Plamen MILTENOFF

Explaining Andreas Horn’s main argument in simple terms

An analysis of Andreas Horn’s LinkedIn post, the accompanying discussion, and the implications for universities and businesses.
Horn makes a significant observation: AI agents have become capable of completing complex tasks with less human intervention, but many organizations are not prepared to work with them.
An AI agent differs from a conventional chatbot. Rather than answering one question and waiting for another instruction, an agent can carry out several connected activities, use software tools, identify problems, and attempt to correct errors.
For example, an AI agent might receive the following assignment:
“Collect the monthly sales data, check for missing values, analyze the results, prepare charts, and draft a management report.”
The agent might complete these steps with relatively little human supervision.
Horn reports that he increasingly gives agents tasks he expects them to fail, yet they often finish successfully. He also describes agents working through extended tasks overnight.
These are Horn’s personal observations, rather than evidence from a systematic evaluation of agent performance.

The central argument: The organizational process becomes the limitation

Consider a company where preparing a sales report requires five stages of approval. An AI agent finishes the report in ten minutes, but employees still need several days to approve and circulate the results.
The AI performs efficiently. The organization continues to operate according to procedures designed for a different working environment.
Horn argues that improving the AI model will not solve this problem.
Organizations need to reconsider how work is organized, who makes decisions, and which activities genuinely require human supervision.
Illustrative example of Horn’s argument

Existing workflow	Redesigned workflow
AI prepares report	AI prepares report
↓	↓
Manager approval	Automated quality checks
↓	↓
Department approval	Low risk: release
↓	↓
Final authorization	High risk: human approval

Conceptual comparison, not a workflow documented in Horn’s post. Necessary compliance and authorization requirements must remain in place.

Horn’s recommendation is to move from reviewing every activity to reviewing those decisions where human judgment matters.

The concept behind his argument is the target operating model, meaning how an organization distributes tasks, responsibilities, decisions, information, and accountability among people and technology.

His central message: Technological capability is advancing faster than organizations are adapting their processes.

But the comments reveal that this argument needs important qualifications.

2. The LinkedIn comments into themes

The supplied discussion contains agreement, practical implementation experiences, technical suggestions, and substantial criticism. Several commenters question whether faster and more autonomous AI agents are necessarily desirable.
The following eight themes capture the main arguments. They are thematic categories rather than a statistical classification of every comment, and individual comments sometimes address multiple themes.

Theme 1. Redesigning processes rather than automating existing inefficiencies

Many participants agree with Horn that organizations often automate fragmented workflows without improving them.

John Reuben, Alexandre R., and Marcel Zimmermann argue that leadership must rethink processes, cross-functional coordination, and organizational structures. Alexandre goes further, questioning why businesses should retain procedures originally designed around human limitations.

Main takeaway: Before automating a task, examine whether the task and its associated procedures are still necessary.

Theme 2. Governance, decision rights, and human authority

This is one of the strongest themes and provides an important correction to Horn’s argument.

Lynn Räbsamen points out that approvals in regulated industries frequently exist for valid reasons. They protect customers, manage financial exposure, and ensure compliance.

Wayne Couch, Kemi L., Wolfgang Reck, and Paul Souhuwat emphasize that organizations need explicit boundaries defining which decisions agents can make and which decisions require humans.

Main takeaway: Removing unnecessary bureaucracy is desirable. Removing necessary governance is dangerous.

Theme 3. Evidence, verification, and reliability

Alexander Strüver presents one of the strongest criticisms. He argues that generative AI produces statistically inferred outputs, which do not automatically qualify as verified evidence.

Kent Cheong explains that auditors need records showing what an agent checked, which evidence supported its conclusions, and what remained unresolved.

Arsen Doda advocates keeping a verifiable record of each completed step, while Zak Gilyard asks how organizations can detect steps an agent silently skips.

Main takeaway: Successful execution is not proof of correctness. Agents need independent verification and auditable records.

Theme 4. Data quality, organizational knowledge, and technical infrastructure

Oleg Shilovitsky introduces the idea of an engineering harness: the operational environment that supplies agents with tools, context, data, workflows, and safeguards.

Leon Schelhase notes that spreadsheets often contain undocumented assumptions. Employees know which data are outdated or unreliable, but an AI agent cannot infer those informal rules consistently.

Main takeaway: Agents require trusted data sources, explicit rules, appropriate tools, and clearly documented organizational knowledge.

Theme 5. Clear objectives, measurable outputs, and completion criteria

Marco Gundlach notes that many organizations have never defined precisely what a completed task should look like.

Darshan Kumar Satpathy proposes machine-checkable completion criteria, which means expressing success as conditions a system can test.

DeeplyAgentic makes a related point about specifying acceptance criteria before allowing agents to operate independently.

Main takeaway: An agent should not receive full responsibility for a process when nobody can define or verify successful completion.

Theme 6. Risk-based autonomy and human checkpoints

Mujahid Hussain proposes a practical distinction: reversible actions can receive greater autonomy, while actions involving financial transactions, customers, or permanent changes require human approval.

Noor Surani suggests linking supervision thresholds to measurable risks, including financial exposure and compliance requirements.

Main takeaway: The level of human supervision should depend on consequences, reversibility, and the potential cost of errors.

Theme 7. Leadership, institutional learning, and workforce preparation

Alex Kuzmenko argues that organizations should treat AI pilots as learning environments where teams clarify responsibilities and practise managing exceptions.

Shashi Shekhar warns that some existing procedures preserve important informal knowledge.

Jack Rudenko describes a situation where individual engineers improved productivity with AI, yet their independently developed approaches created problems for the wider engineering team.

Main takeaway: Individual AI competence does not automatically produce institutional AI competence. Coordination, training, and leadership remain essential.

Theme 8. Business value, operating costs, and skepticism about AI-driven redesign

Noah Salas emphasizes the cost of agents repeating work while waiting for approvals. Elle Falls calls for operational monitoring to detect failures and assign responsibility.

Alejandro Estrella Gabilondo disputes the assumption that organizations must redesign themselves around AI. Drawing on experiences with enterprise resource planning, he argues that technology should adapt to organizational needs rather than automatically dictate new processes.

Main takeaway: The business case must demonstrate improved outcomes, not simply faster task execution or greater automation.

3. Where the commenters agree and disagree

Question	Main position in the discussion
Are AI agents becoming more capable?	Many report improvement from practical experience.
Should organizations redesign processes?	Broad agreement, but disagreement about how much change is necessary.
Should human approvals be reduced?	Yes, when redundant. Essential controls must remain.
Can successful agent execution be trusted automatically?	No. Verification, evidence, and accountability are required.
Should agents make consequential decisions?	Their authority should be explicitly limited and governed.
Is technology the only major obstacle?	No. Participants also identify data, governance, coordination, skills, and organizational knowledge.

The most important disagreement concerns the relationship between efficiency and institutional responsibility.

Horn emphasizes the limitations imposed by existing processes. His critics emphasize that some limitations exist to protect legitimate interests.

These positions can be reconciled by distinguishing three types of organizational requirements: unnecessary administrative friction, necessary human judgment, and controls that can be implemented through reliable technical mechanisms.

4. Main takeaways for academia

Horn’s argument has implications for universities, although the original post primarily addresses businesses.

Universities operate through academic departments, administrative units, research offices, libraries, quality assurance systems, and regulatory structures. Many processes require coordination among multiple units.

Agentic AI creates opportunities to improve this coordination. It also raises questions about academic judgment, educational quality, and institutional accountability.

Six implications for universities

1. Redesign academic workflows rather than adding AI to existing tasks

A university research office might use an agent to identify funding opportunities, collect application requirements, check deadlines, and prepare documentation.

But if the proposal still moves unnecessarily between several administrative units, the university gains limited efficiency.

Practical action: Map an existing workflow, identify unnecessary handoffs, and redesign the process before introducing an agent.

2. Protect academic judgment and research integrity

AI agents can support literature searching, reference checking, data preparation, coding, and routine quality checks. They cannot independently establish the validity of research claims.

Researchers must retain responsibility for research questions, methodology, interpretation, ethical decisions, and publication conclusions.

Practical action: Require traceable sources, reproducible analyses, and documented human verification of consequential research outputs.

3. Reconsider assessment in an agentic AI environment

When an agent can independently search sources, generate code, analyze data, and prepare a report, traditional assignments may no longer measure the intellectual work instructors intend to assess.

Practical action: Assess students’ research decisions, source evaluation, methodological reasoning, interpretation, and ability to explain or defend their results. Documented AI workflows can provide additional evidence of learning.

4. Move from individual AI literacy to institutional AI fluency

Training faculty and students to write prompts is insufficient when AI systems begin managing entire workflows.

Faculty, researchers, and administrators also need to understand delegation, agent supervision, verification, data access, and accountability.

Practical action: Introduce training in AI workflow design and governance alongside existing AI literacy programs.

5. Establish clear institutional authority over AI agents

An agent may assist with course scheduling, doctoral administration, and routine documentation, but decisions involving admissions, grading, academic misconduct, research ethics, and students’ rights require formal governance.

Practical action: Specify which activities agents may perform, which require human authorization, and who is accountable for errors and corrective action.

6. Treat AI implementation as an institutional research opportunity

Universities can examine how agentic AI changes organizational effectiveness, academic workload, research quality, and educational outcomes.

Practical action: Evaluate pilots using measures such as completion time, error rates, verification effort, academic quality, staff workload, and stakeholder satisfaction.

Example: Agentic AI in the research process

A university might develop an AI-supported research workflow with the following division of responsibilities.

Activity	Potential AI role	Human responsibility
Literature search	Find and organize candidate sources	Assess relevance, coverage, and quality
Reference management	Extract metadata and identify duplicates	Validate sources and citations
Data preparation	Clean data and document transformations	Approve exclusions and analytical decisions
Statistical analysis	Execute approved R or Python code	Verify assumptions, results, and interpretation
Manuscript preparation	Assist with drafting and formatting	Establish and defend scientific claims
Research ethics	Organize documentation	Review and authorize ethical decisions

The central principle for academia: the ability to automate an academic activity does not establish universities should delegate responsibility for its intellectual or ethical consequences.

5. Main takeaways for business

The business implications are more directly aligned with Horn’s original argument.

For businesses, the opportunity involves improved process execution, reduced administrative delays, and more efficient allocation of employees’ time. The principal challenge is ensuring that autonomous execution produces reliable business outcomes.

Six priorities for business leaders

Priority	Recommended action	Expected benefit
1. Process mapping	Document real workflows, including exceptions and informal procedures	Identify where automation offers value
2. Clear ownership	Assign responsibility for every agent and its outputs	Reduce organizational ambiguity
3. Risk-based delegation	Define the actions agents can execute independently	Balance efficiency with control
4. Technical infrastructure	Establish reliable data, approved tools, and transaction logs	Improve reliability and traceability
5. Verification	Introduce measurable completion criteria and evidence checks	Detect errors and incomplete work
6. Performance evaluation	Track cost per verified task, quality, error rates, and human rework	Establish whether automation generates business value

Example: Customer service

Consider an agent responsible for handling customer refund requests.

The agent can retrieve the order, verify the company’s refund policy, and prepare the required documentation. Under predefined conditions, it may issue a small, reversible refund.

When a request involves fraud indicators, a disputed transaction, an exceptional amount, or uncertain eligibility, the agent must escalate the decision.

The operational improvement comes from giving the agent sufficient authority to complete routine work while preserving human control over consequential decisions.

This is a more useful business objective than maximizing the number of tasks completed without human involvement.

6. A practical framework for deciding what AI agents should do

The debate suggests assessing each activity according to its risk, complexity, reversibility, and evidential requirements.

AI Delegation Decision Guide

Select a type of work to see the recommended division of responsibility between the AI agent and the human decision-maker.

Routine and reversible work

Recommended AI autonomy: HIGH

These tasks have clear rules, low consequences, and outputs that can be corrected easily.

AI agent may:

• Execute the task independently
• Use approved tools and data sources
• Record actions and results
• Escalate exceptions

Human responsibility:

• Define the task and acceptance criteria
• Set boundaries and permissions
• Review exceptions rather than every action
• Periodically audit performance

Examples: document formatting, meeting summaries, data classification, routine scheduling, standard information retrieval.

Recommended model: AI executes. Humans monitor exceptions.

Analytical and knowledge-intensive work

Recommended AI autonomy: MODERATE

The agent can perform substantial work, but interpretation and validation remain important.

AI agent may:

• Search and organize information
• Run approved analyses
• Generate alternative explanations
• Prepare preliminary reports
• Document sources and analytical steps

Human responsibility:

• Define the research or business question
• Validate data and assumptions
• Evaluate evidence
• Interpret results
• Approve consequential conclusions

Examples: literature searching, market analysis, statistical analysis, forecasting, research synthesis.

Recommended model: AI performs analysis. Humans verify evidence and interpret results.

Work involving sensitive data

Recommended AI autonomy: LIMITED

The primary concern is access to confidential, personal, proprietary, or regulated information.

AI agent may:

• Operate only within approved systems
• Access only authorized data
• Log data access and transformations
• Apply predefined rules

Human responsibility:

• Approve data access
• Define privacy and security controls
• Review unusual access patterns
• Authorize external disclosure or transfer
• Maintain accountability for compliance

Examples: student records, customer data, employee records, financial information, unpublished research data.

Recommended model: constrained AI execution under explicit access, privacy, and audit controls.

Consequential decisions

Recommended AI autonomy: LOW

The agent may support the decision, but the outcome can significantly affect people, money, rights, reputation, or institutional obligations.

AI agent may:

• Collect relevant evidence
• Check predefined criteria
• Identify anomalies and risks
• Prepare recommendations
• Explain the basis for the recommendation

Human responsibility:

• Evaluate competing considerations
• Review supporting evidence
• Exercise professional judgment
• Authorize the final decision
• Remain accountable for the outcome

Examples: hiring, admissions, grading disputes, credit decisions, disciplinary cases, major financial commitments.

Recommended model: AI advises. Humans decide.

High-risk or irreversible actions

Recommended AI autonomy: MINIMAL

Errors may produce serious, permanent, legal, financial, safety, or ethical consequences.

AI agent may:

• Provide information and recommendations
• Run simulations or checks
• Identify possible risks
• Document evidence
• Stop and escalate when thresholds are reached

Human responsibility:

• Review evidence independently
• Confirm legal and institutional authority
• Approve or reject the proposed action
• Document the reasoning
• Accept responsibility for the decision

Examples: large financial transfers, legally binding commitments, research ethics decisions, termination decisions, safety-critical actions.

Recommended model: AI supports the process. Humans retain decision authority.

Core principle: Increase AI autonomy only when the task is clearly defined, independently verifiable, sufficiently reversible, and governed by explicit decision and escalation rules.

This framework is a synthesis of the comments, not a framework formally proposed or empirically validated by Horn.

7. Overall assessment: What the discussion reveals

Horn’s post identifies a consequential shift: organizations increasingly need to decide how work should function when AI agents can execute connected activities autonomously.

The comments deepen his argument in three ways.

First, process redesign alone is insufficient. Reliable agentic systems need trusted data, technical safeguards, documented evidence, and clearly allocated authority.

Second, existing human procedures are not necessarily inefficient. Some preserve institutional knowledge, protect rights, or provide essential accountability. Organizations should examine their purpose before replacing them.

Third, agentic AI implementation is as much a leadership and organizational design challenge as a technical project.

For universities, this connects directly to the broader discussion about hybrid intelligence and the transition from AI assistance to AI participation in teaching, research, and administration.

For businesses, the priority is moving from isolated automation experiments to governed workflows with measurable outcomes.

The strongest shared principle emerging from the discussion is:

AI agents may execute increasingly complex work, but organizations must determine the conditions under which they act, the evidence required to trust their outputs, and the humans who retain decision authority and accountability.

This leads to a question particularly relevant to university governance and business strategy:

If AI agents can perform much of the work currently assigned to people, which decisions should remain human, and how should we redesign our institutions to preserve those decisions?

Earlier posts in this blog asked which educational decisions should remain human, how agentic AI changes the division of work between people and machines, and why universities are moving from experimentation toward institutional decision-making. The September overview framed this as designing the human-AI relationship, while the October overview pushed the discussion toward evidence of learning and institutional responsibility. The current post extends that progression from education into organizational design more broadly: once AI agents can act, the key issue becomes how institutions define authority, verification, accountability, and the limits of delegation.