The agent makes a decision for which there is no position on the org chart

Four out of ten companies with 20 or more employees use artificial intelligence. One third of them report higher costs than planned. The reason is rarely the technology. It is an organizational structure that was built for a problem which is now disappearing.

Thursday, 3:40 pm. A long-standing customer calls the sales office of an industrial company. Instead of 250 units he now wants 400, and he wants them in eight working days instead of fifteen. A good call, one would think. What it sets in motion inside the company is shown in Fig. 2.

Fig. 1 · The case

	Frame contract	Request
Quantity	250 units	400 units
Delivery time	15 working days	8 working days
Order value	60,000 euros	96,000 euros

Fig. 2 · What the request sets in motion

Department	Rule that applies today	What the rush order changes	Who has to decide
Sales	The frame contract covers 250 units at the agreed price, anything above is priced by the list	Price for 150 extra units and the size of the rush surcharge are open	Head of Sales
Finance	Customer credit limit 250,000 euros, 190,000 euros are outstanding	With the order 286,000 euros, the limit is exceeded, an increase to 300,000 euros is possible	Head of Finance
Plant	Production slots are allocated, sequence follows the plan	Three orders would have to move, one of them carries a contractual penalty	Plant Manager
Purchasing	Main supplier delivers in twelve days, extra costs up to 5,000 euros within own authority	Too late. Second supplier delivers in four days, 14 percent more expensive, around 2,500 euros extra	Head of Purchasing
Logistics	Delivery on the regular route	Full delivery in eight days needs a special transport, 1,400 euros	Logistics

Five departments have to check the request, each with its own rule and its own approval. Together this is less than one hour of work. But the customer waits two days for an answer, because every decision waits for the one before it.

An AI agent with access to order data, credit limits, capacity planning and supplier terms needs one minute for the same five checks. It presents a proposal that respects all five limits at the same time, Fig. 3.

Fig. 3 · What the agent proposes after one minute

Department	Proposal	Result
Sales	Split delivery: 250 units on the original date, 150 units four working days later. Rush surcharge of 1,700 euros on the extra quantity	Customer gets most of the order on time
Finance	Credit limit temporarily raised to 300,000 euros, the credit insurer covers the amount	Limit respected
Plant	Two non-critical orders moved by one day each, the order with the penalty is not touched	No penalty
Purchasing	150 units from the second supplier, 2,520 euros extra cost, of which 1,700 euros are passed on to the customer through the surcharge	820 euros stay with the company
Logistics	Both part deliveries on the regular route	No special transport

The proposal is good. The contribution margin is around 6,700 euros higher than if the extra quantity were refused, the calculation is in Fig. 5. But that is not the point.

The agent has done something that nobody in this company is authorized to do. It has taken a pricing decision, a credit decision, a capacity decision, a purchasing decision and a logistics decision in one step and weighed them against each other. Before, this decision did not exist as one decision. It was five partial decisions, each inside its own limit, and the compromise was negotiated between departments. This negotiation was the organizational structure at work.

The agent takes a decision for which there is no position on the org chart.

Why org charts look the way they look

Every organizational structure we know rests on one assumption: judgment is scarce, because it depends on people. The economist Luis Garicano described 25 years ago what hierarchy does in that situation. The front line learns the frequent cases. The rare cases move up to specialists who should not be occupied with routine. The person at the top of the org chart is the person for the case that the level below does not know.¹

In other words: the org chart lives in the exceptions. The standard process has been running in the ERP system for twenty years, and nobody needs a manager for that. Approval limits, escalation paths and spans of control exist for the rush order on Thursday afternoon.

This exception is exactly what the machine now takes over. It attacks the very function the hierarchy was built for. Clerical work was automated long ago. Now it is about judgment.

What the numbers say, and what they do not say

Adoption is growing fast, returns are not. In companies with 20 or more employees, the share of AI users has more than doubled within one year.² One third of these users report clearly higher costs than expected,³ and only one third see a high effect on productivity, Fig. 4.⁴

The gap between what companies expect and what has actually happened is revealing. In 2025 the ifo Institute asked companies whether they had cut jobs because of AI: 3.9 percent said yes, 93 percent reported no change. Asked about the next five years, 27.1 percent expect job cuts, in manufacturing 37.3 percent.⁵ So one quarter expects a restructuring that almost nobody has carried out so far. If technology were the obstacle, this gap would be smaller.

The reason has a name, it is just an uncomfortable one. Erik Brynjolfsson and Lorin Hitt showed at 527 US companies that IT investments bring only normal returns in the short term, but up to five times higher productivity contributions over five to seven years. They explain the difference with large and time-consuming investments in what they call organizational capital.⁶ Later, Brynjolfsson, Rock and Syverson turned this into the productivity J-curve: general-purpose technologies require complementary investments in processes, roles and skills, which cause costs first and returns only later.⁷

Translated for company leadership: whoever measures the return on AI spending today is measuring in the dip. The organizational investment already shows up as cost, but not yet as return. Whoever concludes from this that AI does not work draws the wrong conclusion from a correct observation. And the other side: whoever never makes the organizational investment never gets out of the dip. That company only has the costs. This is the situation of the one third that complains about higher costs.

Three answers from the organization

In practice there are three organizational answers to the rush order, and the most common one is the worst.

Answer one: everything stays with the departments, the agent gets limits. Each department defines the limit for its own part. Each sets it conservatively, because it alone is blamed for mistakes and gains nothing from the overall optimum. The agent may then only do what the narrowest of the five limits allows. The rush order falls out completely, although it brings the company 6,680 euros, Fig. 5. The system runs, but it produces exactly the results gap everyone complains about. This is the normal case, and it is the reason why pilot projects work technically and deliver nothing economically.

Answer two: everything goes to IT. This fails for one single but sufficient reason. IT does not carry the economic consequence. Whoever does not carry the damage cannot be responsible for the limit.

Answer three: a named responsibility for the process as a whole, with its own responsibility for results. A function that owns the path from the request to the incoming payment, with a target system based on contribution margin, delivery reliability and capital tied up, instead of departmental metrics. This function sets the limits for the agent, escalates, and carries the economic consequence. It sits with the managing director or the COO. IT is responsible for something different and no less important: that the agent is traceable, can be switched off, and is logged.

The European AI Act follows the same logic. For high-risk systems, Article 26 requires that operators assign human oversight to natural persons who have the necessary competence, training and authority. Three properties, all of them required.⁸ Internal approval limits and capacity commitments in business-to-business trade are, by the wording, not in the high-risk class, so the rush order above is not a case for the regulation. The question of responsibility remains. It is then answered by company law and by the organizational duties of the management, and that applies regardless of any deadline.

The obvious solution is the most dangerous one

A common reflex is to put the agent on the org chart like an employee. It sounds tidy. A working paper by Emma Wiles and colleagues from Boston University and BCG, published in 2026, shows what this does. As soon as the same work no longer comes from a tool but from an “AI employee”, managers find fewer errors, feel less responsible, and ask more often for a review by a third party, Fig. 6.⁹

Whoever treats the agent like an employee shifts responsibility away from people in practice, although nothing shifts legally. The organization becomes less safe and slower at the same time, because more review loops are demanded.

The experiment that the Mittelstand has already run

On the question whether existing companies can rebuild their hierarchy when machines take over decisions, there is no long-term study. The only solid evidence comes from newly founded US startups. A Harvard study from June 2026 shows that AI-native startups are 25 percent smaller than comparable startups of the same industry and cohort, with half a hierarchy level less, at comparable valuations.¹⁰ But these companies never had an old structure. About rebuilding an existing one, the study says nothing.

But there is a precedent, and it sits inside every mid-market company. When IT handed over the operation of servers and applications to service providers, it lost the activity that had justified its existence. The execution was gone, the responsibility stayed. Hardly anyone changed the job description for that.

How this turns out is well documented, Fig. 7. Outsourcing research has found the same pattern since the nineties: those who outsource almost everything mostly fail, those who keep control in-house mostly do not. The most frequently named condition for success is the ability to manage service providers.¹¹ And at German companies with revenue of 100 million euros or more, according to Lünendonk, two thirds of the IT managers see managed services as highly important in the future, but only 11 percent have a position that manages the providers.¹²

Two thirds want to outsource more, more than half have no position that manages it. It is the same mistake as an agent without a named owner, only ten years older. Whoever outsources operations without building up the ability to manage them ends up with less control and not one cent less responsibility.

The ability that is missing in both cases is the same: set limits, detect deviations, escalate, carry the economic consequence. That is management, not technology.

Four questions for management

Whoever decides about the use of agents in their own company gets further with four questions than with any tool selection.

1. Which decision should the agent take? Not which process it automates. Which decision.
2. Who carries the economic damage today if this decision is wrong? That person is the candidate for the responsibility, regardless of where they sit on the org chart.
3. Does their target system reflect the decision? If the head of sales is measured on order intake, but the agent optimizes the contribution margin, then he is the wrong owner, no matter how well he fits professionally.
4. How does the agent recognize its limit, and what happens then? A limit that is not machine-readable does not exist.

Fig. 8 · Four questions before the first agent

1. Which decision should the agent take? Not which process it automates. Which decision.
2. Who carries the economic damage today if this decision is wrong? That person is the candidate for the responsibility, regardless of where they sit on the org chart.
3. Does their target system reflect the decision? If the head of sales is measured on order intake, but the agent optimizes the contribution margin, then he is the wrong owner, no matter how well he fits professionally.
4. How does the agent recognize its limit, and what happens then? A limit that is not machine-readable does not exist.

The third question is the most uncomfortable one. In the rush order from the beginning, three of the five departmental metrics get worse through the proposal: the material ratio in purchasing, delivery reliability in the plant, credit risk in finance. What gets better is only the total, and the order intake in sales, which does not own the total. There is nobody whose target system reflects the total. This is exactly the position that is missing, Fig. 9 shows it in numbers.

That these questions are rarely asked is shown by two surveys from this year. Only 21 percent of companies have a mature governance model for agentic AI, according to a Deloitte survey of 3,235 executives.¹³ And McKinsey finds among around 500 organizations that those with clearly assigned responsibility for AI reach a clearly higher maturity level than those without.¹⁴

An organization does not get better by buying better tools. It gets better when somebody owns the decision that the tools can now take. As long as this person does not exist on the org chart, every AI investment remains a cost center with better software.

Footnotes

ExpandCollapse

1. Luis Garicano, “Hierarchies and the Organization of Knowledge in Production”, Journal of Political Economy, vol. 108, no. 5, 2000, pp. 874 to 904. Source ↗ ↩
2. Bitkom, press release “Digitalisierung der Wirtschaft”, 11 March 2026, 604 companies with 20 or more employees, survey January to February 2026. Source ↗ ↩
3. Bitkom, same release, section on effects. Base: companies that already use AI. Source ↗ ↩
4. DIHK, Digitalisierungsumfrage 2026, survey 10 to 28 November 2025, 4,686 companies, p. 6. Source ↗ ↩
5. Klaus Wohlrabe, “Mehr KI, weniger Jobs? Was Unternehmen in Deutschland erwarten”, ifo Schnelldienst Digital 8/2025, 16 June 2025, table 2. Source ↗ ↩
6. Erik Brynjolfsson and Lorin M. Hitt, “Computing Productivity: Firm-Level Evidence”, Review of Economics and Statistics, vol. 85, no. 4, 2003, pp. 793 to 808. Source ↗ ↩
7. Erik Brynjolfsson, Daniel Rock and Chad Syverson, “The Productivity J-Curve”, American Economic Journal: Macroeconomics, vol. 13, no. 1, 2021, pp. 333 to 372. Source ↗ ↩
8. Regulation (EU) 2024/1689, Article 26(2). The Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, does not change this article. Source ↗ ↩
9. Emma Wiles, Hsu, Bedard and Kropp, “Putting AI on the Org Chart: Evidence on Delegation and Accountability”, working paper, Boston University and BCG, version of 19 September 2026. The effects apply to managers in organizations that already list AI agents on their org chart. Source ↗ ↩
10. Kim and Koning, “AI-Native Firms”, Harvard Business School Working Paper 26-090, 9 June 2026. The figure applies to the Y Combinator sample, in the wider sample it is 12 percent. Source ↗ ↩
11. Lacity, Khan and Willcocks, “A review of the IT outsourcing literature: Insights for practice”, Journal of Strategic Information Systems, vol. 18, no. 3, 2009, section 4.1, citing Lacity and Willcocks 1998. Source ↗ ↩
12. Lünendonk, “IT-Sourcing-Trends 2025/2026”, 121 user companies with revenue of 100 million euros or more, a quarter of them above 1 billion euros, interviews March to May 2025, p. 22 and figure 26. Source ↗ ↩
13. Deloitte, “Business and IT leaders report AI agents are scaling faster than their guardrails”, 24 April 2026, 3,235 respondents in 24 countries. Source ↗ ↩
14. McKinsey, “State of AI trust in 2026: Shifting to the agentic era”, 25 March 2026, around 500 organizations. Maturity score 2.6 with clear ownership versus 1.8 without. Source ↗ ↩

About the author

Operations consultant · Mid-market companies · formerly COO

Christopher von Könemann advises mid-market companies on improving results and working capital and on setting up steering systems. He was previously COO of a mid-market group of companies.

LinkedIn · Contact

Share on LinkedIn