Technology releaseNovember 20, 2026Редакция 105.kz

72% of manufacturers use AI, but only 10% at scale: three shifts in automation

AutomationWorld analysed three shifts in industrial automation over a 12–18 month horizon: moving quality control to the start of the cycle, spatial AI for the digital twin, and adaptive robots. The key barrier is legacy systems and fragmented data.

ManufacturingEnergyDigital Twins

72% of manufacturers use AI, but only 10% at scale: three shifts in automation

Key takeaway. AutomationWorld (20.11.2026) records the gap: 72% of manufacturers use AI in some form, 10% at industrial scale. Three directions for the next 12–18 months: quality control at the start of the cycle with a closed loop, spatial AI for the digital twin, and adaptive robots for low-volume production. For a plant in Kazakhstan the practical conclusion is a pilot on one line, clean time-stamped data, bounded system authority, and operation inside the plant perimeter.

What happened

On 20 November 2026 the publication AutomationWorld released an overview of three shifts which, in the wording of the material, will determine manufacturers' competitiveness over the next 12–18 months. The starting point is survey results: 72% of manufacturers use AI in some form, but only 10% have brought that use to industrial scale. The second pressure is workforce-related: if the labour gap persists, up to 1.9 million manufacturing jobs could remain unfilled by 2033.

The first shift is moving quality to the start of the life cycle. According to the source, more than 70% of a part's cost is fixed before production and procurement get the opportunity to influence it; tools that read geometry flag thin walls, undercuts and tolerances tighter than required, and estimate cost drivers before tooling is calculated. On the shop floor the emphasis moves from detecting a defect after the operation to preventing it during the operation: a closed loop feeds measurements from vision systems, metrology and sensors into the machine controller and adjusts feed, thermal limits or injection pressure; individual deployments report reductions in scrap costs of up to 25%. The second shift is spatial AI: a 3D model of the shop floor is continuously updated from cameras and lidar, and a schedule or layout change is checked against the current state of the floor rather than against a survey from last quarter. The third is physical AI: adaptive robots with perception instead of hard-programmed cells; in one pilot a humanoid robot processed 90,000 components at an automotive plant in the USA. The main barrier is stated plainly: 48% of industrial respondents cite integration with legacy systems and fragmented data.

What this means for enterprises in Kazakhstan

Plants: machining, casting, stamping

The gap between experiment and scale (72% versus 10%) for Kazakhstani machine-building sites is most likely comparable (estimate). Two classes of tasks are of practical interest. The first is pre-production verification of a part's geometry and cost from CAD before launch: relevant for repair and tooling shops that process dozens of new part numbers per month, where the cost of an error in tooling exceeds the cost of a design discussion. The second is a closed quality loop in an area with measurable scrap: machining, die casting, stamping.

Data for both classes: CAD models and design and process documentation, controller and SCADA logs with time stamps, QC and metrology measurement results, video from the area, process parameters. The limitation is named in the source itself: exchanging data with legacy SCADA and MES is the most labour-intensive part. For national companies and subsurface users, information security requirements and the data localisation rule are added, so collection and processing are usually kept inside the plant perimeter, on-premise (estimate). A system that writes to machine parameters expands the attack surface — this is the source's point, not an industry scare story.

Energy: repair and mechanical services

For the power sector what is relevant is not mass assembly but repair and tooling production, manufacture of spare parts and steel structures, warehouses and repair sites. Here the third block of the material is closer — an updatable 3D model built from cameras and lidar: it gives a current picture of a site where equipment, walkways and storage zones change, and allows a layout or schedule change to be checked against the current state rather than against a past survey. Data: video, laser scanning results, layouts, SCADA events. For continuous production, automatic writing to process setpoints is a separate level of risk; the article describes closed loops only for discrete manufacturing.

How to apply: an 8–12 week pilot

  1. Choose one area with measurable pain. Not "AI at the plant", but one line where the reporting already contains a number: scrap rate, rework cost, changeover time. The material states this condition directly — start with the line where the pain is measurable.
  2. Run a data audit. What is recorded, at what frequency, whether time stamps exist, how the batch, part, shift and process parameters are linked. Without clean time-linked data there is nothing to attach models to.
  3. Define the limits of authority. What the system changes itself, what it only proposes, who owns the result, how operator override works and how every change is logged. For regulated environments this is the condition for preserving traceability.
  4. Deploy a narrow loop inside the perimeter. Storage and processing on-premise, exchange with existing SCADA/MES through native interfaces, operator interfaces in Russian and Kazakh where the site requires it. Sector context — Industrial enterprises.
  5. Measure the effect and only then scale. A baseline before the start, a period of stable operation (in the source — several months), a separate assessment of the load on the information security function and on staff. For layout and updatable-model tasks — Digital twins.

Limitations and risks

  • 72%, 10% and 48% are survey results. The sampling methodology is not disclosed in the text; these are stated figures, not measurements taken at specific plants.
  • "Up to 25% reduction in scrap costs" is phrased as "individual deployments report". Without specifying the enterprise and the calculation method, this is an order-of-magnitude reference point, not a guaranteed effect.
  • 480 autonomous robots and 90,000 components are single sites and pilots under supervision. The author himself notes this: it proves that adaptive systems work in a human environment, not that universal robots are ready for any task.
  • There are no Kazakhstan statistics in the source. Any transfer of the conclusions to local practice is an estimate.
  • Data requirements are not disputed, but neither are they detailed. Without time stamps and batch linkage the pilot degrades into a demonstration.
  • Cybersecurity. A loop that writes to equipment parameters requires a change procedure, roles and a log; this is additional load on SCADA and information security.
  • Staff. An operator supervising an autonomous system is a different role from an operator responding to an alarm; training is planned within the same project, not after it.

Sources

  1. Three Industrial AI Trends Reshaping Manufacturing Automation in 2027 — AutomationWorld · 2026-11-20
    Manufacturers entered 2026 with no shortage of AI enthusiasm, yet a recent survey found that 72% have adopted AI in some form while only 10% have deployed it at scale.
  2. Фото: Globetrotter19 — Korridor , Auszeichnungen, Rathaus, 2024 Kisújszállás.jpg (CC BY-SA 4.0, Wikimedia Commons)

This material was prepared with AI from the sources listed above. Spotted an inaccuracy? Write to info@105.kz.

FAQ

Frequently asked questions

Short, straight answers to the questions we hear most often.

We use AI in two areas but have not scaled it — is that lagging behind?

According to the survey cited in the material, 72% of manufacturers use AI in some form and only 10% at industrial scale. The gap between experimentation and sustained operation is a common situation, not a sign that a particular plant is lagging.

Should we start with data or with choosing the area?

With the area where the pain is measurable — scrap or changeover — and immediately after that with an audit of data with time stamps. In the source, successful manufacturers start with one line and clean data, and add models later.

Can we allow the system to write to machine setpoints?

Only with explicit boundaries: the list of parameters it is permitted to change, the owner of the result, mandatory operator override and logging of every change. That is exactly how the strong deployments in the material are described.

Do we need a static digital twin before an updatable one?

The source describes the transition from static, rule-driven twins to a model continuously updated from cameras and lidar. The logic is simple: if the model does not reflect the current state of the shop floor, decisions are made on outdated data.

Next step

Discuss a pilot project

Tell us about your task — we'll come back with a data audit plan and an effect estimate within two business days.

  1. 1A 30-minute call: the task, your current systems, who makes the decision
  2. 2A review of data and processes, a staged plan with time and cost ranges
  3. 3A 2–4 week prototype or an 8–12 week pilot with success criteria agreed upfront

NDA from the first contact

Reply within one business day