The sequence a plant follows to get off paper and become ready for AI, written for the person who owns the P&L and the spend. Two arcs, three phases, in the one order that works. Do them out of order and you pay for AI that has nothing to read.
Read this first · Who this is for
This is a decision brief, not a floor manual. If you are the owner, CEO, or COO weighing whether to spend on manufacturing AI this year, the honest first question is not which platform to buy. It is whether your plant holds data an AI could read at all. Most high-production floors still run on handwriting, one person's spreadsheet, and machines that talk to nothing, and no model can act on a record that only exists on paper.
There are no invented numbers on this page. The one figure here is a public rule with a link. The pilot terms further down are Harmony's own offer, stated plainly. Everything else is sequence, because the order you do the work in is the part that decides whether the money returns.
The order is the decision
Manufacturing AI is sold as though the data already exists: records in systems, machines reporting numbers, figures that are current. On most floors that is not the case, and the gap is the whole problem. An agent that cannot read the record cannot act on the record. A copilot answering from a system that holds a third of the week gives confident answers about a third of the plant. That is not a model failure. It is a foundation that was never laid.
So the roadmap below is not a menu you pick from. It is an order. The plants that got a return from AI did the same unglamorous work in the same sequence first, and the ones that wrote off AI almost always skipped it. For an executive the takeaway is blunt: the sequence, not the model, is what determines whether the spend comes back.
What is actually on the line
Three things move at the board level while the floor stays on paper. Audit and recall exposure: when records live in binders, a regulator's clock and a recall's blast radius are set by how fast you can trace a lot, and paper cannot meet a modern trace window. 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 a recall or public health threat must provide them in an electronic sortable spreadsheet within that same window (21 CFR 1.1455). Cost that compounds: every number re-keyed by hand, every figure that is only current at month end, every giveaway you cannot see until after the run is a cost that repeats every shift. Competitive position: the plants laying this foundation now are the ones that will be able to buy AI that works, while the plants still on paper will keep buying AI that does not stick.
The two arcs
The whole journey splits into two arcs. The order between them is not negotiable: the first has to be substantially done before the second returns anything.
Arc 1
The digital transformation
Get every record off paper, get the machines and systems sharing what they already know, and unify all of it into one live layer that is current and entered once. This is the foundation. It is where most of a high-production plant actually is today, and it is the least glamorous and most valuable work on the list.
Foundation · records a system can read
Arc 2
AI
Once the live layer exists, AI has something to stand on: proactive operations, scheduling and predictive maintenance, then agents and copilots that act on the live data with humans approving. Buying this arc before Arc 1 is where the failed AI purchases come from.
Payoff · agents that act on the layer
The three phases
The two arcs unfold across three phases. Arc 1 is Phase 1. Arc 2 is Phases 2 and 3. You cannot buy your way past Phase 1, and every plant moves through them in this order whatever it has been sold.
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.
Phase 2
Production & Operations Scale
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 back office act on the live layer: quality signals, reports, copilots. Humans approve.
Why the order matters
Each phase is the ground the next one stands on. Phase 2 scheduling and predictive maintenance need a live signal off the machines, which Phase 1 creates. Phase 3 agents need one current, readable layer to act on, which Phase 1 unifies. Skip ahead and you are pointing sophisticated tools at data that is partial, stale, or trapped in a panel, and they behave exactly as badly as that input. The reason failed AI purchases trace back to missing data rather than a bad model is that most plants tried to start at Phase 3. The good news for the P&L: Phase 1 is the most fixable position on the list, because the work is known and mechanical rather than experimental.
Not sure which phase your plant is actually in? The AI Readiness Assessment is the interactive version: twelve questions, and it returns your score, the phase you are in, and your next three actions on screen, free.
The full roadmap
Get the step-by-step inside Phase 1.
You have the two arcs, the three phases, and why the order matters. Enter your work email and the executable part opens right here on this page, with a copy to your inbox: the exact sequence inside Phase 1, and the sequencing mistakes that sink AI budgets.
Which record to digitize first, and why it is not the easy one
Connecting machines over OPC UA, and unifying into one live layer
The six sequencing mistakes executives make, and the fix for each
Harmony's fixed-scope Phase 1 pilot terms, stated plainly
Work email only. We use it to send the roadmap and nothing else you did not ask for. Unsubscribe anytime.
Unlocked. The full roadmap 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 map the sequence to your plant.
Phase 1 · Step 1
Digitize the record that carries your worst liability first
The instinct is to digitize the easiest record, or to try to digitize everything at once. Both are wrong. The record to move off paper first is the one whose absence costs you most in an audit or a recall. On most high-production floors that is the batch or production record and the quality checks a lot trace has to cross, because those are what a regulator asks for and what a recall depends on.
Digitizing means capturing it at the station, where the work happens, so it becomes data the moment the operator records it rather than hours later when someone re-types a clipboard. A record written by hand does not exist to any software until it is re-keyed, and the re-keying is both a cost and a place errors enter. Start where the exposure is highest, prove the pattern on one line, then extend it. You are not buying a system here. You are removing the paper between the floor and the record.
The executive test
Ask which single paper record, if it went missing tomorrow, would hurt you most in an audit or a recall. Digitize that one first. If the answer is a binder, you have found Phase 1.
Phase 1 · Step 2
Connect the machines over OPC UA
Your machines already know their numbers. Counts, states, faults and cycle times exist inside the controller and, on most floors, stop at the panel: an operator reads them off an HMI screen or a clicker and writes them down. The gap is not measurement. It is that the measurement never leaves the machine.
Many controllers can already share that data over OPC UA, the standard connection for industrial equipment, so the common case is not a plant that needs new sensors on everything. It is a plant whose existing signal is trapped behind the panel. Connecting it gives you a live feed instead of a clipboard, and that live feed is the thing Phase 2 scheduling and predictive maintenance later run on. The executive framing: this is usually a connection problem, not a capital-equipment problem, which is why it is cheaper and faster than most vendors imply.
The executive test
If someone is writing machine counts on paper, the machine can almost certainly hand them to a system directly. Ask your team whether your PLCs can speak OPC UA before anyone quotes you new hardware.
Phase 1 · Step 3
Unify everything into one live layer
Digitized records and a machine signal are not yet a foundation if they sit in separate places. The last step of Phase 1 is to pull them into one live layer that is current and entered once. One place where the number is entered a single time, stays current, and can be read by a system today and by an AI agent later.
This is what kills the two costs that compound quietly: the hours a week your team spends re-typing figures that already exist somewhere else, and the drift that happens when the same number lives in three systems and slowly disagrees until nobody trusts any of it. When the layer is unified, a lot trace stops being phone calls and binders and becomes a query, which is what turns the audit and recall exposure from Step 1 into a solved problem. This is the deliverable of Arc 1. Everything in Arc 2 stands on it.
The executive test
Ask how many places the same number lives, and how many hours a week your people spend moving numbers between systems. Both should trend to near zero when the layer is unified.
The pitfalls
Six sequencing mistakes executives make
The roadmap is simple. The way it goes wrong is predictable, and almost always a sequencing error made above the floor, not on it. These are the six that sink budgets.
01Buying Arc 2 before Arc 1 exists. Signing for a copilot or agents while the plant is still on paper. The tool then reads a fraction of the operation and the pilot is judged a failure. Fix: fund Phase 1 first, treat AI as the second purchase, not the first.
02Digitizing the easy record instead of the risky one. Starting where it is convenient rather than where the audit and recall exposure lives. Fix: sequence by liability, per Step 1.
03Boiling the ocean. Trying to digitize every record and connect every machine at once, which stalls under its own weight. Fix: prove the pattern on one line, then extend.
04Treating it as an IT project. Handing Phase 1 to a systems team detached from the floor produces software nobody on the line will use. Fix: run it as an operations project with engineers on the floor.
05Skipping unify and buying dashboards. Pointing a reporting tool at disconnected sources gives confident charts built on stale, partial data. Fix: unify the layer first, then visualize.
06Letting one spreadsheet become the system. The daily report or schedule that one person rebuilds by hand is a single point of failure and a source no AI can read. Fix: make the live layer the source, not a person's file.
Regulatory source referenced above: 21 CFR 1.1455, paragraphs (c)(1) and (c)(3)(ii), FDA Food Traceability Rule.
Harmony's offer
Phase 1 as a fixed-scope pilot
$15,000 to $20,000 one time4 to 6 weeksWorking software by the end of the pilot
Harmony runs Phase 1 as a fixed-scope pilot: forward-deployed engineers on-site digitizing the record that carries your liability, connecting the machines over OPC UA, and unifying it into one live layer. The terms are $15,000 to $20,000 one time, 4 to 6 weeks, with working software in your hands by the end of the pilot. It is deliberately small and bounded so the decision is easy to make and the foundation is real before anyone talks about Arc 2.
Want the sequence mapped to your own plant before you commit? Run the AI Readiness Assessment for a score and your next three actions, or book a call below.
Rather have someone map the sequence to your plant?
The roadmap above is the first weeks of a Harmony pilot: on-site, forward-deployed engineers doing the digitizing, connecting and unifying alongside your team. Phase 1 first, because that is the order it has to happen in. See what the live layer looks like, or run the AI Readiness Assessment first.
