56 % of CEOs see no ROI from AI, and they’re seriously surprised?

Harald Weiss published a sober analysis in the German technology publication heise: 56 % of CEOs see neither revenue nor cost benefits from their AI investments. Only 12 % report measurable success on both sides. At the same time, 67 % of companies are increasing their GenAI budgets.

Anyone working in manufacturing is not surprised by this. Anyone who is surprised has not understood the problem.

The basic misunderstanding: agentic AI is not automation 2.0

Weiss puts his finger on it. Expectations of agentic AI rest on experience with classical process automation. A stable process is digitalised, a manual step disappears, the effect is measurable.

That is the mental model of the past twenty years. For agentic AI it is wrong.

An image for it. Classical automation is a transfer line. Workpiece in, workpiece out, always the same. Every cycle predictable, every fault reproducible.

Agentic AI is a workshop with a journeyman who decides for himself. He picks the tool, interprets the drawing, chooses the sequence. Usually he is right. Sometimes he is not. And when he is wrong, nobody notices until the part fails at the customer.

A transfer line you can set up and forget. A journeyman you have to lead, continuously. That is the difference most companies have not priced in.

Three hidden cost blocks, translated to the factory

Weiss identifies three cost drivers that appear in no business case. In the language of manufacturing:

One: data integration, or the master data disaster

Weiss writes of context stitching, the merging of inconsistent data from ERP, CRM, MES and proprietary systems. Of semantic inconsistencies, missing identifiers and mismatched time references.

On the floor it sounds like this:

The material master says the part weighs 2.4 kg. The logistics spreadsheet says 2.7 kg. The scale at the loading dock reads 3.1 kg, because nobody ever entered the packaging.

I described exactly this case in an earlier article: an automotive supplier near Stuttgart who could not get a transport management system running, not because of the software, but because nobody knew what the parts weighed.

Agentic AI multiplies that problem. A classical system throws an error when data is missing. An agent improvises. It takes the nearest available value, interpolates, estimates, and nobody sees the mistake until the pallet does not fit on the truck or freight costs explode.

Context stitching is not an IT project. It is an operational overhaul. Skip it before the first agent goes live and you are building on sand.

Two: interface logic, or when the agent writes instead of reads

The most dangerous passage in the article is this one: because an agent does not only read but also writes, closing tickets, triggering orders, changing system parameters, transaction safety becomes the critical factor.

Translated into the supply chain:

A purchasing agent rates a supplier as reliable on the basis of available data and automatically releases a call-off of 50,000 parts against a framework agreement. What the agent does not know: last week that supplier moved its payment terms from 30 to 90 days, a classic early warning sign of liquidity trouble. The information sits in an email from the head of purchasing, not in a structured data field.

Result: 50,000 parts with a supplier who may not be delivering in three months. And the agent did everything correctly, within its decision space.

Weiss describes this technically as idempotency, rollback strategies and race conditions. In a factory it means: giving an agent write access to the ERP without understanding the chain of consequences is roulette with the supply chain.

That is not a prompt engineering problem. That is a lack of industrial experience.

Three: the cost of assurance, or the myth of headcount savings

The most seductive aspect of agentic AI is staff reduction. Salesforce cut 4,000 positions. The headline works. Weiss dismantles the myth: human-in-the-loop is frequently replaced by human-on-call.

Translated: the planner who used to process 200 orders a day manually now handles 20 escalations the agent could not resolve. Those 20 are the hardest, the most ambiguous, the riskiest. The planner works less and under more pressure, with higher complexity and without routine as an anchor.

I know this pattern from lean work. In the 2000s we automated production lines in the furniture industry and reduced headcount. What remained were the disruptions. The remaining people needed more qualification, not less. Whoever had let the best people go stood in front of machines nobody could run.

With agentic AI the same pattern repeats, faster and with higher stakes.

And then comes the cost block that appears in no business case: monitoring, incident handling, compliance checks, drift detection, version conflicts between model, prompt and tooling. Weiss calls it observability. I call it the invisible factory behind the AI factory. It needs people, processes and budget, permanently.

Why the majority sees no return

The figures tell a clear story. Most companies systematically underestimated the total cost of agentic AI. Not because the technology fails. Because the mental model is wrong.

Companies calculate like this:

Cost = licence + inference + integration Benefit = positions saved × annual salary

That is the arithmetic of a transfer line, not of an autonomous system.

The correct calculation would be:

Cost = licence + inference + integration + master data remediation + interface assurance + monitoring infrastructure + escalation staff + compliance + continuous training Benefit = decision speed × decision quality × scalability, minus the cost of errors when the model hallucinates

Only someone who understands both the technology and the operational reality can draw up that calculation. Someone who knows what a wrong disposition costs. What a supply failure in week 37 means. What happens when an agent reads blocked stock as available.

That is the business quotient, and it is missing in most AI projects. Not on the technical side. On the decision-making side.

Why this hits mid-sized industry hardest

Large corporations can afford the cost of assurance. They have their own ML-Ops teams, compliance departments, test infrastructure. When an agent misbehaves, it is an incident. Not an existential risk.

Mid-sized manufacturers do not have those buffers. A failed purchasing agent that releases five wrong orders overnight can strain the cash flow of a fifty-million company for a quarter. And there is nobody in the building who spots model drift before the damage happens.

The answer is not to avoid agentic AI. The answer is adult supervision: experienced people who do not just admire the engine but know which chassis it belongs in. Who calculate total cost before the budget is released. Who see the agent not as a replacement for the planner, but as a tool the planner operates.

The return is not in the model

Agentic AI is not self-running. It is a leadership task, not an IT task.

The CEOs who see no return did not buy the wrong tool. They asked the wrong question. The question was what AI can do for us. The right question would have been whether our processes, our data and our people are ready for a system that decides on its own.

Anyone who cannot answer that should not release a budget. Anyone who can usually needs less budget than planned, because half the intended agents become unnecessary once the processes are understood.

That is industrial translation. And what manufacturing needs now is not more agents. It is the right people steering them.

---

#AgenticAI #AI #ROI