BMW thins its management with AI: has the time come for the supervised artificial manager?

On September 30, 2026, during its Capital Market Day, BMW Group presented a program to regain profitability and agility. Its official statement mentions a 20% reduction in the number of divisions and associated management positions by mid-2027, with a comparable reduction across lower organizational levels. It also plans to simplify processes and expand the use of artificial intelligence in development, purchasing, production, sales, marketing, and aftersales.

The company expects digital agents to analyze complex information, automate routine tasks, and accelerate decision-making. In engineering, BMW was explicit: AI outputs will be monitored, reviewed, and ultimately approved by developers. That phrase defines the responsible model better than an easy headline: broad automation, yes; identifiable human authority, too.

BMW links these measures to a financial goal: bringing the automotive business's EBIT margin to between 3% and 5% in 2028 and returning to the historical range of 8% to 10% by the beginning of the next decade. It does not attribute this entire outcome to AI; it will also cut variants, regionalize products, and reorganize its structure.

Operational honesty. An AI does not need to doctor metrics to protect its career or conceal a mistake to preserve influence. It can log what data it read, what rule it applied, and why it recommended an action. But that honesty depends on data integrity and on ensuring no one configures the system to reward a convenient conclusion.

Data-driven decisions. It can consolidate signals from ERP, CRM, production, support, and treasury into a single reading; model scenarios and update priorities with up-to-date data. It does not replace judgment when a problem is ambiguous, but it reduces decisions made solely on intuition or driven by the most persuasive presentation.

Fairness and absence of personal favoritism. A machine has no friendships, grudges, or protégés. It can apply the same documented criteria to equivalent cases. However, it can also inherit historical biases from evaluations, compensation, or hiring; this is why fairness must be measured through periodic audits, the right to an explanation, and human review.

Speed. It can process thousands of transactions, detect exceptions, and prepare options in minutes. The advantage is not rushing every decision, but rather reserving human time for irreversible decisions while repetitive workflows advance within pre-approved thresholds.

Connection with ERP and other applications. Integrated via least-privilege permissions, AI can transform scattered information into coordinated alerts and actions. Every access and every modification must be logged; no artificial manager should have a master password or unchecked authority to execute payments, dismiss, hire, or alter records.

Continuous improvement. Decisions can be compared against real-world outcomes to fine-tune forecasts and identify obsolete rules. Learning must not mean modifying itself without oversight: new versions, data sources, and policies require testing, approval, and rollback capabilities.

Preservation of know-how. Procedures, exceptions, historical context, and expert solutions can be turned into searchable institutional memory. When someone retires or moves to another company, that knowledge does not vanish inside their inbox. The company, in turn, must respect authorship, confidentiality, and access rights over that expertise.

Continuity and consistency. The system can operate through shift changes, vacations, or peak workloads without losing context. This does not eliminate leaders; it prevents an entire department from relying on the memory or availability of a single individual.

It is worth dispelling an illusion: a model possesses neither ethics, loyalty, nor a sense of justice. It produces results based on data, instructions, tools, and objectives. If the metric rewards cutting costs without protecting quality, safety, or decent employment, it will optimize the wrong figure with formidable discipline.

Nor is it enough to say that it has no favorites. If it learns from a company that for years promoted one group more than another, it can turn that inequality into a statistical pattern. And if its explanations are impossible to audit, apparent neutrality can conceal decisions that are harder to challenge than those of a human boss.

That is why the decisive word is supervised. A person or committee must be accountable for the objective, permitted data, autonomy thresholds, and consequences. The company needs to separate recommendations, reversible actions, and high-impact decisions; the latter must retain effective human review.

In a commercial department, for example, AI could review sales, margins, inventory, accounts receivable, complaints, and delivery capacity every morning. It would detect that a promotion increases revenue but erodes margins, or that a region is selling well while accumulating returns. It would then propose reallocating budget, adjusting the forecast, or escalating an exception.

Low-risk actions—creating an alert, requesting information, scheduling a report—could execute autonomously. Medium-risk actions would require approval from a manager. High-impact actions—contracts, disciplinary measures, salary changes, credit approvals, or fund transfers—would remain under human decision-making and segregated controls.

AI performance would be evaluated just like that of any manager: forecast accuracy, value generated, errors, compliance, fairness, team satisfaction, and the ability to explain decisions. If only speed or cost savings are measured, the system will learn to sacrifice whatever the company forgot to put on the dashboard.