Picture the Monday after you finish this level. You have read a vision claim skeptically, you can tell a camera that grades parts from a model that predicts a failure from a chatbot that drafts a procedure, you know the false-reject rate is real money, and you understand why a safety-critical control loop is not where AI goes. You are, in the language of this program, an AI-Aware Manufacturer. And then your plant manager, who toured a competitor's "smart factory" last week, stops you at the production board and says: "You've been doing that AI training. What are we actually going to do with it?" That question is where most reskilling dies. The honest answer for the person who has only finished L1 is not "let me architect a plant-wide digital twin." It is "give me ninety days and I will show you one verified workflow on one real loss, with a number attached." This lesson is the ninety-day plan that turns a freshly aware manufacturer into someone who has shipped something. It is built in three thirty-day phases, each phase ending in a tangible artifact you can put in front of that plant manager, because the entire forcing function of this program is that structured programs see three to four times the adoption of self-directed learning, and the difference is almost always whether the learner shipped something real or just watched videos.

Why Ninety Days, and Why One Loss

Ninety days is not an arbitrary quarter. It is the longest a plant manager will wait for evidence before the "AI plan" loses momentum, and it is the shortest honest window to go from skeptical reader to shipped workflow without skipping the verification discipline that keeps you out of trouble. The plan deliberately refuses to do two things that wreck most floor-AI starts. It refuses to boil the ocean, because a plant has many losses and a beginner who tries to fix all of them fixes none. And it refuses to start with a tool, because starting with a vendor demo means you let the tool pick the problem, and the tool always picks the problem it is good at, not the one costing you the most.

So the plan starts where every loss-reduction effort on a real floor starts: the loss chart. OEE (Overall Equipment Effectiveness, the single number that multiplies availability times performance times quality to express how much good product a line actually makes against its theoretical best) is the scoreboard, and underneath it sits a Pareto of why the number is not higher. On almost every plant that Pareto is dominated by two bars: unplanned downtime and scrap or rework. You are going to pick exactly one loss, the tallest bar you can realistically touch, and spend ninety days on it. Not because the others do not matter, but because a single shipped result on one loss is worth more to your credibility and your plant than five half-built experiments.

Here is the worked example that anchors the whole lesson, and every phase will return to it. Suppose your downtime Pareto shows a packaging line that loses roughly six hours a month to a recurring jam on one infeed, and your plant accountant has told you a stopped line on that product costs about one thousand dollars an hour in lost throughput and idle labor. That is around six thousand dollars a month, seventy-two thousand dollars a year, sitting in a single bar. You do not need AI to read a Pareto. But you do need the AI-aware judgment this level gave you to ask the right next question: is this loss a knowledge problem wearing a technology costume, the kind AI can multiply a thin crew against, or is it a mechanical problem AI cannot touch? That question, asked well, is the first deliverable.

Do not start with a tool, because the tool will pick the problem it is good at. Start with the loss chart, because it picks the problem that is costing you the most.

Phase One, Days 1 to 30: See Clearly

The first thirty days are not about building anything. They are about seeing your plant the way an AI project actually has to see it, which is almost always worse than you assumed. The deliverable at the end of Phase One is a one-page AI Opportunity and Risk Memo on your one chosen loss, and it is the same artifact the L1 capstone asks for, so this phase doubles as your capstone work.

Week 1, place the loss on the chart. Pull the real numbers, not the remembered ones. Get the downtime minutes from the historian or the line's downtime log, get the cost-per-hour from finance, and write the annualized dollar figure down. In the worked example that is the seventy-two thousand dollar jam. If your plant cannot produce these numbers, that is itself a finding, and it goes in the memo, because a loss you cannot measure is a loss you cannot prove you fixed.

Week 2, ask which of the three AIs even applies. This level taught you that vision, prediction, and generation are three different tools. Hold your loss up against each. The recurring infeed jam might be a vision problem if a camera could catch the misfeed forming, a prediction problem if sensor data trends toward the jam, or, very often, neither, because the real cause is a worn guide rail that a mechanic should replace. Being willing to write "no AI use case here, this is a maintenance fix" is a sign of competence, not failure. The memo names the candidate AI, or names honestly that there is none.

Week 3, audit the data and the floor reality. This is where most beginners' plans quietly die, so do it early and on purpose. For a vision idea, ask: do we have labeled images of the defect, under the lighting and material conditions of every shift, or would we have to create them? For a prediction idea, ask: does the historian actually log the relevant sensor, at a useful rate, with tags anyone can find? Remember the brownfield reality this level hammered: most plants run a 1990s PLC and a historian nobody has queried in years, and roughly seventy-eight percent of plants cannot even centrally monitor their OT (Operational Technology, the control-system network of PLCs, SCADA, and historians that actually runs the equipment, as opposed to the IT business network) environment. If the data does not exist, your ninety-day deliverable is the data-collection plan, and that is a perfectly legitimate, shippable result.

Week 4, run the safety and OT check, then write the memo. Apply the boundary from the safety lesson: is your idea advisory or does it touch a safety-critical control loop? If it touches one, it is dead in this form and the memo says so. Apply the OT boundary: does the idea require putting a model near the PLC, or can it live safely on the IT side reading a copy of the data? Then write the one page. It states the loss in dollars, the candidate AI or the honest "none," the data and verification needs, the safety and OT check, and a clear go or no-go recommendation. That memo is what you hand the plant manager at day 30. It demonstrates that you can read a loss skeptically, which is exactly the L1 capability, and it costs the plant nothing but your judgment.

Phase Two, Days 31 to 60: Prove Small

If Phase One ended in a go, Phase Two is where you do the smallest possible real thing, by hand, before anyone buys anything. The deliverable is a verified manual proof: a tiny, human-run version of the workflow that shows the AI-aware concept works on your actual data, with the verification step built in from the first hour. The point is to learn the failure modes on a scale where a mistake costs an afternoon, not a containment.

The cardinal discipline of this entire program governs Phase Two: the job shifted from producing the draft to verifying the draft against the drawing, the standard, and the historian. So whatever small thing you prove, you prove it with a human verifying every AI output, and you measure how often the AI is wrong, because that number is the whole game.

Take the most common Phase Two for an L1 graduate, the generative case, because it needs no new hardware and no OT access at all. Suppose your chosen loss is the slow, error-prone drafting of work instructions and maintenance write-ups that eats your green crew's time and produces inconsistent records. Your manual proof is this: take ten real maintenance notes a tech scribbled this month, use an AI assistant to turn each into a clean, structured CMMS (Computerized Maintenance Management System, the software that holds work orders, asset histories, and the preventive-maintenance schedule) work-order record, and then sit with the tech and verify each one against what actually happened. Count the errors. Maybe the model invents a torque spec it was never given, or names the wrong part number, or omits the safety step the tech mentioned. You now have a real, measured error rate on your real data, and you have proven both that the drafting saves time and that the verification is non-negotiable. That is a far stronger thing to show a plant manager than any vendor's accuracy slide, because it is your data and your number.

For a vision-flavored loss, the small proof might be running a batch of your own labeled defect images through a trial of an off-the-shelf vision tool and computing, on a holdout set the tool never saw, how many good parts it would have falsely rejected and how many real defects it would have missed. You are not deploying anything. You are getting the false-reject and false-accept numbers, in dollars, on your parts, under your lighting, exactly the skepticism this level trained. If the false-reject rate would cost more than the escapes it catches, you just saved the plant a bad purchase, and that, written up, is a shipped Phase Two deliverable that is arguably more valuable than a success.

The rule for Phase Two is to keep the blast radius tiny and the verification total. Ten records, not ten thousand. One product, one defect, one shift's lighting. A human checking every single output. The artifact at day 60 is a short write-up with the measured numbers: time saved per record, the AI error rate you found, the verification step that catches those errors, and a recommendation about whether this is worth scaling. You are now someone who has run a real AI experiment on a real plant loss and knows precisely how often the AI lies.

Phase Three, Days 61 to 90: Ship One Workflow

Phase Three turns the proven concept into one workflow that one team actually uses, with the human verification step baked in as a required gate, not an optional courtesy. "Ship" does not mean plant-wide rollout. It means one real, repeatable, documented workflow running on one line or one team, producing a logged result. The deliverable is the workflow itself plus the first evidence that it moved the number on your chosen loss.

Returning to the maintenance-record example: shipping means the work-order drafting flow is now the standard way that one maintenance team writes its records. The tech speaks or types rough notes, the AI drafts the structured CMMS record, and a defined person verifies it against the job before it is saved, with that verification logged. You write a one-page standard work that names who does what, where the verification gate sits, and what to do when the AI output is wrong. You run it for the back half of the month and you measure: average minutes saved per record times records per month, and the consistency improvement in the records themselves. If you saved each tech twenty minutes a record across two hundred records a month, that is roughly sixty-six hours of skilled time returned to the floor every month, time a chronically short maintenance crew can spend on the preventive work that prevents the next hot-afternoon breakdown.

Three disciplines make a Phase Three workflow real instead of a demo, and all three come straight from this level. First, the human gate is non-negotiable and visible. Anyone looking at the workflow can point to the exact step where a named human verifies the AI output before it has any consequence. That is what keeps you compliant with the cardinal rule that the customer audits you, not the vendor, and it is what an auditor or a customer will look for first. Second, you log the result. A workflow that saves time but produces no record of having done so cannot defend itself in the next budget meeting, so you keep the simple before-and-after numbers, the same way a predictive-maintenance program logs the avoided downtime when a save lands in the CMMS. Third, you watch for drift. Even a generative workflow degrades as the work changes, and a vision workflow degrades as lighting and material drift between shifts, so the standard work includes a periodic re-check of the AI's error rate, because the moment a workflow stops being verified is the moment it stops being trustworthy.

The artifact at day 90 is the thing the plant manager asked for at day zero, made concrete: one workflow in use, a documented human verification gate, a logged result with a dollar or hours number attached, and an honest note on its limits and its drift plan. You have gone from reading a vision claim skeptically to shipping one verified workflow, exactly the arc this level promised, and you have done it on a real loss with a real number, which is the only kind of evidence that earns the next ninety days.

The Traps That Derail the Ninety Days

Most ninety-day plans fail in the same handful of ways, and naming them is how you avoid them. The boil-the-ocean trap is trying to fix three losses at once because they all look urgent; you end with three unfinished pilots and no shipped result, and a half-built workflow is worth nothing to a budget meeting. Pick one loss and starve the rest until day 90. The tool-first trap is letting a vendor demo set your agenda; the demo will be impressive and will quietly redefine your problem to match the product, so you keep the loss chart, not the demo, as your north star. The skip-verification trap is the most dangerous, because it is the one that ships a hallucinated torque spec or a wrong root cause to a customer; the verification step is not the slow part of the workflow you optimize away, it is the part that makes the workflow legal and safe to run at all.

The no-number trap is shipping a workflow that "feels faster" without a measured before-and-after; feelings do not survive the next downturn's budget review, but seventy-two thousand dollars of addressed annual loss does. The data-denial trap is assuming the historian has clean, labeled, queryable data when it does not, then discovering in week eight that the project was impossible from day one; the Phase One data audit exists precisely to surface this in week three when it is cheap. And the safety-and-OT shortcut is the trap that ends careers rather than projects: skipping the advisory-versus-in-the-loop check or bolting a model onto the control network to save an integration step. The ninety-day plan front-loads both checks into Phase One for a reason. A workflow that is fast and impressive and crosses the safety line is not a win; it is the incident you spend the next ninety days explaining.

What You Can Show, and What It Earns

At day 90 you can stand in front of the plant manager and say the sentence this whole program is built to let you say: "Here is the loss I picked and what it costs us, here is the AI error rate I measured on our own data, here is the verification step that catches those errors, here is the one workflow now running, and here is the time or money it has returned in its first month." That sentence is the credential. It is worth more than any certificate, because it is evidence, on your plant, in your numbers, that you can deploy AI for a real loss and verify it well enough to defend in front of a customer.

It also positions you for the roles this level's career lessons described. The quality engineer who shipped a verified vision triage, the reliability tech who logged a real avoided-downtime save, the continuous-improvement lead who turned a kaizen analysis into a shipped workflow, the plant data person who built the data-collection plan that made the rest possible: these are the titles that pay for this skill, and a single shipped ninety-day workflow is the proof that you can fill one of them. The talent cliff that drives this whole program, roughly two million workers needing reskilling against about five hundred thousand unfilled roles, is not just a threat to the plant; it is the reason a manufacturer who can demonstrably ship and verify floor AI is suddenly one of the most valuable people in the building.

And the ninety days are a loop, not a finish line. The workflow you shipped becomes the foundation for the next loss on the Pareto, and the verification discipline you built becomes the habit you carry into L2, where you start using AI for quality triage, maintenance records, work instructions, and root cause in earnest. The on-ramp does exactly what an on-ramp does: it gets you up to the speed of the road. What you do once you are on it is the rest of this program.

Key Takeaways

• The honest answer to "what will we do with this AI training" after L1 is not a plant-wide platform; it is one verified workflow on one real loss in ninety days, with a number attached. Structured, ship-something programs see three to four times the adoption of self-directed video-watching, and shipping is the difference.
• Start with the loss chart, never with a tool. The OEE Pareto picks the problem that costs the most; a vendor demo picks the problem the product is good at. Choose one loss, the tallest bar you can realistically touch, and starve the rest until day 90.
• Phase One (days 1 to 30): see clearly. Place the loss in dollars, test it against vision, prediction, and generation, audit whether the data and labels even exist in a brownfield plant, run the safety and OT check, and deliver a one-page AI Opportunity and Risk Memo, which is also the L1 capstone. Writing "no AI use case, this is a maintenance fix" is competence.
• Phase Two (days 31 to 60): prove small. Run the smallest manual version on your real data (for example, ten maintenance notes drafted into CMMS records and verified by the tech) and measure the AI's actual error rate. A trial that shows a vision tool's false-reject rate would cost more than the escapes it catches is a valuable shipped result, not a failure.
• Phase Three (days 61 to 90): ship one workflow on one team or line, with the human verification gate baked in as a required, visible step, the result logged with a dollar or hours number, and a drift re-check scheduled. Twenty minutes saved across two hundred records a month is about sixty-six hours of skilled time returned to a short crew.
• Six traps derail the plan: boil-the-ocean, tool-first, skip-verification, no-number, data-denial, and the career-ending safety-and-OT shortcut. The plan front-loads the data audit and the safety and OT checks into Phase One precisely so they surface while they are still cheap.
• The day-90 credential is a single sentence backed by evidence: the loss and its cost, the measured AI error rate on your own data, the verification step, the running workflow, and the time or money returned. That evidence, not a certificate, is what earns the next ninety days and the roles that pay for this skill.
• The on-ramp is a loop. The shipped workflow seeds the next loss on the Pareto, the verification habit carries into L2, and the manufacturer who can demonstrably ship and verify floor AI becomes one of the most valuable people in a building short five hundred thousand workers.