helping you. Respect it as a second move. Put the extraction review on the sequence row of the pack with a date. That is the compromise a reasonable architect can sign. If the architect cannot sign anything short of a 14-month program, you have a political problem, not a technical one, and you should say so in the interview as a risk with an owner — not as a rant.

You will get this from counsel, from InfoSec, from a hiring-panel actor, and from your own fear. Sometimes it is true: restricted data, no approved path, no DPA, no. Sometimes it is a freeze that means "we have not looked." Your job is not to freelance a consumer API past them. Your job is to find the lightest approved path, put a sanitized POC on a clock, and if that still fails, stop without pretending you will sneak it in.

Spoken answer, steal and rewrite: "If Atlas is internal and not restricted, the lightest path is usually the enterprise-hosted endpoint we already have a DPA for, with logging, no-training clause, and no consumer tools. I will bring a data-flow that shows copies — index, logs, vendor, UI — and ask Security for that path, not for a new platform. If they say no, we do a sanitized corpus in an approved tenant with a death date. If that still fails, we do not do this use case. I will not put SOP text into a consumer chatbot and hope."

Bring a data-flow, a class, a retention story, and a banned-tools line. That is the security conversation from week 3, compressed. Interviewers who have been burned want to hear copies named. "The index is a copy. The prompt log is a copy. The vendor may retain for abuse monitoring." Then the control: default-deny raw logs for higher classes, locked eval store, named owner, burn date. You do not need to win a law degree. You need to show you will not be the person who pastes a claims file into a personal account.

If legal is a hard no on any hosted model, the path is VPC / on-prem / existing approved tenant, or kill. Do not offer to "use an open-source model on a laptop" as if that removed the data problem. A laptop is a copy. An unmanaged laptop is an incident. Say: "open-source on an unmanaged machine is not a control. If we have an approved on-prem path, we can talk. If we do not, this slice waits. There are other jobs-to-be-done that are public or already-approved. I would rather move one of those than freelance this."

Sanitized POC with a death date is the compromise. Dummy SOPs, no real employee names, a tenant that dies on day 11, eval questions that do not contain live data. You learn retrieve quality. You do not learn whether the real ACL map works — say that limitation out loud so nobody promotes the sanitized POC to production. Promotion requires the real class, the real ACL, and security on the real path. You have seen pilots become production. Put the death date in the SOW and in the steering table.

In an interview, they are scoring whether you protect the company without sounding thrilled to block work. Wear the face you wear when you stop an untested release. Calm. Specific. An alternative if one exists. No speech about regulation as a vibe. If you do not know the company's actual approved path, say so: "I would start by asking Security what the lightest already-approved path is for internal text. I would not open with a new vendor." That is a first-week move in the 90-day plan. It is also a correct interview answer.

Yesterday's demo story is how the slice works. Today's operating pack is how the workstream is governed, funded, and killed. FDE and solutions first rounds usually want the first. Delivery, transformation, and many second rounds want the second. Some rooms want both: demo first, pack if they lean in. Prepare both. Pick in the first 90 seconds of the room from the interviewer's title and the first question. Do not dump both unprompted — that is a status meeting, not an interview.

Signals you should lead with the demo: "walk me through something you built," "draw the architecture," "how would you implement this for a customer," a take-home that is a RAG, an interviewer who is an engineer or an FDE manager. Signals you should lead with the pack: "you have five minutes with the committee," "a VP wants a chatbot," "we have twelve POCs," "how do you handle legal," an interviewer who is a director of delivery or a transformation lead. If you are unsure, ask: "I can walk the slice design or the steering pack — which is useful here?" That question is itself positioning.

Store them as two documents, not one blob. Demo script with timestamps and the six-box diagram. Pack table plus five-minute script plus the three pushbacks. Day 28 will put both in the folder. Day 29 will use six minutes of demo and four of pack in the twelve-minute walkthrough. If they live in one endless note, you will ramble. You already know what happens when a RAID and a design doc are the same PowerPoint. Do not do that to yourself.

A single use case told two ways is stronger than two weak use cases. Do not invent a second product for the pack. Atlas wiki copilot is enough. Depth beats variety. Interviewers can probe one specific thing. They cannot probe a portfolio of slogans. If you have a real work program from your job that is more specific than Atlas, you may use that as the pack's "job" row — as long as you are allowed to, and as long as the eval/HITL/cost rows are still honest. Mixed proof is fine. Two shallow stories are not.

If your primary is FDE, you may be tempted to skip today. Do not. The second round will put you in a customer steering sim. The people who only prepped the six boxes freeze when a fake CFO asks about 10× cost. If your primary is Transformation, you may be tempted to skip yesterday. Do not. Someone on the panel will ask how the slice actually works, and "I do not do technical" is not a transformation answer in 2026. Both stories are in the job. Both stories are in the loop.

End today by writing two lines at the top of each artifact: "Use this when…" and "Do not use this when…" Example, demo: use when they ask how it works; do not use when they handed you a steering clock. Example, pack: use when they want a decision; do not use when they handed you a whiteboard marker. That is the whole skill. You will need it again on day 26 when the loop shape changes by seat.