DigestNovember 22, 2026Редакция 105.kz

Production data gap slows physical AI on the shop floor

Robotics and physical AI attracted almost $48 billion in 2026, but they will not reach the shop floor without connected production data. A review of the column by the CTO of VulcanForms: what exactly gets in the way and which pilot steps make sense for a plant.

ManufacturingDigital Twins

Production data gap slows physical AI on the shop floor

Key takeaway. The main barrier to physical AI in the plant is not a shortage of robots but disconnected data: process parameters live in the process control system, inspection results in the quality system, and the technologist's knowledge in logs. In the Manufacturing Leadership Council 2026 survey, 37% of manufacturers named data incompatibility as an obstacle to the smart factory. It is worth starting with one processing stage and a “parameter—inspection—quality” link rather than buying new sensors.

What happened

On November 21, 2026, The Robot Report published a column by Michael Kenworthy, CTO of VulcanForms. The starting point: robotics and physical AI companies attracted almost $48 billion in 2026, yet the arrival of fully autonomous machines on the production floor is not guaranteed. To train robots, data on real production conditions is needed, and most sites are not set up to provide it.

An example from the text is welding. For an autonomous system to improve the technique itself, it needs to understand how welding angle, travel speed, workpiece position and other parameters affect the finished result. Experienced technologists accumulate this understanding over years, but a significant part of the knowledge is not recorded in a form suitable for training. The author cites data from a Manufacturing Leadership Council survey (2026 Smart Factories and Digital Production Survey): 37% of respondents named a lack of data compatibility as a serious obstacle to implementing smart factory strategies, whereas in 2025 this figure was 22%. Separately quoted is Lior Susan (Eclipse): “If your physical asset isn’t generating real-time operational data, it’s a dead asset”.

What this means for enterprises in Kazakhstan

The source describes the plant environment in general, not industry specifics. For Kazakhstani enterprises the practical conclusion is more modest than it appears: this is not yet about autonomous robots but about basic traceability of “mode — result”. Let us break it down by types of production where this gap is most visible (assessment).

Welding and workpiece preparation areas. Process parameters are usually available in the controller or in the process sheet, while the result is in the visual and dimensional inspection report and in the QCD acceptance certificate. As a rule, there is no link between them in a form suitable for analysis. Without it, neither robot training nor offline analysis of defect causes relies on fact.

Heat treatment and furnaces. The author explicitly notes the asymmetry: image recognition systems in metal additive manufacturing machines are becoming the norm, while integrating downstream equipment, such as heat treatment furnaces, into the same system is noticeably more difficult. For a plant this means that the “tail” of the process — the most expensive in terms of consequences — most often falls out of the digital history of the part.

Additive manufacturing and pilot production areas. The source lists factors affecting part properties: powder or wire characteristics, processing temperatures, environmental conditions. Linking these variables to inspection results is a separate engineering task, and it cannot be closed by purchasing sensors alone.

Data and the IT landscape. At Kazakhstani enterprises, process parameters usually live in the process control system and SCADA, inspection results in the quality system or in a paper log, and the batch ID often does not match between systems. Data localization and information security requirements of national companies mean that the link is built on-premise, with Russian and Kazakh in documentation and reference data, taking into account GOST requirements and internal standards. Legacy process control systems without open interfaces are a typical constraint, and it must be taken into account at the audit stage, not at the implementation stage.

Let us emphasize separately: the applicability of physical AI at Kazakhstani plants is still a hypothesis (assessment). The source contains no country-specific data; the Manufacturing Leadership Council survey reflects a different sample.

How to apply: an 8–12-week pilot

The logic of the pilot below is managerial, without implementation-level engineering recipes.

  1. Choose one processing stage and one measurable output. Not “the plant” but a specific operation with a clear quality indicator: defect rate by inspection, parameter spread, repeatability. The source advises starting with the target application so as not to rebuild the entire infrastructure at once.
  2. Conduct a data audit. Inventory the sources: controllers and SCADA, operator logs, quality system records, inspection reports, the reference data for batches and workpieces. Record: whether there is a single part ID, whether time is synchronized, how often records are actually reviewed.
  3. Define the link and acceptance rules. Record which process parameters must be linked to which inspection result, who is responsible for the correctness of the record, and by what criterion the pilot is considered successful or failed.
  4. Build a minimal digital thread in the selected area. According to the source, where there is no “parameter—quality” link, it is more useful to connect existing records than to add new sensors. The approach to describing a physical asset and its history is on the Digital twins page; industry context is on the Manufacturing enterprises page.
  5. Set autonomy boundaries and escalation rules. Before any automatic adjustments, define permissible limits for parameter changes, how the result is checked, and the conditions under which engineering review is required. A system that changes the mode but does not evaluate the final quality is not ready for independent decisions — this is the source's thesis.

The economic effect of the pilot is measured only by the pre-selected indicator and in the selected area; transferring it to the entire plant is incorrect (assessment).

Limitations and risks

The material in the source is an authored column by a vendor's CTO, not independent research. The figures 37% and 22% are taken from an industry survey; the sampling methodology is not disclosed in the text; $48 billion is the volume of attracted investment, and it does not confirm the maturity of the technology. The source provides no direct measurements of the effect of closing the described gap: the statement “data is needed” remains an argument, not a measured result.

Nor is it proven that linking process parameters to inspection results is sufficient for autonomous adjustments: the author himself limits the conclusion — without evaluating final quality, the system is not ready to make decisions. In regulated production (the source's example of medical devices), process changes require additional validation; at Kazakhstani enterprises, a similar constraining loop is the work permit, the technologist's responsibility and internal standards, not only technical readiness (assessment).

The data without which the pilot cannot start: an end-to-end part or batch ID, traceability across processing stages, time synchronization between systems, readable records of actual modes and inspection results in machine-readable form. If these elements are absent, the first stage is not a model but putting the records in order. It is impossible to guarantee an economic effect from implementing physical AI based on any of the steps described in the source.

Sources

  1. The manufacturing data gap holding back physical AI · 2026-11-21T22:25:35.000Z
    37% of surveyed manufacturers cited a lack of data interoperability as a major roadblock to implementing their smart factory strategies, up from 22% in 2025.
  2. Фото: WireCrafters — Robotic Welding Cell.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.

Do we need to buy new sensors to prepare the plant for physical AI?

Not necessarily and not first. The source states directly: where there is no link between mode and quality, it is more useful to connect existing process records with inspection results than to increase the volume of machine data.

What does the digital thread give if there are no robots in the area yet?

Traceability of “parameter—inspection—quality” is needed both for ordinary defect analysis and for justified mode changes by the technologist. Robotization here is the next step, not a condition.

Can a system be allowed to change process parameters automatically?

Only after permissible limits for changes, the method of checking the result and criteria for engineering review have been defined. In regulated production, process changes additionally require validation.

How applicable are the source's conclusions to Kazakhstani plants?

As industry practice, they are applicable: problems of data compatibility and disparate systems are universal. There is no quantitative data on Kazakhstan in the source, so the applicability of specific steps at your enterprise remains a hypothesis until an audit.

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