Who Acts, Who Decides: Four Patterns for Human–AI Work

TL;DR for operators

AI autonomy and AI initiative are different design decisions. An AI system can execute work independently while waiting for instructions, or proactively offer help while remaining limited to a narrow task.

Michael Weiss identifies four collaboration patterns: Instruction, Delegation, Assistance, and Co-Creation. Each suits a different kind of work, and each introduces a recurring tension. Delegation can increase capacity while reducing oversight; Assistance can improve individual components while weakening their fit with the whole task.

For enterprise workflow designers, the framework suggests specifying not only what AI may do, but also when humans must clarify requirements, verify results, reintegrate outputs, or reaffirm ownership. These are conceptual design heuristics, not experimentally established performance improvements.

One report can require four different AI relationships

Consider a business analyst preparing a report with AI.

Initially, the analyst may struggle to articulate the question. The AI’s responses help reveal what the analyst actually wants to investigate. Once the question becomes clear, the analyst may ask AI to collect information independently. While writing, the analyst might welcome proactive suggestions for improving a chart. When interpreting the results, the analyst may want AI to challenge assumptions and propose alternative explanations.

The same AI system could support all four activities, but the appropriate distribution of initiative and responsibility changes.

Granting more autonomy does not necessarily improve the arrangement. Independent execution may be desirable when requirements are stable, yet inappropriate when the analyst is still determining the task. Proactive assistance may be valuable even when AI should have little authority over the broader work.

In Human-AI Collaboration: From Paradoxes to Patterns, Michael Weiss develops a framework for distinguishing these situations and identifying the tensions each creates.¹

The central contribution is a classification of collaboration arrangements based on two independent dimensions, followed by an account of why none permanently resolves the competing demands of human-AI work.

Autonomy and initiative define four collaboration patterns

Autonomy describes how independently AI performs assigned work. Initiative describes whether AI waits for human instructions or initiates relevant actions.

Crossing low and high levels of these dimensions produces four patterns, presented in Section 3, Figure 1 and Table 1.

These configurations are not successive levels of AI maturity.

Assistance, for example, combines high initiative with low autonomy. An AI-enabled editor might proactively suggest a revision while the human retains control over the document. The system does not require authority to complete the entire writing task simply because it initiated a suggestion.

Delegation reverses that combination. A human assigns a defined research task, and AI performs it independently without deciding when the broader project should begin.

The operational choice therefore concerns two permissions: what AI can execute independently and what AI can initiate without being asked.

Each collaboration pattern creates a tension that must be managed

Weiss draws on paradox theory to explain why selecting a suitable collaboration pattern does not eliminate its associated risks.

A paradoxical tension exists when two desirable properties conflict but remain interdependent. Increasing one may create consequences that eventually require renewed attention to the other.

The paper describes these recurring adjustments as circular responses.

Instruction: clarify requirements without closing exploration too early

Instruction applies when the human understands something about the task but cannot yet express it completely.

Iterative prompting helps externalize that tacit knowledge. The human probes AI, evaluates its responses, and progressively refines the instructions.

Greater structure makes the task easier to specify. However, structure can also narrow attention, reducing sensitivity to unexpected interpretations or behavior.

The proposed response alternates between probing and reflection. Exploration reveals information that a premature specification might exclude; reflection converts useful discoveries into clearer requirements.

Delegation: increase capacity while preserving authority

Delegation transfers execution of a self-contained task to AI.

The benefit is additional work capacity. The corresponding concern is reduced human visibility into execution and possible disengagement from the results.

Weiss frames this as a tension between centralized and decentralized control.

The pattern proposes selective verification and reclaiming authority when outputs depart from expectations. Delegation therefore does not imply eliminating human oversight. It changes how and when oversight is exercised.

Assistance: improve components without losing task coherence

Assistance isolates a supporting activity within a larger human task.

AI may efficiently improve an individual passage, calculation, or component. Yet a locally satisfactory result may be poorly aligned with the purpose of the whole task.

The proposed response is reintegration: the human evaluates the AI-supported component in its broader context and adjusts it when necessary.

This is a different intervention from checking whether an independently delegated task was completed correctly. Component quality and whole-task coherence are distinct evaluation targets.

Co-Creation: welcome alternative perspectives while retaining ownership

Co-Creation gives AI both substantial autonomy and initiative in an activity requiring multiple perspectives.

AI can contribute interpretations that the human had not considered. But greater incorporation of AI perspectives may weaken the human participant’s sense of authorship or ownership.

The proposed response combines inviting external perspectives with reasserting human judgment, voice, and intent.

In this framework, active AI participation does not remove the human’s responsibility for deciding which contributions become part of the work.

The eight-step method turns tensions into design patterns

The paper’s second contribution is a procedure for constructing patterns from recurring tensions, rather than treating the four categories as an exhaustive collection of ready-made answers.

Section 4.3 and Table 2 demonstrate eight steps using Instruction:

1. Scenario: Identify the collaboration arrangement.
2. Context: Specify the task conditions.
3. Paradoxical tension: Name the opposing but interdependent forces.
4. Problem statement: Express the conflict those forces create.
5. Forces: Describe why each side matters and how they interact.
6. Solution: Propose actions that navigate the conflict.
7. Resulting context: Explain the consequences of those actions.
8. Nested paradoxes: Identify further tensions that may emerge.

The Instruction example begins with tacit task knowledge. Making requirements explicit requires structure, but additional structure may restrict sensitivity to unanticipated possibilities.

The proposed solution—repeated probing and reflection—responds to both demands without assuming either can permanently dominate.

This procedure matters because it connects a design recommendation to the underlying reason for that recommendation. A pattern description identifies its context, problem, solution, and consequences, allowing designers to examine whether the proposed interaction fits their own circumstances.

The paper provides a conceptual derivation method, not an optimization algorithm or a trained policy for automatically allocating decision rights.

Clarifying requirements can change the appropriate collaboration mode

The four patterns need not remain separate throughout a project.

Instruction may help a domain expert articulate previously tacit knowledge. As requirements become more explicit, the task may become suitable for Delegation, Assistance, or Co-Creation.

The paper illustrates this progression through a fictitious three-prompt conversation in Table 3.

A user initially requests help writing a pattern. The meaning of pattern remains ambiguous. The user then clarifies that it means a recurring solution to a problem in context, before asking AI to collect the necessary information through questions.

The example shows how successive interactions can clarify both terminology and the process required to perform a task. The author reports informal testing with a memoryless Perplexity configuration, but the conversation is explicitly illustrative, not an empirical validation.

For business processes, the transition suggests a possible staged approach: use exploratory interaction to establish task requirements before assigning independent execution.

However, a transition also introduces new tensions. Better task specification can make delegation possible while creating a fresh need to balance execution autonomy against oversight.

Enterprise design should specify intervention points

Cognaptus interpretation: The framework offers a way to structure workflow-design discussions around distinct human and AI responsibilities.

For a reporting team, that could mean using Instruction while defining analytical requirements, Delegation for bounded information-gathering tasks, Assistance for contextual suggestions, and Co-Creation when exploring competing interpretations.

The resulting workflow would need to specify which activities AI may initiate, which outputs require verification, how component changes are evaluated against the complete report, and who approves the resulting interpretation.

This goes beyond choosing a collaboration label. It requires mechanisms that support the complementary human action associated with each pattern.

A team introducing Delegation, for instance, could establish review checkpoints and conditions for returning execution authority to the analyst. A team introducing Assistance could require proposed edits to be evaluated within the surrounding analysis before acceptance.

These are design implications inferred from the paper’s framework, not controls whose effectiveness the study has measured.

The appropriate verification depth, intervention frequency, and allocation of responsibility would still require task-specific evaluation.

The framework identifies design tensions, not superior outcomes

The evidence boundary is substantial.

Weiss develops four conceptual patterns, explicit paradox mappings, an eight-step derivation procedure, and illustrative human-agent-tool workflows. The paper does not conduct controlled experiments comparing collaboration arrangements.

Consequently, it does not establish that pattern-guided design improves accuracy, productivity, creativity, or collaboration quality. Nor does it show that the proposed responses reliably mitigate their associated tensions in deployed organizations.

The framework also assumes domain-expert humans who can evaluate AI outputs and have already decided to use AI. It does not address whether AI should be adopted for a particular task, systematically analyze AI-directed human workflows, or establish applicability to novice users.

Broader testing of the derivation procedure and formal representations of paradoxes remain future research directions.

For organizations, that makes the taxonomy suitable for generating and documenting design hypotheses. Demonstrating effectiveness would require comparing candidate arrangements under explicit task objectives, failure conditions, and accountability requirements.

Collaboration design does not end when autonomy is assigned

The paper’s contribution is to treat human-AI collaboration as a continuing allocation of action and judgment.

Autonomy determines how independently AI works. Initiative determines whether it begins relevant work. Their combinations create distinct collaboration arrangements, each with competing demands that cannot be permanently settled by selecting a category.

The framework gives enterprise designers a vocabulary for those demands and a process for converting them into reusable interaction patterns.

Its strongest application today is as a structured way to decide where human reflection, verification, integration, and ownership should remain visible in AI-enabled work. Whether these arrangements outperform alternative designs is a question for subsequent empirical testing.

Cognaptus: Automate the Present, Incubate the Future.