AI in Manufacturing Without Stopping the Line: A Brownfield Playbook

Last verified: October 27, 2026. We re-check every figure in this playbook against its primary source each quarter, starting with the US Census Bureau data that opens it.

TL;DR

• In the US Census Bureau’s business survey for October 5 to 18, 2026, 22.9% of manufacturing firms used AI, below the 24.6% rate for all US firms.
• Deloitte’s 2026 survey of more than 140 manufacturers found that 84% report measurable value from AI, while only 20% of their use cases run at scale across sites.
• In most plants, the hard part is attachment: connecting a model to equipment of several generations on a line that cannot stop for an experiment.
• Build the data foundation and a read-only attachment first, then let each use case climb the trust ladder one tier at a time on measured evidence.
• AI never gets control of safety functions. Emergency stops, interlocks, safety instrumented systems and other protective functions stay outside its reach by design, as functional-safety practice requires.

This playbook on AI in manufacturing is for operations and engineering leaders who run brownfield plants. A brownfield plant is an existing facility where new lines work beside equipment installed decades earlier. The playbook shows where an AI layer attaches to such a plant and what that layer is allowed to change, including the functions it never touches. Use cases such as predictive maintenance appear here only as examples of attachment, because the architecture decides whether any of them reach production.

What government data says about AI adoption in manufacturing

In the US Census Bureau’s survey round for October 5 to 18, 2026, 22.9% of US manufacturing firms used AI in at least one business function, below 24.6% for all US firms. The survey is the Business Trends and Outlook Survey (BTOS), a nationally representative survey of US firms that runs every two weeks. Manufacturing trails information at 46.4% and finance and insurance at 40.5%. More than three in four US manufacturers did not report AI use in any business function, while much of the vendor literature describes an industry already transformed.

On November 17, 2025, the Census Bureau changed its core AI question from use “in producing goods or services” to use “in any of its business functions.” Census Bureau researchers write that the revision precludes a direct comparison of earlier and later rates. A Federal Reserve analysis found that manufacturing had the largest relative jump between the two wordings among the industries it examined: 159%, or 7.5 percentage points. Any growth figure that spans both versions therefore overstates the climb. On the new wording alone, manufacturing rose from 12.7% in late November 2025 to 22.9% in mid-October 2026.

The second detail is firm size, since adoption rises with it across the economy. In its May 2026 analysis, the Census Bureau reported AI use at 37% of firms with 250 or more employees. The industry figure therefore understates use among the larger manufacturers this playbook addresses.

A second survey reframes the first: Deloitte’s AI in Manufacturing 2026 study covered more than 140 manufacturing organizations. It found that 84% already generate measurable value from AI, yet only 20% of their use cases are scaled across sites or enterprise-wide. The two surveys count different populations, since Census samples every US firm, including the smallest, while Deloitte surveyed organizations already working with AI. Read together, they show AI paying off where plants have attached it, in a small share of plants and use cases. Deloitte names execution and industrialization as the key challenge.

The limiting factor in both surveys is attachment, and four conditions make a plant harder to attach to than an office. The line cannot stop, and the equipment spans several generations. Outputs move physical things, and the operational technology (OT) network has an owner who guards it. The rest of this playbook turns those conditions into an attachment map and a trust ladder, with a firm line around what never gets automated. It draws on how SumatoSoft scopes IoT for manufacturing engagements, where the AI layer goes on top of the systems a plant already runs.

Why plants are different: four constraints that change the architecture

Each of the four constraints rules out a design that works in office software. Together, they decide where an AI layer can connect.

The line cannot stop. A plant has no staging environment that matches production, and every operations leader carries a downtime number personally. Any architecture whose failure mode is a stopped line fails before anyone evaluates its accuracy. Model training and testing therefore happen away from the control system, on copies of plant data. A model that crashes has to leave the line running.

The equipment spans several generations. SumatoSoft’s manufacturing practice starts from a plain observation on its service page: “Most manufacturing environments are mixed.” In a typical brownfield plant, a 2021 line runs next to a 1998 press. The historian, the database that records time-stamped machine readings, holds tags (named data points) that its current engineers cannot fully explain. Data availability varies per asset as well. A new machine exposes hundreds of tags over OPC UA, an industrial data standard. An older press exposes a single run signal, and a standalone pump exposes nothing until someone fits a sensor. The architecture has to tolerate that unevenness. SumatoSoft’s IoT development work therefore normalizes data from every asset before any model uses it.

Outputs move physical things. Software errors in an office end up as records that someone corrects later, which gives every office system a cheap path back. On a plant floor, a wrong write can move a tool, change a temperature, release a batch or stop a line. That difference is why the trust ladder later in this playbook has a tier that never opens.

The OT network has an owner. OT is the network of controllers and machines that runs production, and plants divide it into security zones on purpose. The ISA/IEC 62443 standards for industrial automation security organize protection around those zones and the conduits between them. NIST’s Guide to Operational Technology Security (SP 800-82 Rev. 3) gives the US reference. The plant’s OT team decides what crosses each boundary and refuses AI vendors direct access to the control layer, and it is right to. Every IT/OT convergence project, AI included, crosses that boundary under the OT team’s rules. An AI layer therefore works through conduits the OT team already approves. Those include read-only interfaces, mirrored network ports, a historian copy outside the control zone, and data diodes. Where the architecture calls for one, a data diode lets data travel in only one direction, out of the control zone, so nothing comes back in. SumatoSoft holds ISO 27001 certification and aligns its plant security approach with IEC 62443.

Together, the four constraints lead to one design decision, which the next section maps: the AI layer works outside the control loop.

The non-invasive attachment map: four places AI connects to a plant

The non-invasive attachment map shows the four places an AI layer connects to a plant: telemetry and the historian, the quality station, planning and scheduling, and field service.

SumatoSoft’s IoT for manufacturing page states the principle in one sentence: “We add AI on top of your existing IoT system without touching production-critical components.” The same page adds that “the AI layer operates independently, consuming data through secure connectors and APIs.” Existing devices, gateways, data pipelines, and dashboards remain intact. In architecture terms, the AI layer attaches beside the control system and stays outside the control loop.

“Beside” needs a definition that an OT engineer can test. The AI layer reads plant data through read-only interfaces, so it has no way to write to a controller. Some use cases later earn the right to change something, such as a set-point, the target value a controller holds, or a batch hold. The change then enters as a request through an interface the plant already exposes to operators. The controller’s own limits still apply to that request, and the safety system stays independent of both.

SumatoSoft applied this pattern for a multi-plant industrial machinery components manufacturer in Poland. The team added an explainable machine learning layer to the client’s existing IoT platform without stopping production. An 8-week pilot on selected assets came first, and the rollout then moved line by line. The results appear under the Advise tier below.

Each layer below answers the same four questions: what data it needs, what it may touch, which tier it starts in, and what breaks when it is wrong.

Layer 1: Telemetry and the historian

Most plants start here, because the data already exists in sensor streams, historian tags, SCADA feeds, and maintenance records. SCADA (supervisory control and data acquisition) is the software operators use to watch and steer the process. The difficulty lies in meaning more than in volume. Tag names follow a convention set three retirements ago, units differ per line, clocks drift apart, and gaps go unlogged. A pilot that never gets a steady feed from the historian and the maintenance database stalls, however good its model. For that reason, custom machine learning development for plants starts with data plumbing. This layer may touch nothing, because it is read-only by construction, and it starts in the Observe tier. When it goes wrong, the model learns an artifact of the historian instead of the process. Typical artifacts are a compression setting that dropped data points and a tag renamed during a controls upgrade.

Layer 2: The quality station

Vision inspection and in-process measurement attach at a point where a decision already exists and a person already makes it. Every accept or reject call becomes a label, so the ground truth is unusually clean and the feedback loop is short. The layer may touch a reject gate or a hold flag, and either one already counts as an Act-tier write. The trust ladder therefore applies here earlier than anywhere else. The layer starts in Observe, scoring parts in parallel with the inspector without acting on them. When it goes wrong in one direction, a false accept sends a defect to a customer. In the other direction, a false reject costs yield and erodes operator trust fast.

Layer 3: Planning and scheduling

Planning data lives in IT systems: the ERP (enterprise resource planning) system, the MES (manufacturing execution system), routings and the labor calendar. That location makes attachment easier and change management harder, because the system competes with a planner’s judgment. Deloitte’s 2025 Smart Manufacturing and Operations Survey shows the demand, with 35% of the executives surveyed at large manufacturers with US operations ranking advanced production scheduling as a top-two investment priority. The layer may touch a proposed schedule that a planner accepts or overrides, and it starts in Advise. When it goes wrong, the plan looks optimal on paper but ignores a constraint the planner carries in their head. One example is the single setter qualified on a particular die. Once plans miss constraints like that, planners start routing around the system.

Layer 4: Field service and the feedback loop

Field service is where alerts turn into work. SumatoSoft’s AIoT development services page describes the loop in one line. It reads: “AI alerts create work orders with the right asset IDs and parts, completions feed back to retrain AI models and refine thresholds.” A computerized maintenance management system (CMMS) holds those work orders. Each completed job then becomes a labeled example that keeps the model current as equipment wears. The layer may touch the work-order queue, and it starts in Advise, with drafted work orders that a planner releases. When it goes wrong, the failure is alert fatigue. Technicians stop opening alerts once too many prove false, and a muted system stops improving because the loop no longer closes.

The four attachment layers at a glance

The plant-floor trust ladder: Observe, Advise, Act and Never

On a plant floor, a write permission decides whether software can move equipment. Each tier of the ladder is therefore granted per use case and per asset class, with evidence, and never to a system as a whole. Our playbook for governing agent write access in enterprise IT uses a similar ladder. The plant-floor version differs in one respect: its top tier never opens.

• Observe: the model watches production and writes nothing.
• Advise: the model recommends, and a person decides.
• Act: the model makes bounded, reversible changes inside a fence the OT team defines.
• Never: safety functions stay outside the automation envelope, whatever the model’s accuracy.

Every use case starts in Observe and earns Advise and then Act on measured evidence, while safety functions stay in Never permanently.

Observe: The model watches and writes nothing

In Observe, the model runs read-only in parallel with production. It is scored against outcomes that happen anyway: failures, rejects, unplanned stops and late orders. The model stays here until its false-positive rate is a measured figure from production data. Model performance starts to drift once it meets live data, and in a plant the drift is physical. Vibration baselines move as bearings wear, tooling changes, the product mix shifts and ambient temperature follows the seasons. A model validated during a demo week has seen none of those shifts. Promotion to Advise therefore requires a measured error profile across a full production cycle, including at least one changeover. Our guide to predictive maintenance in practice shows how anomaly detection covers machines that have no failure history yet.

Advise: The model talks to a person

In Advise, the model issues alerts, recommended set-points, ranked work orders and flagged parts, and a person makes each decision. Its main design problem is alert economics. Operators stop reading alerts from a noisy advisory system, and once they stop, its accuracy no longer changes anything on the floor.

The Polish manufacturer’s results show which metric to watch. False alerts fell by 40%, and unplanned downtime fell by 50%. Early fault detection improved by 35%, and the project delivered a 2.5x return on investment within 8 months. For a maintenance manager, a falling false-alert count carries more weight than any accuracy figure. Every false alert costs a technician’s walk to the machine. Promotion to Act requires operators to act on the advice at a measured rate. That condition ties to SumatoSoft’s AI readiness research, which surveyed 72 business and technology leaders. Of those, 61% named rebuilding processes around AI as the top factor that moved them from pilot to production.

Act: The model changes something inside a fence

In Act, the model writes, but only bounded and reversible changes. Examples include a set-point inside a validated band, a hold on a suspect batch, a reject at the quality gate, and a work order created automatically. Each write is allowlisted per asset, with hard caps, a rate limit, a tested stop and a logged reason. Writes travel through the supervisory interface the plant already uses, so the controller’s own limits still apply.

OT managers look first at how a promoted model gets demoted. SumatoSoft’s AIoT practice uses “signed and staged OTA for firmware and AI models, as well as health checks and rollback.” OTA means over-the-air updates. A promoted model can therefore return to its previous version without a site visit. Promotion to Act requires a monitored Advise period, measured agreement between model and operator, a tested rollback and OT sign-off. This tier is also where two of the seven enterprise AI adoption risks turn physical: the expanding attack surface and ungoverned agentic autonomy.

Never: The tier that does not open

AI should never control safety instrumented systems, interlocks, emergency stops or the other protective functions that make a plant fail safe. A safety instrumented system brings a process to a safe state when a hazard appears. An interlock stops a machine when, for example, a guard opens. These functions remain outside the automation boundary by design, under engineering practice that predates AI. Functional-safety standards govern how they are engineered and validated. Examples are ISO 13849-1 for the safety-related parts of machine control systems and IEC 61511 for safety instrumented systems in process plants. Connecting a model to a validated safety function would change that function. The change would send it back through validation with the model inside the safety case, so the AI layer gets no path to it. The AI layer can still read trip events once the historian has recorded them. When a proposed use case needs a path into a safety function, SumatoSoft says no during scoping.

Sequence the work so the cheapest step comes first

Survey data supports building the data foundation before the AI. Deloitte’s 2025 Smart Manufacturing and Operations Survey covered 600 executives at manufacturers with US$500 million or more in revenue and headquarters or operations in the US. At the facility or network level, 57% used cloud computing and 57% used data analytics. Industrial IoT stood at 46%, while AI and machine learning stood at 29%. The survey is a snapshot fielded in August and September 2024, so it records where large manufacturers stood at one moment. The case for that order rests on dependency, because a model needs clean, connected data before it needs anything else.

The attachment sequence runs in five steps, and each step produces the evidence that justifies the next:

1. Map what data exists and what each tag means.
2. Attach read-only and observe.
3. Advise a person.
4. Act inside a fence.
5. Extend line by line, then plant by plant.

Each step costs less than the one after it. A plant that stops after step 2 has spent little and has learned what its data supports.

The sequence also exposes use cases with no data behind them. If the asset exposes no data and the failure mode has no labels, the first project is instrumentation, and the AI work starts after it. Running that project as an AI pilot means paying to discover the gap in week three. An AI readiness assessment finds the gap before the pilot starts.

The wider evidence points the same way. In SumatoSoft’s AI readiness research, 58% of respondents named data quality and consistency as their biggest readiness gap. In a plant, that gap lives in the historian and the maintenance log. Deloitte’s State of AI in the Enterprise survey of 3,235 business and IT leaders adds the scale problem. Only 25% of them had moved 40% or more of their AI pilots into production. Our analysis of why AI pilots never reach production covers the causes outside manufacturing.

Frequently asked questions about AI in manufacturing

What percentage of manufacturers use AI?

According to the US Census Bureau, 22.9% of US manufacturing firms used AI in at least one business function in the survey collected from October 5 to 18, 2026. The rate for all US firms was 24.6%. Census changed the question wording in November 2025, so earlier figures do not compare with these.

Can you add AI to old manufacturing equipment without replacing it?

Yes, an AI layer can read legacy equipment through the least invasive route each machine allows. Routes include a read-only OPC UA connection, a protocol gateway that listens passively, a current clamp on a motor feed, and a retrofit sensor. Edge software normalizes those signals into one data model. The AI layer then works from that model without touching the controllers.

Does AI need to connect to the PLC or control system?

No, an AI layer needs no connection to the control loop. It reads data from a copy of the historian, from read-only interfaces, from mirrored network traffic, or from added sensors. The PLC (programmable logic controller) keeps running its own logic. Any change a use case later earns enters through an interface the plant already exposes to operators.

Can AI control a production line?

AI can make bounded changes once it has earned that right, but it should never take over a line’s control logic. On a plant floor, a wrong write can move a tool, change a temperature, release a batch, or stop a line. Changes the model is allowed to make therefore stay allowlisted per asset, with hard caps, rate limits, a tested stop and a logged reason.

What should never be automated with AI in a plant?

Safety functions should never be automated with AI, from emergency stops and guard interlocks to safety instrumented systems and other protective functions. Every use case starts in Observe and earns Advise and then Act on measured evidence, while safety functions stay in Never permanently. Functional-safety standards govern those functions, so the AI layer gets no path to them.

What is the most common first AI project in manufacturing?

Predictive maintenance is among the most common starting points. In Augury’s State of Production Health 2026 survey of 501 manufacturing leaders, it was the leading AI use case, deployed by 57% of respondents, though other surveys rank quality control or process optimization first. Whatever the use case, the strongest first project starts at the telemetry and historian layer, in the Observe tier, because the data already exists there. A read-only model scored against failures that happen anyway carries little risk. It also produces the error profile that later promotion requires.

How long before an AI system pays back in a plant?

Payback depends on the downtime cost of the assets involved, so this playbook gives no generic range. As one documented example, SumatoSoft’s predictive maintenance layer for a multi-plant manufacturer delivered a 2.5x return on investment within 8 months. The project began with an 8-week pilot on selected assets. Unplanned downtime fell by 50% and false alerts by 40%.

Why do manufacturing AI pilots stay stuck on one line?

Deloitte found that only 20% of manufacturers’ AI use cases run at scale across sites. Pilots stay on one line when the attachment was built for that line alone, with hand-mapped tags, a custom connector, a single-line model and alerts routed to one supervisor. A shared data model and reusable read-only connectors let the second line reuse the first line’s work.

Conclusion

Manufacturing AI adoption is early because attaching models to working plants is hard. The models themselves already deliver, since 84% of the manufacturers Deloitte surveyed report measurable value. Most of the work in AI for manufacturing goes into attaching a model to a plant that cannot stop, and choosing the use case settles very little of it. The plants that scale build that attachment once and move each use case up the trust ladder on evidence, with a fixed line around safety functions. The AI layer attaches beside the control system and stays outside the control loop.

SumatoSoft has delivered 350+ custom solutions for clients in 25+ countries since 2012, with a 98% client satisfaction rate. The company holds ISO 27001 and ISO 9001 certifications. Its manufacturing teams add the AI layer alongside the systems a plant already runs, without touching production-critical components. They keep safety functions out of scope from the first scoping call.

Map where AI can attach in your plant

Bring your equipment list and whatever data you already collect. Our lead architects will map where AI can attach without touching production-critical systems and which tier each use case starts in. They will also mark which safety functions are out of scope and which use cases have no data yet. SumatoSoft is ISO 27001 certified and has delivered 350+ custom solutions since 2012.

Book your free IoT strategy call.

Sources

• US Census Bureau, Business Trends and Outlook Survey, sector estimates for AI use in any business function, collection periods November 17 to 30, 2025 and October 5 to 18, 2026
• US Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users, May 26, 2026
• Bonney, K., Breaux, C. L., Dinlersoz, E., Foster, L. S., Haltiwanger, J. C., & Pande, A. A., The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks, NBER Working Paper 35141, April 2026
• Board of Governors of the Federal Reserve System, Monitoring AI Adoption in the U.S. Economy, FEDS Notes, April 3, 2026
• Deloitte, AI in Manufacturing 2026: From pilot value to scaled industrial impact, June 30, 2026
• Augury, State of Production Health 2026, March 2026
• Deloitte Insights, 2025 Smart Manufacturing and Operations Survey, May 1, 2025
• Deloitte, The State of AI in the Enterprise: The Untapped Edge, January 21, 2026
• Merzlova, K., & Funtikov, K., AI Readiness: How Companies Move from AI Pilots to Production, SumatoSoft Research, April 29, 2026
• SumatoSoft, AI-powered predictive maintenance for a large industrial manufacturer, case study
• International Society of Automation, ISA/IEC 62443 Series of Standards
• National Institute of Standards and Technology, SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security, September 2023
• International Organization for Standardization, International Electrotechnical Commission, IEC 61511, Functional safety — Safety instrumented systems for the process industry sector
• International Electrotechnical Commission, IEC 61511, Functional safety — Safety instrumented systems for the process industry sector

About the author

Yury Shamrei founded SumatoSoft in 2012 and leads the company as CEO. He co-authored SumatoSoft’s predictive maintenance playbook and wrote its analysis of enterprise AI adoption risks.