Skip to content

Lesson 1 of 9 in AI in manufacturing

Where AI pays first on the factory floor, and the bar it has to clear

Start where AI pays on the factory floor: predict part failures or catch defects, and fund the project when downtime costs more than the model.

2-minute read, 20-minute exercise. Updated . Free lesson, no account needed.

You leave with: First floor AI project card (blank card, worked example), three AI prompts.Blank card (PDF)The card as text, with a worked example

In 60 seconds

  • Start with the expensive machine. Predictive maintenance pays wherever there are fleets or expensive equipment: predict part failures, service only when needed, keep spares on hand.
  • Weigh defect detection by its reach. Training on historical data broadens what can be detected, and you don't have to spell out exactly what to look for.
  • Go live with minimal disruption. Availability and throughput come first on the floor, so the model goes live through disciplined machine learning operations.
  • Watch for drift from day one. Plan for continuous drift and performance monitoring from the day the model goes live.
  • Hold irreversible steps to first-time-right. Where a step cannot be redone, AI assistance must pass rigorous first-time-right validation rather than iterate and fix.

A free lesson from AI in manufacturing, one of its 9 lessons. Read another free lesson

What an hour of downtime costs

An hour of downtime in a large automotive plant now costs $2.3 million, Siemens' 2024 True Cost of Downtime report finds. That stopped hour is the money your first project on the floor should chase. Two candidates compete for the budget line: predicting part failures and detecting defects.

1. Start where the machine's downtime costs more than the model

Predictive maintenance pays off wherever there are fleets or expensive equipment, from mining to aviation to local government. You service the machine only when needed, predict part failures, and keep spares on hand. When to act on each prediction is a decision for later. The pick itself is settled by one test: the machine's downtime has to cost more than the model.

Defect detection is the other candidate. Training on historical data broadens what can be detected. You no longer need to spell out exactly what you are looking for.

2. What changed on the floor

In classic machine learning, a moved sensor or a shifted use case meant a restart: the model went back to square one. Foundation models generalize. The practical physical-AI wins include anomaly detection and predicting when to take a million-dollar machine offline for service. If an analytics project on your floor stalled years ago, ask whether restarts killed it.

3. Go live with minimal disruption to production

On the floor, availability and throughput come first. Your AI project has to respect that.

  • Go live through disciplined machine learning operations, so the disruption to production is minimal.
  • Monitor continuously for drift and performance, from the day the model goes live.

4. The bar for steps you cannot undo

Once a chip design goes to manufacturing, there is no patching. It works or it does not. A lot of money rides on it. For a step like that, hold AI assistance to rigorous first-time-right validation rather than iterate and fix.

Physical operations run on low margins and are often life-or-death. The quality bar is far higher than for a consumer app. A system that misses too many failures, raises too many false alarms or invents facts is unacceptable there. The potential economic and social value is bigger too. Before a model touches an irreversible step, name the first-time-right validation it passed.

Exercise (20 minutes): fill in the first floor project card

Pick the machine or defect detector you would actually fund first, and fill in the card for it. You won't know every line; some numbers sit with finance, the plant controller or maintenance. Every line you can't fill gets an owner and a date before you close the card. The card download includes a filled-in example of a project that isn't ready yet, and why.

First floor AI project card (PDF)

A blank card and a filled-in example, one page each.

Open the card (PDF)
Show the three prompts: fill the card, test it, turn it into a plan

Use AI to fill in and test your card

These three prompts go into your company's approved AI assistant. Use the first to fill in the card one row at a time; it won't guess a number for you. Use the second when the card is done, to hear the case against funding yet. The third asks about your industry and your plant and turns the lesson into a 30-day plan, with a checklist, rubric or brief for your boss if you want one. Use placeholders for company, system and people names; the help doesn't depend on them.

Prompt 1: Fill in my card with me

Paste this, then add your candidate project in one or two sentences and the card below.

Prompt
Before you start: use your company's approved AI assistant, not a personal account. Replace company, system, supplier and people names with placeholders (Company A, System X) and round any figures. The help you get does not depend on them.

You are helping me fill in a card for the first AI project I will fund on the factory floor. The card is pasted at the end.

Rules:
- Ask me about one row at a time, in order, and wait for my answer before moving on.
- Write down only what I tell you. Never suggest, estimate or invent a number, a name, a yes/no answer or a decision for me, even if I ask you to guess. If I ask, say that this row needs a real answer and ask who could give it.
- If I don't know an answer, write "Not known" and ask who should find out and by when. Put that in the Owner / by when column.
- After the last row, show the whole card as a table with three columns: Question, Answer, Owner / by when. Then list every row marked "Not known" with its owner, and remind me that any unanswered row means the project is not ready to fund.

My candidate project in one or two sentences: [DESCRIBE THE MACHINE OR DEFECT DETECTOR]

The card:
[PASTE THE CARD]

Prompt from The AI Decision Course by Dr. Tony Hoang, host of The Artificial Intelligence Podcast (Spotify, Apple Podcasts): tonyphoang.com/learn

Prompt 2: Stress-test my card

Paste this, then your filled card.

Prompt
Before you start: use your company's approved AI assistant, not a personal account. Replace company, system, supplier and people names with placeholders (Company A, System X) and round any figures. The review does not depend on them.

Below is a card I filled in for the first AI project on a factory floor. Act as a skeptical reviewer at the decision meeting and argue the case for NOT YET.

For each row:
- Say whether the answer is specific enough to act on (a number, a named role, a date) or too vague.
- Point out where answers contradict each other, for example the downtime cost against the model cost, or an irreversible step with no validation named.
- Say what evidence you would ask for before accepting it.

Do not rewrite my answers and do not invent facts about my plant. If a row is blank or says "Not known", treat it as not ready.

Finish with: your recommendation (FUND PREDICTIVE MAINTENANCE, FUND DEFECT DETECTION or NOT YET), the three rows that most need work, and one question I should ask my team on Monday.

My card:
[PASTE YOUR FILLED CARD]

Prompt from The AI Decision Course by Dr. Tony Hoang, host of The Artificial Intelligence Podcast (Spotify, Apple Podcasts): tonyphoang.com/learn

To paste the card into prompt 1, copy its rows from the PDF.

Prompt 3: Make it yours

Paste this and answer its four questions.

Prompt
Before you start: use your company's approved AI assistant, not a personal account. Replace company, system, supplier and people names with placeholders (Company A, System X, Supplier B).

You are helping a leader apply one lesson to their own situation. The lesson's five decisions are below. Use only them and what I tell you. Never invent facts, numbers, vendors, laws or results about my organization or my industry. Where a number is needed and I have not given one, say what to measure and who should measure it, instead of guessing.

First, ask me these four questions, one at a time, and wait for each answer:
1. My industry, and roughly how many people work in my organization.
2. The floor project I have in mind, in one or two sentences.
3. Where it stands today: an idea, a pilot running, or about to launch.
4. What worries me most about it.

Then give me the following. Keep it short and practical:
A. My plan. For each of the five decisions: one action written for my project, the role that should own it (a role, not a name), and what should be done within 30 days. If my industry makes a decision harder, for example regulated production, safety-critical steps or a single expensive line, say which decision and why in one sentence.
B. Three questions to ask my team on Monday, specific to my project.
C. Ask me which one I want: a checklist for my project, a rubric to score the project against the bar, or a one-page brief for my boss. Then write it, and finish with a prompt I can paste later to update it.
Mark anything that rests on an assumption with "(check)".

The five decisions:
1. Start with the expensive machine. Predictive maintenance pays wherever there are fleets or expensive equipment: predict part failures, service only when needed, keep spares on hand.
2. Weigh defect detection by its reach. Training on historical data broadens what can be detected, and you don't have to spell out exactly what to look for.
3. Go live with minimal disruption. Availability and throughput come first on the floor, so the model goes live through disciplined machine learning operations.
4. Watch for drift from day one. Plan for continuous drift and performance monitoring from the day the model goes live.
5. Hold irreversible steps to first-time-right. Where a step cannot be redone, AI assistance must pass rigorous first-time-right validation rather than iterate and fix.

Prompt from The AI Decision Course by Dr. Tony Hoang, host of The Artificial Intelligence Podcast (Spotify, Apple Podcasts): tonyphoang.com/learn

Your company's assistant may answer differently, so read what it writes before you use it.

Check yourself

Two proposals share one budget line: part-failure prediction for the line's most expensive machine, or a defect detector trained on historical data. Which gets funded first?

Show the answer

The part-failure prediction, if that machine's downtime costs more than the model. Run the test on real numbers.

The model passes acceptance the same week as a production push. Run both together?

Show the answer

No. Availability and throughput come first on the floor. Move the go-live so the interruption to production stays minimal.

The model has run clean for a quarter. Can drift monitoring move down the list?

Show the answer

No. Drift and performance monitoring runs continuously from the day the model goes live.

A vendor's AI step would scrap suspect product automatically, and the scrap cannot be undone. What comes first?

Show the answer

Evidence of rigorous first-time-right validation, before the step touches product.

A floor assistant demos with consumer-app accuracy. Good enough for the floor?

Show the answer

No. A system that misses too many failures, raises too many false alarms or invents facts is unacceptable there; the bar is far higher than for a consumer app.

Sources

The advice draws on interviews with an AI leader at a professional association, the chief executive of an enterprise technology services firm, the co-founders of a physical-AI startup, the chief product officer of an AI startup, the chief product officer of a large software company, and a vice president at a large semiconductor company.

  • Siemens (Senseye Predictive Maintenance), "The True Cost of Downtime 2024", published mid-2024; trade coverage dated 28 June 2024. (https://assets.new.siemens.com/siemens/assets/api/uuid:1b43afb5-2d07-47f7-9eb7-893fe7d0bc59/tcod-2024.pdf) The figures come from an ongoing Siemens survey of 181 maintenance, engineering and IT professionals at large industrial organizations in four sectors (automotive, fast-moving consumer goods, heavy industry, oil and gas), worldwide, covering April 2019 to March 2023. They are respondents' own estimates, Siemens sells predictive maintenance and the report estimates savings from it, and the report says its combined results "should be seen as indicative only".

Quiz

Join to take the quiz and see your score.

  1. Your quality team writes a detection rule for every new defect type. A vendor proposes a defect detector trained on your historical inspection data instead. What is the advantage?

  2. Your data science team wants to go live without any machine learning operations setup and let engineers manage the model by hand. What do you require?

  3. Months after your foundation model went live, maintenance moves a sensor on the line. What shows whether the move has hurt the model?

  4. An AI assistant drafts chip designs, and once a design goes to manufacturing there is no patching. What must you demand before using it there?

This is one lesson from The AI Decision Course.

$199 a year (founding price)

Blank card (PDF)The card as text, with a worked example

Membership adds every lesson and its card, the chat on your own case, your notes and progress saved, and a certificate per course.

Try it for 14 days. If it isn't useful, the full $199 comes back.

Paying with a company card? Ask your manager. Your receipt comes from Link, the checkout run by Stripe, and names the membership.

Made by Dr. Tony Hoang, host of The Artificial Intelligence Podcast.

Ask about this lesson

Where AI pays first on the factory floor, and the bar it has to clear

    Free preview: 3 questions a day, answered from this lesson. Follow-ups remember your earlier questions on this visit. Don't paste confidential details.

    Ask about this lesson