The 70-point gap: manufacturers have an AI strategy, and it is nowhere near the floor

The AI is already in the building, in the wrong room

The Census Bureau rewrote its AI question in November 2025, from AI used in producing goods or services to AI used in any business function. The Federal Reserve’s Jeffrey S. Allen documented the effect: manufacturing’s reported adoption jumped 159%, roughly from 4.7% to 12.2%, the largest proportional rise of any sector.

Same firms. Same fortnight. Wider question. Four times more AI.

Manufacturing’s AI was always outside production, which is why the narrow question kept missing it. The same rewritten question exposed it across every physical sector, and manufacturing showed the effect most sharply.

Census’s own working paper says where it lives. Among firms using AI at all, the functions are sales and marketing at 52%, strategy and business development at 45%, IT at 41%. And 57% run it in three or fewer functions.

Your competitors are in the same position. The one who moves it to the floor first gets a cost structure the others cannot match on price.

Waiting is the expensive option

The comfortable assumption is that a strategy sitting on paper is neutral, and the cost of waiting is zero. Census microdata says otherwise.

Kristina McElheran of the University of Toronto and the MIT Initiative on the Digital Economy, working with Erik Brynjolfsson and colleagues, studied what happened inside manufacturing firms after they adopted AI. Established firms did something nobody predicted. They let go of the management practices that were already working.

“Old firms actually saw declines in the use of structured management practices after adopting AI,” McElheran told MIT Sloan. The paper attributes roughly one third of the short-run productivity losses to exactly that: firms abandoning structured production-management practices such as monitoring key performance indicators and production targets.

Her summary is the sentence to keep:

An unembedded strategy arrives on the floor as a second system running alongside the first, and the first is what quietly gets dropped. That is a live cost, accruing now, in plants that believe they are simply being patient.

The valve at the frontline

PwC and the Manufacturing Institute surveyed 102 HR and operations leaders in Q3 2025, specifically about frontline leadership. Frontline leaders here means the salaried people running production: supervisors, line leaders, manufacturing team leaders.

The integration numbers are single digits. AI is meaningfully integrated into daily operations for 8% of frontline workers and 2% of frontline leaders.

Against a 90% strategy rate, the 2% is the number that should stop a plant manager. The people running production are the least likely group in the entire company to be working with the thing the company says it has a strategy for.

Three findings explain why it holds.

Nineteen percent offer any AI-related training at all, and 49% say their existing programs prepare frontline employees for AI-driven role change in no way whatsoever. Forty-five percent say the exclusion of frontline leaders from design and rollout contributed to unsuccessful AI initiatives — PwC’s framing of the mechanism is exact: frontline leaders “are often engaged downstream after key technology decisions have been made, limiting their ability to shape solutions or surface operational constraints before deployment.” And the one that closes the loop: 3% report that most frontline AI suggestions get implemented.

That is a valve. Decisions travel down after they are made, and almost nothing travels up. Which makes it entirely rational that 62% of frontline workers are described as skeptical of AI against 24% described as excited. They have watched this movie five times.

“As AI becomes more important to manufacturing, leaders need to offer additional training so that workers are equipped to utilize AI on the shop floor.”

— Gardner Carrick, Chief Program Officer, The Manufacturing Institute

Four full-time people, writing down work that was already done

Here is what the gap costs in headcount.

OFS surveyed 500 manufacturing professionals in June 2026 and found operators losing an average of 21 minutes a shift to manually logging information, transcribing data and preparing reports. Its CEO James Magee made the obvious point: “21 minutes does not sound dramatic until you scale it.”

So scale it. A hundred people on shift, 250 working days: 8,750 hours a year. Four full-time equivalents, occupied entirely with writing down work that had already been done. In a plant of 300, it is closer to twelve.

That is before quality. Quality is where the documentation load stops being minutes and starts being days.

Hold the four-people figure for a moment, because the number that matters next is what it costs to check it. A bounded first step on quality and production starts at $15,000, priced once the scope is clear, and produces a measured answer in six weeks. Four full-time equivalents cost a plant more than that every month.

A part is inspected, then the non-conformance is written up. A corrective action is agreed, then the 8D report is built. A CAPA is opened, then the investigation is documented into a QMS designed to satisfy an auditor rather than to help the engineer filling it in. Every one of those is a senior person, doing a second job, describing a first job they have already finished.

In regulated manufacturing, where a federal inspector keeps score, you can see what that does. The single most-cited device finding in the FDA’s fiscal year 2025 inspection observations was inadequate or non-existent CAPA procedures. On the drug side, procedures not in writing or not followed by the quality unit ranked first. The most common thing a regulator finds in a plant is a quality process that failed to get properly written down.

Everywhere else, nobody has counted it, because nobody sells the counting. That is the opportunity, not a hole in it. Quality is the function where the most time disappears into writing and the least of it gets measured, which also makes it the easiest place in the plant to prove a change. The output is a document with a date on it. You can show a before-and-after in six weeks without instrumenting a single machine.

That is why what AI-native manufacturing actually looks like starts there rather than with a company-wide rollout.

The layer after that is planning

Sales, inventory and operations planning is the second place the gap closes, and it closes differently. Quality is a documentation problem solved one department at a time. SIOP is a reconciliation problem rebuilt end to end, because the cost is spread across ERP, supplier EDI feeds and a set of spreadsheets that exactly one person fully understands.

Every published benchmark for how long a mid-market SIOP cycle takes comes from a company selling planning software, which is the argument for rebuilding the planning cycle end to end as an embedded engagement instead of a tool purchase. The only way to find out what a cycle is made of is to work inside one, with the person whose spreadsheet it is.

For a company running several plants on different ERP instances, there is a third layer above both, where the same question gets asked once and answered the same way everywhere. That is a longer engagement and a later one. Most manufacturers should start below it.

What “embedded” actually means

PwC’s survey carries one more number worth holding. Only 4% of respondents reported comprehensive success across all the AI and digital maturity indicators measured, while 89% said their technology investments have failed to fully deliver expected results and 87% said poor data quality hampered the value of digital initiatives.

So “embedded” is a working state rather than a maturity badge. The table below is our own framework, drawn from what the two states look like in the engagements we run.

Five questions worth asking before anyone calls a strategy embedded:

1. Which department can show a measured before-and-after, and on what data?
2. What did a frontline supervisor change about how they work this month?
3. Which manual step was retired, rather than kept running in parallel?
4. When a line leader made a suggestion, what happened to it?
5. If the vendor changed tomorrow, what would the company keep?

A company that can answer all five is in the 20%. A company that can answer the first two is closer than most.

Where to start, and what it costs

Manufacturing’s AI is already in the building. It is in sales, in marketing, in IT, and the plant has been on the far side of it since before anyone measured properly. That is true across every industry that runs the physical world, and manufacturing shows it most sharply, because manufacturing is where the distance between the boardroom and the work is longest.

Closing it is a sequencing problem rather than a budget one. It runs in three moves, and only the first one needs deciding now.

One department, measured. A bounded first step, one department, from $15,000, in the function where the writing is heaviest and the proof is fastest.

Concretely: a week timing how your 8Ds, non-conformances and CAPA records actually get produced now, two weeks configuring agents against your own QMS and your own past reports, three weeks of quality engineers using it on live inspections, and a written before-and-after at the end. Your QMS stays where it is. The decision about going further gets made on the comparison rather than on a demonstration.

One workflow, rebuilt. An embedded pod takes a single high-value workflow end to end, against live production data, with a go or no-go gate before anything scales.

Then the layer across plants, once the first two have earned it.

Start where the writing is heaviest. In almost every plant, that is quality.