Written for the person who owns the P&L and the AI decision. Not a floor manual. A decision framework for the CEO, COO, or owner deciding whether to spend on AI at all, what to buy, and in what order, so the money lands on something that holds.
Read this first · Who this is for
This is written for you, the executive, not for your controls engineer. It is about the decision you own: whether AI is worth spending on this year, what a real vendor should be able to answer, and how to keep a pilot from becoming another tool nobody uses. The floor detail is here so it rings true, but the frame is the board, not the panel.
There are no invented market numbers on this page. The one figure with a number attached is a public rule with a link. Every ROI number in the full guide is one you put in from your own plant, because an industry average will not tell you what your gaps cost you.
The one thing to understand before you spend a dollar
AI is being sold into manufacturing as though the hard part is choosing the model. It is not. The hard part is that most of a high-production plant still runs on paper, one person's spreadsheet, and machines that talk to nothing. An agent, a copilot, or a forecast can only act on data a system can read. A batch sheet in a binder, a count clicked off an HMI and written down, a schedule rebuilt by hand every morning: none of that exists to software until a person re-keys it.
So the honest first question is not which AI to buy. It is how much of your floor an AI could even see today. If the answer is a third, you are about to pay for a tool that is confident about a third of your plant and blind to the rest. That is the single most expensive mistake in this category, and it is an executive mistake, not a technical one, because it is made at the moment of the purchase decision.
You cannot do AI on paper. The foundation comes first.
The plants that get value from AI did the same unglamorous thing in the same order first: digitize, connect, unify, before any AI. Digitize gets records off paper at the station. Connect gets machines and systems sharing what they already know. Unify pulls it into one live layer that is current and entered once. Only then does AI have something to stand on. Skipping that order is why AI purchases stall, and the bill for skipping it shows up as a failed rollout, not a line item you can see coming.
What the full guide covers
Below are the section headings of the full guide. The thesis above is free. The rest, the vendor questions, the traps, the pilot scope, the ROI math on your own numbers, and the questions to bring to your board, opens when you enter your work email.
Section 01
Should you spend on AI at all this year
The go or wait decision
How to tell whether your plant is ready to buy AI, ready to buy the foundation first, or better off waiting a quarter. The test you can run before any vendor call.
Section 02
What to ask a vendor before you sign
The questions that separate real from demo
The short list of questions a serious vendor answers in one sentence and a weak one deflects. Built to expose whether the tool needs a data layer you do not have yet.
Section 03
The traps that burn the budget
Where the money goes to die
Buying agents before the data layer exists, paying for a pilot that never touches production, and the giveaway hidden in numbers nobody trusts. The patterns behind most failed AI spend.
Section 04
How to scope a pilot that proves something
Small, real, on the clock
What a pilot has to include to tell you anything: a real line, a real trace, working software in weeks not quarters, and a definition of done you set before it starts.
Section 05
How to judge ROI on your own numbers
Your inputs, not an industry average
The four costs that are already on your P&L and how to price them yourself: re-keying hours, trace time, giveaway, and unplanned downtime. No benchmark required.
Section 06
The questions to bring to the board
What to say when they ask about AI
The handful of questions that turn an AI conversation with your board from hype into a capital decision, framed around risk, sequence, and what happens if you wait.
Want the same readiness question answered on your own floor before you read further? The AI Readiness Assessment returns your score and the phase you are actually in, free, and the AI Readiness Checklist is the plain yes-or-no version you work through by hand.
The full guide
Get the full buyer's guide.
You have the thesis and the six section headings. Enter your work email and the full guide opens right here on this page, and a copy goes to your inbox to forward to your team or bring to the board.
The go, wait, or foundation-first decision, with the test you run before any vendor call
The vendor questions, the traps, and how to scope a pilot that proves something
How to price ROI on your own numbers, plus the questions to bring to your board
Work email only. We use it to send the guide and nothing else you did not ask for. Unsubscribe anytime.
Unlocked. The full guide is open below, and a copy is on its way to your inbox. If you checked the box, a Harmony engineer will reach out to walk your readiness and a pilot scope with you.
Section 01
Should you spend on AI at all this year
The question is not whether AI will matter to manufacturing. It is whether your plant can turn a purchase into a result this year, or whether the money is better spent on the foundation that makes AI work later. There is a fast test you can run yourself, before any vendor call.
Pick one real lot and trace it end to end, on the clock. Receiving to production to quality to shipping, one pass, no rehearsal. If that trace means pulling binders and making phone calls and it takes hours, your records are not connected, and no AI you buy this year will change that, because the AI would be reading the same disconnected records. If the trace is minutes and lives in one system, you have a foundation worth putting AI on.
That single test sorts you into three answers. Go: your data is digitized, connected, and unified, so a scoped AI pilot can pay off now. Foundation first: the value is real but it starts with Phase 1, digitization, not with an agent. Wait: nothing on the floor is captured yet, and the honest move is to fund the capture, not the AI. Most high-production plants land in the middle, which is good news, because the middle is the most fixable position there is.
Section 02
What to ask a vendor before you sign
A serious vendor answers each of these in a sentence. A weak one changes the subject to the model. The point of the list is to surface, before you sign, whether the tool assumes a data layer you do not have.
What data does this actually read, and how much of my floor is captured in a form it can read today?
If my batch sheets are on paper and my machines are not connected, what does your tool do on day one, before any of that changes?
Do you digitize and connect the source records, or do you assume that work is already done?
Can my PLCs share data over a standard like OPC UA, and do you set that up, or is it on us?
Show me working software on my line in weeks. What exactly runs, and by when?
Who owns the data layer when the pilot ends, and can I keep it if I do not renew?
What is the one number on my P&L this changes, and how will we measure it together?
The pattern to listen for: a vendor who can only talk about the AI, and goes quiet on where the data comes from, is selling you the roof without the walls. The right partner starts with the foundation and treats the AI as what it is, the last layer, not the first.
Section 03
The traps that burn the budget
Trap 01 · Buying agents before the data layer exists
This is the expensive one. Agents and copilots are the last phase, not the first. Pointed at a plant that still runs on paper, an agent answers confidently from the sliver of data it can see and stays blind to the rest. You do not find out until the rollout stalls, and by then the money and the credibility are both spent. The fix is order: digitize and connect first, then put agents on top of a layer they can actually read.
Trap 02 · A pilot that never touches production
A demo on sample data proves nothing about your floor. If the pilot does not run on a real line with real records and a real trace, it is theater. A pilot that cannot show working software on your own line in weeks is telling you it is not ready to. Insist the pilot touch production, or do not call it a pilot.
Trap 03 · The giveaway hidden in numbers nobody trusts
When the same figure lives in three places and drifts apart, people stop trusting any of it and pad every decision to be safe. That padding is giveaway: extra material, extra labor, extra buffer, quietly on your P&L. AI does not fix this. A single trusted live layer does, and that is a foundation cost, not a model cost. If a vendor promises the AI will fix your numbers, they have the order backward.
Trap 04 · Confusing a model problem with a data problem
When an AI purchase underdelivers, the instinct is to blame the model and shop for a better one. In manufacturing the cause is almost always missing or disconnected data, not a bad model. Buying a second tool onto the same broken foundation spends more to get the same result. The failure is upstream of the AI, so the fix is upstream of the AI.
Section 04
How to scope a pilot that proves something
A good pilot is small, real, and on the clock. It runs on one line you already care about, it touches production, and it has a definition of done you set before it starts, not one the vendor sets after. The two tests it has to pass: does working software run on your floor in weeks, and does a real lot trace get faster because of it.
For reference, here is what Harmony offers, stated as a first-party number so you have a concrete shape to compare any vendor against, not an industry benchmark:
$15,000–$20,000
One-time pilot cost
4–6 weeks
Full pilot window
By the end of the pilot
Working software running
The shape matters more than the exact figures. A pilot priced as a one-time cost, measured in weeks, with working software before the halfway mark, is one designed to prove something and then hand you a decision. A pilot priced as an open-ended engagement with value promised at the end is one designed to keep billing. When you scope yours, write the definition of done first: the line, the trace, the number on the P&L, and the date. The first week of a real pilot is the foundation walk, digitizing at the station and connecting what your machines already know, because that is the order it has to happen in.
Section 05
How to judge ROI on your own numbers
You do not need a benchmark to build the business case, and you should distrust any vendor who hands you one. The four costs below are already on your P&L. Price them with your own inputs and you have an ROI case no industry average can argue with.
The four costs already on your P&L
Re-keying hours. Count the hours each week your team spends typing figures that already exist somewhere else, multiply by loaded labor cost, annualize. This is the most legible number on the list because it converts straight into a budget line.
Trace time. Time one real lot trace, then weigh it against what a trace is now expected to take under the rule below. The gap is your exposure in an audit or a recall.
Giveaway. The extra material and labor you spend padding against numbers you do not trust. Estimate the buffer as a share of material cost and annualize it. Trusted numbers shrink the buffer.
Unplanned downtime. Put your own cost on an hour of unplanned downtime and multiply by the hours you lose to failures you did not see coming. Live machine data is what turns that reactive.
Add those four on your own numbers and you have the cost of the gap, in dollars, at your scale. Then weigh it against a pilot priced as a one-time cost. The case for the foundation is almost always made by numbers you already have, not by a benchmark you have to trust.
On trace time specifically, what a trace is expected to take is not a matter of opinion. Under the FDA Food Traceability Rule, which implements section 204 of the Food Safety Modernization Act, covered firms must make required traceability records available to an authorized FDA representative within 24 hours of a request, and during an outbreak, recall, or other public health threat must provide the required information in an electronic sortable spreadsheet within that same window. A paper trace can meet neither condition reliably, which is why trace time is both an ROI number and a risk number.
Source: 21 CFR 1.1455, paragraphs (c)(1) and (c)(3)(ii), FDA Food Traceability Rule.
Want these priced for you? The ROI Calculators & Tools put dollar figures on the gaps using only your own inputs.
Section 06
The questions to bring to the board
When your board asks what you are doing about AI, these are the questions that turn the conversation from hype into a capital decision. Bring them and you frame the discussion around risk, sequence, and cost, which is where the real decision lives.
How much of our floor could any AI actually read today, and what is the honest number?
What does a paper-based trace cost us in an audit or a recall, and what is that exposure worth?
Are we buying AI, or are we buying the data foundation that AI needs, and do we know which one we need first?
What is the one-time cost to lay the foundation, and how does it compare to the four costs already on our P&L?
If a competitor lays this foundation and we do not, what do they get that we cannot answer in twelve months?
What is the smallest pilot that would prove or kill this, and what is our definition of done before it starts?
The through line for the board is order and risk. The plants that win with AI are not the ones that bought the flashiest model. They are the ones that laid the data foundation first, in the right sequence, and put AI on top of something that holds. That is a capital decision you can defend, and it is the one this guide is built to help you make.
The sequence behind the whole decision
Everything above maps onto the same three phases every plant moves through. The order is not a preference, it is the constraint: each phase is what makes the next one possible. Buying out of order is the trap the whole guide is built to help you avoid.
Phase 1
Lay the Data Foundation · Digitization
Every pen-and-paper record digitized at the station, every software system connected, and all of the data unified into one live layer. The digital transformation starts here.
Phase 2
Production & Operations Scale
Factory operations turn proactive: live sensors and machine data, the AI scheduling board, predictive maintenance before failure.
Phase 3
AI-Native Operations
Agents across the floor and the back office act on the live layer: quality signals, reports, copilots. Humans approve.
The first week of a Harmony pilot is the foundation walk, on-site, with forward-deployed engineers doing the counting alongside your team. Phase 1 first, because that is the order it has to happen in. See what the live layer looks like.
