Barriers to AI adoption: why data and integration slow you down

Barriers to AI adoption: why data and integration slow you down

The biggest barriers to AI adoption in manufacturing are data access, IT/OT integration and a lack of operational responsibility, not the technology. In our analysis of 52 practitioner conversations, 42 of 206 pain point statements concern data, IT/OT and integration. In surveys, 77 % name data protection requirements as the number one obstacle (Bitkom 2026, the German digital industry association).

For decision makers in brief

• Effort: The largest block comes before the first model. According to Capgemini 2025, 80 % of initiatives in industry first require investment in data architecture and OT/IT integration.
• Cost: These foundation costs rarely appear in the quote. If you hide them in the business case, you create the disappointment that is later booked as a failed AI project.
• Risk: In 39 % of manufacturing companies, AI is merely tolerated (ZEW 2026, the Leibniz Centre for European Economic Research). There, nobody knows which data flows where.
• Lever: Clear data classification, one approved tool, a narrowly scoped first use case and defined operations.

What surveys name as obstacles

The obstacles are well measured. The ranking is striking: data protection comes before skilled staff, skilled staff before time, time before security, and security before money. The problem is rarely that a model does not work. The problem is that it cannot reach the data and nobody has time to run it.

Data protection as an AI barrier is thus clearly in first place. Behind this lies a concrete concern: design data, cost calculations and customer information could end up in public services. If you take this concern seriously, you quickly arrive at the question of where a model runs. One answer is AI in your own data centre, where data leaves neither your premises nor the EU.

What practitioners say: 206 statements by topic

Surveys measure agreement with predefined categories. To check whether the findings also appear where practitioners speak freely, we analysed 52 published practitioner conversations from industry (September 2025 to August 2026) and coded 206 individual pain point statements. Multiple assignments were possible.

The largest topic area covers data silos, missing standards, heterogeneous machine fleets, retrofitting old equipment, OPC UA and PackML as a gap, and IT landscapes that have grown over time and carry technical debt. No other topic comes even close. Technology as such barely appears as a problem. Why this adds up to an implementation problem rather than a technology problem is explained in our article on why AI projects fail.

One methodological caveat belongs here: the analysis reflects how the speakers present themselves and serves as a qualitative cross-check, not as representative statistics.

Two perspectives, one finding

At first glance, the two data levels contradict each other. The surveys name data protection and skilled staff first, the conversations data and integration. In fact, both describe the same thing: the survey names the risk, the conversation names the effort. If you want to understand the barriers to AI adoption, you therefore need both perspectives.

The gap also shows in how widely AI is used. In machinery and plant engineering, 43 % use AI somewhere in the company (VDMA 2025, the German Mechanical Engineering Industry Association), yet only 13 % in their own production (Fraunhofer ISI 2024). On the shop floor, machine data, OT networks and safety requirements stand in the way. A manufacturer study without disclosed methodology provides an indication: 93 % of manufacturers have an MES, yet only 28 % have scaled it across the company.

The technical layer

Data quality in manufacturing is rarely a problem of volume. It is a problem of access: which data sits in which system at which frequency, and who releases it? This question comes before any choice of model. Regulation adds to this. Since 12 September 2026, the Data Act obligation of access by design has applied to newly placed connected products. The VDMA points to OPC UA as the technical implementation path. In our reading, this makes a clean data connection a mandatory task for machinery manufacturers in any case.

The organisational layer

A ZEW survey from spring 2026 reveals a governance vacuum. In manufacturing, AI use is explicitly prohibited in 8 % of companies, only 30 % actively provide AI applications, and 39 % merely tolerate its use. Tolerance is not governance. Bans do not solve the problem either, they push it into shadow IT. What works is a short policy plus one approved tool. We describe which technical and contractual evidence this requires on our page on GDPR compliant AI.

On the subject of co-determination, Bitkom published a guide in February 2026. Involving the works council before the pilot, not before the rollout, has proven effective. Works agreements negotiated after the fact regularly cost more time and money.

Overcoming barriers to AI adoption: five rules

The industry report summarises five rules that have proven themselves in projects. They address the barriers to AI adoption directly.

The last rule deserves particular attention. In the conversations we analysed, there is a lot of talk about use cases and almost none about where the model runs, who operates it and how operations are secured. According to the overall picture of the analysis, this is exactly where scaling fails. We therefore take on not only the build, but also AI operations as a managed service: monitoring model quality, retraining, documentation. Our general operations target is 99.9 %.

A further finding from the report: solutions developed through strategic partnerships reach productive operation in around 67 % of cases, purely in-house developments in only around 33 %. The bottleneck is not resources, but the organisational approach. According to Schneider Electric (April 2026), 80 to 90 % of the work in an AI transformation is change management, training and redesigning workflows.