The map: where a delivery manager actually lands

When people say they want to get into AI, they usually picture training models from scratch or joining a research lab. That market is real. It is also not the market that is hiring delivery managers in volume. The seats that keep showing up in 2025 and 2026 requisitions are the ones that make a model useful inside a company: scoping the first slice, getting data into a shape the system can use, writing evals a VP will believe, managing vendors who will change the model name twice a year, and getting something live in a customer's environment without lighting the RAID log on fire. Those jobs have titles. None of them are 'AI researcher.'

Five seats show up again and again. AI Delivery / Program Lead is the closest leap: same muscles (plan, RAID, stakeholders, go-live), new object (a model-backed workflow). Forward Deployed Engineer is the highest-leverage seat if you will write a little code and sit at the customer. Solutions Engineer / Architect fits if you already like pre-sales, whiteboards, and turning a vague RFP into a thin slice. AI Product Manager fits if you already sit with engineering on scope and can write a spec that survives contact with a model. AI Transformation / Value Lead fits if you already run portfolios and can say no to executives who want a chatbot of everything.

You do not need to become a researcher. You need to become the person who can hear 'we want a chatbot' and answer with a thin slice, an eval, a cost, and a kill criteria — in language both a VP and an engineer respect. That sentence is the whole job. The VP hears: we will not boil the ocean, we will measure, we will stop if it does not work. The engineer hears: you will not ask them to 'just hook up GPT' with no retrieval, no eval, and no owner. If you can say that sentence without hedging, you are already ahead of most career-switch decks.

The wrong default is 'AI engineer' because that seat optimizes for training, serving, and deep systems work. Job posts that say 'AI engineer' often mean three different things, which is why people get confused. Sometimes they mean ML engineer (train and deploy models). Sometimes they mean someone who can call an API and wire a RAG. Sometimes they mean the delivery lead and they used the wrong title. Read the actual bullets. If the post asks for CUDA, research publications, or years of PyTorch, it is not your post. If it asks for customer workshops, evals, data readiness, and a first working slice, it is.

Pick one primary and one backup. Primary is the seat you will write artifacts for this month. Backup is the seat you can walk into if the primary market is slow in your city or your company. Do not pick all five. A hiring manager who hears 'I could do delivery, or FDE, or PM, or transformation' hears someone who has not chosen. Choosing is a delivery skill. You already force sponsors to pick a first slice. Do that to yourself today.

Your unfair advantage is the mix. Engineers under-index on stakeholders, RAID, and change. MBAs under-index on how the system actually fails — hallucination, retrieval miss, unbounded agent loops, a prompt with no owner. A delivery manager who can talk tokens, RAG, and evals is rare. That rarity is the whole bet of this month. You are not competing with people who trained models at a lab. You are competing with other operators who cannot yet draw the loop on a whiteboard.

AI literacy is knowing what a model is, what a token costs, why RAG exists, and when an agent is a bad idea. It is the ability to sit in a steering meeting and not get snowed by a vendor slide that says 'our model is smarter.' Literacy is not a certificate. It is a set of sentences you can say without looking at notes: next-token, context window, grounding, eval, unit cost. This week is mostly literacy. If you skip it, week 2's thin slice will feel like magic and you will not be able to debug it.

AI building is calling an API, chunking documents, writing an eval set, wiring a tool. Week 2 is a deliberately thin slice of building — enough to be dangerous, not enough to fake a CS degree. You will read Python the way you already read a process pack: find the named steps, find the gates, find the one constant you would change. You will not grind LeetCode. You will not train a model. You will produce a path you can talk through: load, chunk, retrieve, generate, score. That is the building bar for FDE and solutions. Delivery leads can sit slightly lower; they still need to read the path.

AI delivery is the job you already have, pointed at a new object: use-case intake, data readiness, human-in-the-loop, vendor lock, SLA, go-live. Week 3 is where you stop being a student and start being the lead. The work looks like every other program — RAID, cutover, comms, hypercare — except the object fails in new ways. It invents a policy. It blows a token budget. It cannot see last Tuesday's SharePoint. Delivery is the layer that notices those failures before the demo becomes a production incident.

Confusing the three layers is how people waste a year. They stay in literacy (podcasts, newsletters, 'I follow the space') and never produce an artifact. Or they jump to building without literacy and copy a RAG tutorial they cannot explain. Or they jump to delivery theatre — a transformation office, a 40-use-case heatmap — with no eval and no thin slice. This month is stacked on purpose: literacy, then a little building, then delivery, then packaging for a hiring loop. Do not rearrange it because building feels more impressive on LinkedIn.

In a meeting, you should be able to name which layer a request belongs to. 'Can you explain why it made that up?' is literacy. 'Can you add our policy pack to the prompt?' is building (retrieval). 'Can we go live for 2,000 users next month?' is delivery (evals, cost, change, support). When a sponsor mixes them — 'just make it know our data and roll it out' — your job is to unmix. Write the three layers on the whiteboard. Put the request in one box. That move alone will save you from a chatbot-of-everything program.

Week 4 packages whatever you actually did. Packaging is not a fourth technical layer. It is the hiring loop: resume, walkthrough, 90-day plan, the artifacts from the first three weeks. If you skip literacy, your walkthrough will sound like a blog. If you skip building, an FDE interviewer will notice in ten minutes. If you skip delivery, you will look like a prompt hobbyist. Stay stacked. The walkthrough is a test of the stack, not of your ability to list tools. A hiring manager will ask you to draw the path. Draw it from the stack you actually studied.

Two hours a day × 30 days = 60 hours. That is roughly a 3-credit university course, compressed. It is enough to change what you can say in an interview. It is not enough to make you a research scientist, and you should not pretend it is. The honest sentence is: after this month you can design a thin slice, write a prompt library, estimate cost, sit in a vendor meeting without getting rolled, and walk a hiring manager through a worked use case. That is a pivot, not a new identity.

Each day has a session plan that already adds to 120 minutes. Follow it. Reading without the practice is how people 'learn AI' for six months and still freeze when asked to design a thin slice. The practice is not homework in the school sense. It is the thing a hiring manager can probe. If you only have 90 minutes, cut further reading, not the write. The write is the point. A half-finished charter you can talk through beats a perfectly highlighted article you cannot. Protect the write the way you protect a decision log on a live program.

You will produce six artifacts, not certificates: a prompt library, a wiki-copilot design, a use-case charter, a role-targeted resume, a 10-minute walkthrough script, and a 90-day plan. Those are what a hiring manager can probe. A Coursera badge without a worked use case does not move a delivery-manager pivot. If you only have time for one thing this month, protect the artifacts, not extra videos. Videos feel like progress. Artifacts survive a screen. Put the six on a list today and tick them as they appear. The list is the month.

The certificate trap is specific. Platforms sell completion. Hiring managers buy evidence. Completion is 'I finished 17 hours of LLM Ops.' Evidence is 'here is the eval set I used to kill a chatbot-of-everything, here is the cost at 10×, here is the human-in-the-loop on the write path.' You can collect both. If you have to choose, collect evidence. The trap is using certificates as a substitute for the awkward work of writing a charter with a kill criteria. Awkward work is the job. A badge is a receipt that you sat down. Keep the receipt if a bench requires it later. Do not confuse it with proof.

Sixty hours is also a constraint you should say out loud. You will not learn CUDA. You will not become the person who trains the next model. You will not master every vendor. You will get a working vocabulary, a small set of diagrams you can redraw, and six artifacts. That is the deal. People who secretly want the degree will be unhappy by day 9 and will wander off into a 40-hour specialization they will not finish. Stay on the 60-hour path.

Lab days (7, 14, 21, and the capstone stretch) are where the artifacts get made. Do not skip lab days to 'catch up on reading.' Reading is cheaper dopamine. The lab is the product. If you miss a lab, you do not have a week. You have notes. Notes do not get you into the interview. The prompt library on day 7 is the first thing that can leave the study and enter the job on Monday. Protect it the way you protect a go-live window: reading can slip a day; the artifact date cannot. If you are behind on Tuesday, skip further reading, not Friday's lab.

Same time each day. Phone in another room. The session plan is the agenda — you do not 'see how you feel.' If you miss a day, you do not double up the next day past 2.5 hours; you slip the calendar by one and still finish. Doubling up is how people create a six-hour Saturday they will resent and then abandon. A delivery manager already knows this from programs: a slipped workshop is recoverable, a death-march weekend is not a plan.

Each block is either read, practice, or write. The write is the point. Notes in this app are persisted on this device — treat them as the start of your portfolio, not a diary. Date them. Put the day number at the top. When you rewrite positioning on day 25 you will want v0 from today, not a vague memory of having been motivated. The study OS is a filing system plus a calendar, not a mood. If a note has no date and no day number, it is a diary entry. Diary entries do not survive a screen-share. File it as if a deputy had to find v0 in thirty seconds.

Weekly lab days (7, 14, 21, and the capstone stretch) are where the artifacts get made. Do not skip lab days to catch up on reading. If you are behind, skip further reading on a non-lab day, not the lab. The session blocks already add to 120 minutes. The first block today is a slow read on purpose. Later days keep that pattern: a long read, a diagram redraw, then the write. The redraw is not artistic. It is a check that the picture is in your head, not only on the page.

Protect the two hours the way you protect a client workshop. Title the calendar hold with the day number. If a real fire eats the hold, reschedule within 24 hours. Do not 'catch up later this week' without a new hold. You already know what happens to tasks that live on a someday list. This month is a 30-day program with a named end. Treat it like one. The people who finish are not more talented. They have holds that survived day 9, when the novelty is gone and the content gets more technical.

What to do with energy that is not two hours. If you have a spare 20 minutes, do not start a new course. Re-draw yesterday's diagram, or re-read the check questions, or say the VP paragraph out loud. Spaced recall beats extra input. If you have a spare Saturday, do not add a specialization. Polish the artifact that is currently ugly. The protocol is also a no: no second MOOC, no 'I'll also learn LangChain this weekend,' no certificate side-quest until the six artifacts exist.

You will want to skip the redraw block because it feels like school. Do not. Whiteboard interviews and steering meetings are redraw tests. The person who can draw next-token, context budget, and the four interfaces without notes is the person who gets believed. Ten minutes a day of redraw is cheaper than looking at the ceiling when a VP says 'so it knows our data, right?' If the drawing is wrong, the meeting is already lost, no matter how many articles you have read. Redraw is the cheapest rehearsal you will get all month.

The hiring market for 'someone who can make AI useful here' is loud and uneven. Headcount showed up first in the big consultancies, the model vendors' own deployed teams, and the enterprises that already had a data office. Then it showed up in every company that bought a ChatGPT Enterprise seat and discovered that a seat is not a program. Requisitions use a mess of titles: AI Delivery Lead, GenAI Program Manager, Forward Deployed, Solutions, Copilot PM, Transformation. Read past the title. Read whether they want a builder in the customer environment, a PMO with evals, or a slide owner.

What they are actually screening for, in this order: (1) Have you shipped something with a messy stakeholder set. (2) Can you talk about failure modes without magic words. (3) Do you have an artifact. (4) Will you pick a first slice and kill the rest. Domain helps — a delivery manager who has run finance close, claims, or a plant outage will beat a generic 'AI enthusiast' on a use case in that domain. Do not throw away your domain to look more AI. The domain is why they can put you in front of the customer on week two.

What they are not screening for, despite the job-post noise: a master's in ML, a Kaggle rank, a fine-tune you ran on a laptop. Some posts copy-paste an ML engineer description onto a delivery role. That is a broken post, not your homework list. If the interview panel is a hiring manager from the business plus one engineer, bring the charter and the cost envelope. If the panel is three ML engineers and they want you to derive attention, you are in the wrong loop. Walk away faster than you think.

Compensation follows the seat, not the word AI. FDE and solutions at a vendor or a serious consultancy pay like implementation plus scarcity. Internal AI delivery leads often pay like the PMO / program band you already know, sometimes with a bump, sometimes with more politics than money. Transformation titles can be senior and vague. Ask what budget they own and what they have shipped. A title with no shipped slice is a committee. You already know how to detect that. Detect it in the interview, not six months in.

Internal mobility is a real path and a slower one. If you are inside a company that is 'doing AI,' the fastest proof is a prompt library and a thin slice on a workstream you already own — status, RAID, knowledge search — with an eval and a cost. That is how you get air cover to move. Applying cold to FDE roles without a single worked example is how you join a pile of operators who all say they are 'passionate about GenAI.' Passion is not a filter. A walkthrough is.

The market will change while you study. Model names will change. A vendor will announce an agent platform. None of that changes the seat map. The work remains: slice, retrieve or tool, eval, cost, change, kill criteria. Study the work. Skim the announcements. If you find yourself refreshing launch blogs instead of writing the positioning sentence, you have left the protocol. Launch blogs are designed to make you feel behind. The protocol is designed to make you hireable. Stay on the protocol even when a keynote says everything is different this week. It is not.

Enough technical depth is not 'I can train a model.' It is 'I can stay in the room when the builder talks, and I can tell when they are hand-waving.' Concretely: you can draw next-token prediction, you know what a token costs, you know why stuffing 200 pages into the window fails, you can tell chat from extraction from search from a write-tool, and you can read a 40-line Python script that calls an API. That is the bar this month is built for. Past that bar, you have specialists.

In an interview they will test depth without announcing it. They will ask why the copilot invented a clause. They will ask what you do when legal will not send data to a US endpoint. They will ask what happens to cost if the agent takes eight steps. They will put a prompt on the table and ask what is missing. None of those require you to implement attention. All of them require you to have a mechanical picture, not a blog-post picture. If your answers start with 'the AI thinks' or 'we just need a better model,' you have failed the depth test.

Enough also has a ceiling you should respect. You will not be the person who debugs CUDA OOM errors. You will not be the person who chooses a PEFT method. You will not win an argument with a research scientist about tokenization edge cases, and you should not try. The failure mode of this month is fake depth: jargon used to hide that you cannot draw the picture. Prefer a short accurate sentence over a long impressive one. Delivery people who try to sound like researchers get found out. Delivery people who sound like adults with a whiteboard get hired.

A useful test: take a vendor claim and rewrite it as a delivery claim. Vendor: 'Our model has a million-token context.' Delivery: 'We still retrieve the two right pages because stuffing is slow, expensive, and noisy; the window is a budget, not a filing cabinet.' Vendor: 'Our agent can use tools.' Delivery: 'Write-tools need authz, idempotency, audit, and a human on money-moving actions.' If you can do that rewrite in thirty seconds, you have enough depth for the seats on today's map.

Another test: sit with an engineer for fifteen minutes on a failing answer. Ask where the retrieved chunks came from, what the system prompt actually says, what the eval set looks like, and what the trace shows for that call. If those questions are fluent, you are useful on an FDE or AI PM squad. If your only move is 'maybe we fine-tune,' you are not. Fine-tune is the last lever, not a personality. Enough depth includes knowing which lever is first.

Write your personal depth sentence today and put it under your positioning line. Example: 'I can design the slice, the eval, and the cost; I can read the path; I will not pretend I train models.' That sentence keeps you honest in week 2 when building feels like you should already know more Python. You are not behind. You are at the bar this month set. The bar is enough to be the adult in the AI room. It is not enough to be the model owner. Do not confuse those jobs.

An artifact is a document or working path a hiring manager can open and argue with. A badge is a record that you completed a vendor's module sequence. Both can exist. Only one survives a 45-minute loop. This month is an artifact factory with some literacy attached. If you finish the month with six completed modules and no charter, you did it backwards. The loop will not ask you to recite a module list. It will ask you to walk a use case, name a kill criteria, and say what you would do in week one. Badges do not help you answer that.

The six artifacts, in order of appearance: (1) positioning v0, rewritten on day 25. (2) Delivery prompt library, day 7. (3) Wiki-copilot design, week 2. (4) Use-case charter with eval and kill criteria, week 3. (5) Role-targeted resume plus 10-minute walkthrough script, week 4. (6) 90-day plan. Each one is short on purpose. A hiring manager will not read a 40-page playbook you wrote for yourself. They will read a one-page charter and ask you to walk the diagram. If an artifact is longer than two pages, it is not an artifact yet. It is a draft of a playbook. Cut it until a stranger can probe it in five minutes.

What makes an artifact probe-able: a specific use case, named constraints, numbers, a failure mode, and a decision you would make. 'A chatbot for HR' is not an artifact. 'A policy copilot for 200 managers, retrieve-then-generate, citations required, human review on anything that changes a contract, kill if groundedness on the eval set is under 80 percent, cost envelope $X at 10×' is an artifact. You already know how to write a charter. This is that genre with new fields.

Badges still have a narrow use. If a company requires a vendor's certification to get on the partner bench, do it later, after the artifacts, for that bench. Do not collect badges in parallel with this month 'just in case.' Parallel collection is how the two-hour hold dies. The study OS has a no-list. Extra certificates belong on it until day 30. After day 30, if a named bench requires a named badge, schedule it as a separate two-week job with its own holds. Do not smuggle it into this month and then wonder why the charter is still a heading.

Show artifacts at the job you already have, not only in interviews. The prompt library can reduce the time you spend on steering packs next Monday. That is not a side quest. That is how you get a story that is not hypothetical. 'I standardized the status digest and cut an hour off the pack, here is the eval I used when it tried to paint us green' is a better story than 'I am upskilling in GenAI.' Delivery people who reduce pack time get air cover. Delivery people who talk about upskilling get a lunch-and-learn slot.

File them like a program. One folder, six named documents, dates, versions. When you rewrite positioning on day 25, keep v0. Interviewers like people who can show a revision. It proves you treat this like work, not like a mood. You already know how to version a RAID log. Version the portfolio the same way. If a deputy cannot find the current charter in thirty seconds, you do not have a portfolio. You have a desktop. Name the files so a stranger could pick the right one: day, artifact, version. That stranger is a hiring manager on a screen share.