5 mistakes with AI agents, and why manufacturing needs different answers than the startup world

Jamin Mahmood-Wiebe listed five mistakes companies make with AI agents in the German technology magazine t3n. The analysis is accurate. It is also written from the perspective of software teams, and that perspective changes what the fixes have to look like once the agent touches a supply chain.

Mistake one: demo against production, or the Potemkin factory

What the original describes. In a pilot with 500 requests everything runs cleanly: 95 % accuracy, two seconds response time. The board is delighted, the budget is released. In production, at 10,000 requests a day, accuracy drops to 80 % and latency explodes by a factor of twenty. The recommendation is load testing with real data and a staged rollout: 5 % of volume, then 20, then 50.

The industrial reality. The recommendation is right and it stops short, because the problem starts in the pilot.

95 % accuracy sounds like a solid starting point in web development. In a supply chain it is a disaster. At 1,000 automated purchase transactions a week, a 5 % error rate means 50 faulty transactions. Each one can trigger an express shipment, a contractual penalty or a stopped line. Conservatively, the financial impact per error in an automotive setting runs from 200 to 500 euros. That is 10,000 to 25,000 euros a week. Half a million to a million a year.

And the staged rollout? Manufacturing has known that concept for forty years. We call it a ramp-up curve. Every production manager knows you do not run a new machine at full rate on day one. You bring it up shift by shift, measure scrap rates, adjust parameters.

AI agents need the same treatment. Not a software deployment with a release date, but the commissioning of a machine, with a ramp-up curve, quality gates and a supervisor standing next to it.

That is not a rollout plan. It is a commissioning protocol.

Mistake two: vibe coding, or when the intern programs the factory

What the original describes. An AI network was exposed with 4.75 million records lying open, because the founder had let an AI assistant generate the entire platform without a security review. Backslash Security reports that GPT-4o produces vulnerable code in 90 % of cases. The recommendation is automated security scans and mandatory code review.

The industrial reality. In manufacturing the code problem is real, and it is not the primary risk. The primary risk is the process problem.

Picture an internal AI champion, motivated, technically capable, without deep process knowledge, building an agent that generates purchase requisitions in the ERP. It works in test. It gets rolled out. What nobody checked:

• Does the agent know the minimum order quantities from the framework agreements?
• Does it respect blocked-stock logic?
• Does it know that supplier X only ships against prepayment?
• Does it understand that hazardous goods take different freight routes?

In most cases the answer is no. Not because the code is bad, but because the business rules were never made explicit. They live in the head of the planner who has done the job for fifteen years. They appear in no specification and in no prompt.

Web development needs a security review for vibe coding. Industry needs a process review: does the agent understand the business rules nobody ever wrote down?

That is the moment adult supervision stops being a metaphor. Somebody has to stand between the AI system and the ERP system who understands both worlds. Somebody who speaks the language of the engineer and the language of the algorithm.

Mistake three: hidden costs, or the token trap in purchasing

What the original describes. One company started at 500 dollars of agent cost in week one. By week four it was 18,400. Nobody had defined a cost dashboard or a loop budget. A DeepMind and MIT study confirms that multi-agent systems cost several times more per solved task than single agents, while producing worse results.

The industrial reality. In the technology world that cost explosion is an infrastructure problem. In industry it is a management accounting problem, and it belongs in familiar categories.

The decisive question is not what the tokens cost. It is what the agent costs per business transaction.

• What does the agent cost per automated purchase order?
• Per supplier evaluation?
• Per request for quotation processed?

If the answer is four times what the clerk costs doing it by hand, that is not progress. That is innovation theatre with a negative business case.

Manufacturing already has the instrument for this. It is called a cost centre. Every agent needs a virtual cost centre with a budget cap. When the agent has used its token budget for the month, it does not go into overdraft. It escalates to a human.

And the finding that multi-agent systems cost more than single agents? That is the engine-and-bodywork trap in its purest form. Three engines in one car do not make it faster. They make it more expensive, heavier and impossible to maintain. One good engine, built deep into the right chassis, beats three loosely wired engines on a lift. Every time.

Mistake four: multi-agent systems without physics

What the original describes. In 180 experiments, DeepMind and MIT showed that once a single agent solves more than 45 % of a task correctly, additional agents bring almost no improvement. On sequential tasks they make results worse, by 39 to 70 %.

The industrial reality. For a supply chain that finding is brutal, because supply chain processes are sequential.

Demand determination, requisition, quotation comparison, purchase order, order confirmation, goods receipt, invoice verification.

Every step depends on the one before. Every error propagates. That is not a software problem, that is physics. Or more precisely, it is production control logic that every planner knows in their sleep.

Now let one agent perform supplier evaluation at 60 % accuracy and have a second agent review the result. What happens is not what the architects hoped. The second agent does not do it better. It does it differently. And suddenly two agents disagree about the risk rating of supplier X. Who decides?

A human. The same human the exercise was meant to relieve.

The answer is not less ambition. It is more focus: one agent per clearly bounded process step, with defined inputs and outputs, and a human gate between steps until trust has been earned.

In manufacturing the principle is called takt. Each station does one thing. Correctly. Reliably. And when a station fails, everyone knows immediately where the problem is.

AI agents need takt, not teamwork.

Mistake five: context engineering, or nothing works without master data

What the original describes. Anthropic showed that a model initially scored only 42 % on a benchmark, not because of the model but because of rigid evaluation criteria. After fixing the evaluation rather than the model, performance jumped to 95 %. The conclusion: what decides is not the model but the context the agent sees at each step. Context engineering is the real architectural discipline.

The industrial reality. This is where the translator's heart beats loudest, because what is called context engineering is an old and chronically neglected topic in industry: master data quality.

I saw it years ago at an automotive supplier near Stuttgart. We implemented a transport management system, technically flawless, cleanly configured. Then the reality check: the system could not optimise a single route. Not because the algorithm was poor. Because nobody knew what the parts weighed. Weights and dimensions, the most fundamental master data in logistics, were missing. Or wrong. Or in a spreadsheet nobody had updated in years.

That was 2015. It has not improved by 2026. Ask an agent today to evaluate the delivery reliability of supplier X and it needs more than access to goods receipt data. It needs:

• correct supplier classification
• current framework agreement data
• historical deviation rates
• information about alternative suppliers
• the tolerance limits per material group

If one of those is missing, wrong or outdated, the agent does not return a wrong answer. It returns a plausible wrong answer. That is more dangerous than an obvious error, because nobody questions it.

Context engineering in industry is not prompt optimisation. It is the combination of master data management and process knowledge. No prompt engineer who has never touched a bill of material can deliver that combination.

Process knowledge beats architecture

Mahmood-Wiebe writes that architecture beats technology. Agreed, with an addition:

Process knowledge beats architecture.

The best token budgets, load tests and evaluation frameworks are worth nothing if the person designing the agent has never touched a bill of material. If they do not know what a blocked-stock posting is. If they cannot tell a framework call-off from a single order.

The five mistakes are real. Their industrial translation shows that the answer lies neither in better technology nor in better architecture, and certainly not in better prompts.

It lies in adult supervision. In people who understand technology and operational reality at the same time. Who do not just admire the engine but can fit it into the right chassis. Who know that a ramp-up curve is not a rollout plan but a commissioning protocol.

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

---

#AgenticAI #AI #MidSizedIndustry