The price of delegation

Companies address it through incentives, reports, controls and binding commitments. These measures themselves consume resources and do not prevent every deviation. Agentic AI opens up possibilities for making information more accessible, tracking activities and verifying outcomes. Its effects, however, depend on how it is designed and used.

This article's hypothesis is that agentic AI can reduce some of the costs of human delegation. At the same time, decision-making power over its objectives, data and evaluation criteria becomes more important. An economically better organisation must manage both levels: the execution of delegated work and the design of the arrangements that govern it.

Keywords: principal–agent problem; agency costs; delegation; incentives; information asymmetry; agentic AI; corporate management.

The example is hypothetical. It illustrates a possible conflict between objectives: a decision that makes sense from the sales team's perspective can leave the company as a whole worse off. At the same time, sales knows more about the particular negotiation than company management does. Management has only limited ability to assess how much of a discount was actually necessary.

An additional report does not necessarily resolve this conflict. It can make the deviation visible. It can also document the same incomplete definition of success in greater detail. What matters is what is commissioned and rewarded, and the basis on which the outcome can be assessed.

The relationship becomes problematic when interests diverge and behaviour or information can only be verified to a limited extent. Jensen and Meckling describe the resulting costs of incentives and safeguards. Jensen and Meckling, 1976, section 1.4

Within a company, the same manager can be an agent in relation to senior management and a principal in relation to their team. The two roles can create different demands.

The term does not imply bad intent in any particular case. A conflict between objectives may already be built into remuneration, budget responsibility or performance measurement. Equally, a deviation may result from missing knowledge or an error of professional judgement. A disappointing outcome alone does not establish opportunistic behaviour.

This classification follows Jensen and Meckling, 1976.

An important limit applies in practical evaluation: this does not make every meeting, handover or management task an agency cost. Specialist collaboration may be necessary even when everyone involved pursues the same objective. A late outcome may be caused by insufficient capacity. A lower margin may result from market changes.

A sound analysis must therefore establish which effort serves coordination, which limits a divergence of interests and which loss is actually attributable to such a divergence. Otherwise, the term becomes a blanket label for unwanted administration.

In the sales example, adding a margin metric would be one possible step. However, it would need to specify which costs are allocated and when the assessment takes place. If additional services provided later are excluded, the reported margin may still give an incomplete picture.

The organisational task therefore goes beyond selecting another number. The assignment, decision-making discretion and outcome assessment must fit together. Whoever makes a commitment must be able to consider its relevant consequences. Whoever verifies fulfilment needs access to suitable information.

Judgement remains necessary. A weak short-term margin may be part of a deliberately chosen strategic order. A control that treats every deviation in the same way can prevent such decisions. What matters is whether the exception falls within a clearly explained mandate.

The finding comes from a controlled experimental setting. It does not justify a blanket conclusion against operational controls. It does, however, support assessing their effects on behaviour and collaboration as well.

For a company, this raises a practical question: does a control help address a relevant risk or a conflict between objectives, and what additional work does it create? A justified spot check, clearly defined spending authority and exhaustive monitoring of individuals have different consequences. Their value depends on the task, potential harm and what can actually be verified.

Appropriate discretion can enable expertise to be put to use. It requires understandable limits and a way to disclose uncertainty. Employees must be able to raise a conflict between objectives without the act of reporting it already being treated as a personal failure.

In a suitable setting, an agentic system can combine information from accessible sources, track outstanding commitments and present contradictory information for review. It can also undertake specialist work and coordination within a mandate. This requires the necessary capabilities, access and reliable feedback.

In the sales example, a system could check additional commitments against the cost calculation, flag missing cost assumptions and obtain a necessary decision. After delivery, actual performance could be compared with the original commitment. Information would thus become usable throughout the order.

The potential gain lies in smaller information gaps and less manual follow-up. A technical log alone, however, proves neither that observation is complete nor that the outcome is correct. An unrecorded side agreement remains invisible even in a well-documented case.

The decisive question is therefore: which previously hard-to-observe relationship actually becomes easier to verify, and at what additional cost?

Decision-making power may therefore become concentrated among those who design rules, data access and evaluation criteria. These may be internal departments, operators, integrators or external providers. Their interests need not align with the principal's in every respect. An agentic organisation therefore also needs transparency and verifiable mandates for these relationships.

It is not enough to attribute neutrality automatically to an AI system because it has no personal career interests. A study by Sharma and co-authors describes a tendency among five AI assistants available at the time to align with users' beliefs in certain tasks. In some cases, human preference ratings favoured persuasively expressed agreement over correct answers. Presented in 2023, the study examines model behaviour; it establishes neither human motives in the systems nor the performance of all models available today. Sharma et al., 2023

The resulting organisational test is specific: an automatically generated report must also be checked against facts that could disprove its account. The principal's agreement is not a sufficient measure of quality.

At the same time, technical errors and economic conflicts of interest must be distinguished. An incorrect answer may reflect a capability limit. An evaluation criterion deliberately chosen to be incomplete, by contrast, may advance its designers' interests. Both problems require scrutiny, but they require different corrections.

The NIST AI Risk Management Framework 1.0, published in 2023, identifies clear roles, communication channels and accountability for AI risks as matters to address in design. It also calls for explicit allocation of roles in human–AI collaboration. This supports the organisational importance of such allocations; it is not empirical evidence of lower delegation costs. NIST, 2023, GOVERN 2 and 3.2

The hypothesis advanced here is therefore that agentic AI changes the conditions of delegation. Economic progress occurs when better observation and execution of work create more value than the new design, verification and control of the systems cost.

People also act out of responsibility and trust

The principal–agent model captures selected relationships between interests, information and decisions. It does not describe every motivation behind human work. Professional standards, loyalty and collaboration can align objectives without securing every action through an additional incentive.

An implementation that treats employees fundamentally as a source of errors to be monitored can lose important knowledge and voluntary contributions. A good diagnosis must therefore distinguish a documented conflict between objectives, a lack of resources and mere speculation about personal motives.

The principal can set an unsuitable objective

A decision is not necessarily good for the company or those affected simply because it fulfils a stated assignment. Company leadership can also set incomplete objectives or overlook consequences. Several legitimate requirements may conflict.

A system that pursues an unsuitable objective particularly consistently may make the problem worse. Assessment must therefore also address the quality and permissibility of the assignment. Operational adherence to the mandate does not replace this design responsibility.

Greater transparency can turn into continuous surveillance

The observation needed should concern the outcome, the basis for decisions and relevant deviations. A desire for better management does not create a substantive need to record every activity of every employee in full.

The information required must be justified by the service and its risks. Understandable access rights, opportunities to challenge assessments and the ability to correct erroneous evaluations also matter. Extensive data collection can contribute little to resolving the actual conflict between objectives despite its high cost.

A simpler solution may be more economical

Sometimes clearer responsibilities, a different incentive or a reliable reconciliation of data is enough. Additional agentic management may then be unnecessary.

The comparison must therefore include these alternatives as well. Reduced manual control work only counts when considered alongside the effort of setting up, operating, verifying, correcting and maintaining the new solution. Moving work into technology is not yet an economic gain.

Relevant outcome conditions and possible side effects must be defined before a change. A comparison should cover similar cases and periods while accounting for changes in demand, prices and order mix.

Remaining losses of value can often only be estimated using assumptions. They should be reported separately from observed time and cost expenditure. Expressing them in a common currency does not make evidence of different quality equally robust.

The hypothesis is supported if, under comparable conditions, a design supported by agentic AI reduces relevant deviations from objectives and the effort spent on safeguards, while its additional costs and side effects remain smaller than the gains. It is weakened if information gaps persist, errors become harder to detect or a simpler organisational change achieves the same effect at lower cost.

Agentic AI can change the observation and execution of delegated work. At the same time, it makes the design of its own governing rules a central leadership task. The price of delegation can fall if these changes are considered together and tested against actual outcomes.

The key question is: Can we better assess whose interests a decision serves, what it is based on and whether its outcome fulfils the agreed purpose?

Jadey's service-oriented organisational model connects delegated work with agreed outcomes, clear operational accountability and verifiable fulfilment. A service agent leads a service it has accepted, including the contributions required, their verification and any necessary rework. If it commissions further services, it remains accountable for using those contributions in its own outcome.

The economic hypothesis behind this approach is a reduction in information gaps, manual follow-up and avoidable deviations from objectives. It must be tested in each setting against the full effort involved and suitable alternatives. The sources discussed here do not provide blanket evidence of Jadey's effectiveness.

The existing Insight “Management without self-interest” examines the possibility of coordination without personal career or status interests. It leaves accountability for outcomes with people. The current autonomous organisational model additionally assigns service agents operational accountability for outcomes: the obligation to actually lead services they have accepted and verify fulfilment. Management services covered by appropriate mandates may also be included.

This further-reaching model specification does not transfer legal liability or decision rights reserved for people. Designing and reviewing mandates remains part of the organisational task. For precisely this reason, agentic accountability for outcomes must also prove itself through outcomes that can be understood and corrected.

These sources examine different subjects. They establish neither the fundamental superiority of human or agentic management nor the complete elimination of conflicts of interest. Their application to an agentic company organisation is the author's own testable hypothesis. The sales example is hypothetical. Concepts and evaluation criteria do not replace a diagnosis of the actual organisation.

1. Jensen, M. C.; Meckling, W. H. (1976): Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure. Journal of Financial Economics, 3(4), 305–360. Used: section 1.4; a publicly accessible reproduction of the original publication.
2. Holmström, B.; Milgrom, P. (1991): Multitask Principal–Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design. Journal of Law, Economics, and Organization, 7, Special Issue, 24–52. Used: introduction of the version provided by the authors' university; no independent review of the subsequent mathematical formalism.
3. Falk, A.; Kosfeld, M. (2006): The Hidden Costs of Control. American Economic Review, 96(5), 1611–1630. Used: abstract of the journal version covering the experimental design and main finding.
4. Sharma, M. et al. (2023): Towards Understanding Sycophancy in Language Models. Used: Anthropic's research summary of 23 October 2023. Provider research on models available at the time; not a current market comparison.
5. National Institute of Standards and Technology (2023): Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. Used: the description of the voluntary framework, and GOVERN 2 and 3.2. A design framework, not an impact study.