Do we have the knowledge we need? Rethinking human-AI decision-making in corporations

Anne S. R. Marx; Mennatallah El-Assady; Ricardo M. Avelino; Torbj{\o}rn Netland

REVIEW 2 major objections 5 minor 20 references

Safe AI autonomy in companies is limited by knowledge grounding, not compute; agency over decisions should be allocated dynamically from knowledge types and problem attributes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 14:05 UTC pith:622NXZXQ

load-bearing objection Clean synthesis of agency and knowledge for manufacturing AI; useful vocabulary, untested mapping, standard position-paper limits. the 2 major comments →

arxiv 2606.15575 v1 pith:622NXZXQ submitted 2026-04-02 cs.AI cs.HC

Do we have the knowledge we need? Rethinking human-AI decision-making in corporations

Anne S. R. Marx , Ricardo M. Avelino , Torbj{\o}rn Netland , Mennatallah El-Assady This is my paper

classification cs.AI cs.HC

keywords mixed-initiative systemshuman-AI collaborationagencyautonomyknowledge groundingTDE-loopmanufacturingdecision-making

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

As companies hand more decisions to AI, the real bottleneck is not model power but whether the organization has made the right kinds of knowledge accessible to both people and machines. This position paper argues that every action that changes a company world state can be cast as a TDE-loop—trigger, decision, optional control, execution—and that an orchestrator should assign user, mixed, or system agency at each stage according to how well available knowledge lines up with the problem’s regularity, consequences, goal clarity, and solution-space size. It separates descriptive knowledge (historical, semantic, predictive) from prescriptive knowledge (normative, procedural) and shows, via two manufacturing cases, when full automation is safe and when human control must stay in the loop. The practical upshot is a shift from efficiency-first automation toward intentional knowledge capture and dynamic agency negotiation, so that AI can act reliably on novel cases without erasing human skill.

Core claim

The paper’s central claim is that reliable mixed or system agency over world-state-altering actions depends on grounding: the alignment between available descriptive and prescriptive knowledge and four problem attributes (regularity, consequences, goals, solution space). When that alignment is high, system agency is appropriate; when it is incomplete or the stakes are unforgiving, agency must stay mixed or user-held, with control mechanisms as safeguards.

What carries the argument

The TDE-loop (Trigger → Decision (→ Control) → Execution (→ Control)) together with an orchestrator that maps knowledge types (historical/semantic/predictive; normative/procedural) and problem attributes onto three agency levels (user, mixed, system).

Load-bearing premise

That the four problem attributes plus the descriptive/prescriptive knowledge taxonomy are jointly enough for an orchestrator to choose correct agency levels, and that this mapping generalizes beyond the two manufacturing illustrations without a tested decision procedure.

What would settle it

Apply the same orchestrator logic to a third, independent industrial task (for example, real-time production rescheduling under supply-chain shock) and check whether the recommended agency levels and controls match expert judgment and actual safety or error outcomes; systematic mismatch would falsify the framework’s sufficiency.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Clean synthesis of agency and knowledge for manufacturing AI; useful vocabulary, untested mapping, standard position-paper limits.

read the letter

This is a short position paper that packages known pieces—MAPE-K, ReAct, Rasmussen SRK, Nonaka, Perrow coupling, existing human-AI agency taxonomies—into a manufacturing-facing TDE-loop (trigger-decision-execution plus control) and a descriptive/prescriptive knowledge split. The punchline is that the real bottleneck for safe AI autonomy is knowledge grounding, and that an orchestrator should allocate user/mixed/system agency by checking how well available knowledge lines up with four problem attributes (regularity, consequences, goals, solution space).

What is actually new is the packaging and the explicit manufacturing orientation, not a theorem or measured result. The paper does this cleanly: the knowledge taxonomy is readable, the two cases (visual quality inspection vs. factory location) make the intended contrast obvious, and the research agenda (resilience-first design, mining the reasoning layer, dynamic agency negotiation, KGs as infrastructure) is concrete enough to be useful. Citations are appropriate and not padded; the appendices keep the main text short. As conceptual work it is coherent on its own terms.

The soft spot is exactly what the stress-test flags, and it is real but proportionate for a position paper: the agency-mapping logic in Sections 4–5 and Appendix C is asserted, not tested. The cases fit the categories by construction; there is no decision procedure, measurement of the attributes, counter-example, or head-to-head with Parasuraman levels or SRK. That does not make the framework incoherent; it just means the load-bearing claim remains programmatic. No math to check, no data, no free parameters, no circular fitting of the bad kind—just definitional circularity typical of taxonomies.

Who it is for: people working on Industry 5.0, mixed-initiative systems, or knowledge infrastructure for corporate AI who need a shared vocabulary more than a new algorithm. A serious editor should send it to peer review; workshop or short-paper track is the natural home. I would bring it to reading group if we are talking agency or knowledge graphs that week, and I would cite the TDE framing and the knowledge split if I write on manufacturing human-AI collaboration. Engage with it as a clear agenda piece, not as settled method.

Referee Report

2 major / 5 minor

Summary. This position paper argues that the bottleneck for safe AI autonomy and effective human–AI collaboration in corporations is knowledge grounding rather than compute. It formalizes world-state-altering actions via a TDE-loop (Trigger → Decision (→ Control) → Execution (→ Control)), partitions organizational knowledge into descriptive (historical, semantic, predictive) and prescriptive (normative, procedural) types, and proposes that an orchestrator dynamically allocate user, mixed, or system agency according to the alignment of those knowledge types with four problem attributes (regularity, consequences, goal clarity, solution-space boundedness). Two manufacturing illustrations—visual quality inspection and factory-location choice—are used to show how the mapping would work, and the paper closes with research opportunities on resilience-first design, mining reasoning layers, and dynamic agency negotiation.

Significance. If the framing holds, it usefully re-centers industrial AI discussion on knowledge infrastructure rather than model scale alone, and it productively separates cognitive autonomy from executive power while integrating human-centric (SRK, SECI) and system-centric (MAPE-K, ReAct) loops into a mixed-agency setting. The descriptive/prescriptive taxonomy and the explicit call for semantic/normative layers as first-class assets are timely for Industry 5.0 and neuro-symbolic work. As a short workshop position paper the contribution is conceptual and agenda-setting rather than empirical; its value lies in a clean vocabulary and a set of falsifiable research questions (knowledge-graph mining, intentional inefficiency for skill resilience, grounding-gap detection) rather than in a validated decision procedure.

major comments (2)

Sections 4–5 and Appendix C present the agency-allocation logic (system agency when grounding is high + routine + forgiving; mixed for non-routine/high-stakes; user for abstract/novel) as decision rules, yet supply neither an operational measurement of the four attributes nor a procedure for resolving conflicts among them. The two case studies are post-hoc illustrations that fit the categories by construction; they do not test sufficiency, necessity, or performance against existing taxonomies (Parasuraman levels, SRK, MAPE-K). For a position paper this is acceptable if the claim is reframed as a research agenda rather than an actionable mapping; otherwise the central claim remains programmatic.
The paper asserts that the descriptive/prescriptive partition plus the four attributes are jointly the key signals an orchestrator needs (Sections 3–4). No counter-examples, edge cases, or comparison to omitted factors (e.g., liability, real-time constraints, multi-agent coordination) are examined. A short discussion of when the mapping would fail or require additional signals would strengthen the framework without requiring new empirical data.

minor comments (5)

Figure 1 is referenced as the mixed-agency model but is not described in sufficient textual detail for readers who cannot see the figure; a one-sentence caption expansion would help.
Appendix B usefully contrasts TDE with MAPE-K, Sense-Plan-Act and ReAct; a parallel short comparison of the knowledge taxonomy to DIKW and SECI already appears in Appendix A and could be cross-referenced more explicitly in the main text of Section 3.
Typographical inconsistencies: author name “RICARDO MAIA A VELINO” (space), “uninvention” vs. standard “un-invention,” and occasional missing spaces around arrows in the TDE notation.
The factory-location case claims “full system agency is unlikely due to the lack of logical reasoning in AI”; this absolute phrasing should be softened given ongoing neuro-symbolic and LLM-reasoning work the paper itself cites.
References [1] and [4] are arXiv preprints dated 2026; ensure citation stability or note preprint status consistently.

Circularity Check

1 steps flagged

Definitional framework with illustrative cases that fit by construction; no fitted predictions, equations, or load-bearing self-citation chains.

specific steps

other [Section 5 (Case Studies) + Appendix C]
"We apply our framework to two problems recently observed by the authors in manufacturing companies. Visual Quality Control. ... Agency: An AI model trains on the annotated data ... The orchestrator chooses system agency in the standard setting. ... Opening a New Factory. ... The agency can be mixed if procedural knowledge ... Full system agency is unlikely ... C The Orchestrator Logic for Agency Allocation ... System Agency: Appropriate when grounding is high (historical + predictive), the task is routine, and consequences are forgiving."

The problem attributes (routine/forgiving vs. non-routine/unforgiving, etc.) are author-assigned labels that, by the mapping defined in the same paper, immediately yield the stated agency recommendations. The cases therefore instantiate rather than independently test or predict the framework; the 'applicability' result is definitional.

full rationale

This is a position paper proposing a conceptual framework (TDE-loop + descriptive/prescriptive knowledge taxonomy + four problem attributes mapped to user/mixed/system agency). There are no equations, no parameter fits, no numerical predictions, and no uniqueness theorems. The two manufacturing cases in Section 5 are post-hoc illustrations whose attribute labels (routine/forgiving vs. non-routine/unforgiving) are assigned by the authors so that the recommended agency follows immediately from the mapping stated in Sections 4 and Appendix C; confirmation is therefore by design rather than independent test. Self-citations [7,8] (Holter & El-Assady) supply background taxonomies of human-AI collaboration but are not used to force the central mapping or forbid alternatives. The paper is self-contained as a proposal; the mild circularity is only the definitional character of any untested taxonomy-plus-examples argument. Score 2 reflects that single minor definitional loop without elevating it to a forced derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 4 invented entities

The paper is conceptual; it rests on domain assumptions imported from cognitive science, organizational theory, and automation literature, plus several ad-hoc constructs (TDE-loop stages, the four problem attributes, the three agency levels, the orchestrator) introduced without independent empirical handles. No free parameters are fitted because there is no quantitative model. Invented entities are definitional scaffolding rather than physical posits.

axioms (6)

ad hoc to paper Organizational knowledge can be exhaustively partitioned into descriptive (historical, semantic, predictive) and prescriptive (normative, procedural) types that jointly determine decision quality.
Section 3 and Appendix A synthesize Nonaka, Tulving, and Ackoff but assert this particular bipartition as the operative taxonomy for AI agency without proving completeness or mutual exclusivity.
ad hoc to paper Four problem attributes—regularity, consequences, goal clarity, solution-space boundedness—are the key signals an orchestrator needs to allocate agency.
Section 4 presents these as the decisive attributes; no ablation or comparison to alternative attribute sets is given.
ad hoc to paper World-state-altering corporate actions are adequately formalized by the TDE-loop (Trigger → Decision (→ Control) → Execution (→ Control)).
Section 2 generalizes MAPE-K, Sense-Plan-Act, and ReAct; the claim that this loop is the right unit for mixed-agency analysis is definitional.
domain assumption High-stakes or tightly coupled systems require standardized procedures and additional controls (Perrow-style coupling).
Section 4 cites Perrow; accepted in safety engineering but treated as given for manufacturing agency allocation.
domain assumption Rasmussen's SRK framework correctly ranks error rates (procedural < knowledge-based) and maps onto the paper's knowledge types.
Section 4 imports SRK without re-deriving or re-testing the error ranking in AI-mediated settings.
domain assumption Current AI systems lack reliable logical reasoning and nuanced social-norm application, so full system agency is inappropriate for non-routine strategic decisions.
Stated in the factory-location case (Section 5); treated as background fact rather than measured claim.

invented entities (4)

TDE-loop no independent evidence
purpose: Unit of analysis for world-state-altering actions that supports mixed human-AI agency allocation.
Introduced in Section 2 as a generalization of existing loops; no independent formal semantics or empirical validation outside the paper's illustrations.
Company World State (CWS) no independent evidence
purpose: Snapshot of everything the company depends on and can change, which actions alter.
Definitional construct in the introduction; useful framing but not operationalized or measured.
Orchestrator (human or AI) no independent evidence
purpose: Entity that dynamically assigns agency levels based on knowledge grounding and problem attributes.
Central to the framework (Figure 1, Appendix C); no implementation, interface, or evaluation is provided.
Grounding gap no independent evidence
purpose: Detected mismatch between available knowledge and problem attributes that should trigger agency degradation.
Section 6 research opportunity; postulated mechanism without a detection algorithm or threshold.

pith-pipeline@v1.1.0-grok45 · 11992 in / 3604 out tokens · 30390 ms · 2026-07-13T14:05:34.349160+00:00 · methodology

0 comments

read the original abstract

Organizational knowledge is fragmented across a variety of software systems, tacit expertise, and manual documents that have traditionally been designed for human consumption. As AI systems are increasingly deployed and granted decision-making roles, they require access to this knowledge. This raises two questions: how should organizations store and maintain knowledge so that it remains accessible to both humans and future AI systems, and how should agency be allocated between humans and AI across tasks with different risks and levels of uncertainty? In this position paper, we describe how organizational knowledge evolves and contribute a framework that maps task attributes and knowledge availability to recommended agency allocations and control mechanisms. We illustrate the applicability of the framework on two different manufacturing tasks: a routine operation (visual quality inspection) and a one-off strategic decision (factory location), and conclude with opportunities for future research.

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 5 canonical work pages

[1]

Guillermo Gil de Avalle, Laura Maruster, and Christos Emmanouilidis. 2026. Procedural knowledge extraction from industrial troubleshooting guides using vision language models.arXiv preprint arXiv:2601.22754(2026)

arXiv 2026
[2]

Fleuriot

Lauren Nicole DeLong, Ramon Fernández Mir, and Jacques D. Fleuriot. 2024. Neurosymbolic AI for reasoning over knowledge graphs: A survey. IEEE Trans. on Neural Networks and Learning Systems(2024). doi:10.1109/tnnls.2024.3420218

work page doi:10.1109/tnnls.2024.3420218 2024
[3]

Maria Luisa Farnese, Barbara Barbieri, Antonio Chirumbolo, and Gerardo Patriotta. 2019. Managing knowledge in organizations: A Nonaka’s SECI model operationalization.Frontiers in Psychology10 (2019), 2730. doi:10.3389/fpsyg.2019.02730

work page doi:10.3389/fpsyg.2019.02730 2019
[4]

Feng, David W

K.cJ. Feng, David W. McDonald, and Amy X. Zhang. 2025. Levels of autonomy for AI agents.arXiv preprint arXiv:2506.12469(2025)

Pith/arXiv arXiv 2025
[5]

Venkat Ram Reddy Ganuthula. 2024. The paradox of augmentation: A theoretical model of AI-induced skill atrophy.A vailable at SSRN 4974044 (2024). doi:10.2139/ssrn.4974044

work page doi:10.2139/ssrn.4974044 2024
[6]

Michael E. Gorman. 2002. Types of knowledge and their roles in technology transfer.J. Technology Transfer27, 3 (2002), 219–231. doi:10.1023/a: 1015672119590

doi:10.1023/a: 2002
[7]

Steffen Holter and Mennatallah El-Assady. 2024. Deconstructing human-AI collaboration: Agency, interaction, and adaptation.Computer Graphics Forum43 (2024), e15107. doi:10.1111/cgf.15107

doi:10.1111/cgf.15107 2024
[8]

Steffen Holter, Caterina Moruzzi, and Mennatallah El-Assady. 2026. Toward agency in human–AI collaboration.IEEE Computer Graphics and Applications46 (2026), 13–25. doi:10.1109/mcg.2025.3623892

work page doi:10.1109/mcg.2025.3623892 2026
[9]

Daniel Kahneman. 2011. Thinking, fast and slow.Farrar, Straus and Giroux(2011)

2011
[10]

Kephart and David M

Jeffrey O. Kephart and David M. Chess. 2003. The vision of autonomic computing.Computer36, 1 (2003), 41–50. doi:10.1109/mc.2003.1160055

Pith/arXiv arXiv doi:10.1109/mc.2003.1160055 2003
[11]

2019.The Soar cognitive architecture

John E Laird. 2019.The Soar cognitive architecture. MIT press

2019
[12]

Raja Parasuraman, Thomas B Sheridan, and Christopher D Wickens. 2000. A model for types and levels of human interaction with automation. IEEE Trans. on Systems, Man, and Cybernetics(2000), 286–297

2000
[13]

2011.Normal accidents: Living with high risk technologies

Charles Perrow. 2011.Normal accidents: Living with high risk technologies. Princeton University Press

2011
[14]

Jens Rasmussen. 2012. Skills, rules, and knowledge; signals, signs, and symbols, and other distinctions in human performance models.IEEE Trans. Systems, Man, and Cybernetics3 (2012), 257–266. doi:10.1109/tsmc.1983.6313160

doi:10.1109/tsmc.1983.6313160 2012
[15]

Horst W. J. Rittel and Melvin M. Webber. 1973. Dilemmas in a general theory of planning.Policy sciences4, 2 (1973), 155–169

1973
[16]

Lisa Soder, Julia Smakman, Connor Dunlop, Weiwei Pan, Siddharth Swaroop, and Noam Kolt. 2024. Levels of autonomy: Liability in the age of AI agents. InWorkshop on Socially Responsible Language Modelling Research

2024
[17]

Ankit Srivastava. 2019. Sense-plan-act in robotic applications. InIntelligent Robotics Seminar. 1–8

2019
[18]

Endel Tulving et al. 1972. Episodic and semantic memory.Organization of Memory1 (1972), 381–403

1972
[19]

Ilkka Tuomi. 1999. Data is more than knowledge: Implications of the reversed knowledge hierarchy for knowledge management and organizational memory.J. of Management Information Systems16, 3 (1999), 103—-117. doi:10.1080/07421222.1999.11518258

work page doi:10.1080/07421222.1999.11518258 1999
[20]

triggers

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InProc. Int. Conf. on Learning Representations. A Theoretical Foundations of Grounding Our taxonomy of knowledge grounding synthesizes concepts from cognitive science and organizational theory, ad...

2022

[1] [1]

Guillermo Gil de Avalle, Laura Maruster, and Christos Emmanouilidis. 2026. Procedural knowledge extraction from industrial troubleshooting guides using vision language models.arXiv preprint arXiv:2601.22754(2026)

arXiv 2026

[2] [2]

Fleuriot

Lauren Nicole DeLong, Ramon Fernández Mir, and Jacques D. Fleuriot. 2024. Neurosymbolic AI for reasoning over knowledge graphs: A survey. IEEE Trans. on Neural Networks and Learning Systems(2024). doi:10.1109/tnnls.2024.3420218

work page doi:10.1109/tnnls.2024.3420218 2024

[3] [3]

Maria Luisa Farnese, Barbara Barbieri, Antonio Chirumbolo, and Gerardo Patriotta. 2019. Managing knowledge in organizations: A Nonaka’s SECI model operationalization.Frontiers in Psychology10 (2019), 2730. doi:10.3389/fpsyg.2019.02730

work page doi:10.3389/fpsyg.2019.02730 2019

[4] [4]

Feng, David W

K.cJ. Feng, David W. McDonald, and Amy X. Zhang. 2025. Levels of autonomy for AI agents.arXiv preprint arXiv:2506.12469(2025)

Pith/arXiv arXiv 2025

[5] [5]

Venkat Ram Reddy Ganuthula. 2024. The paradox of augmentation: A theoretical model of AI-induced skill atrophy.A vailable at SSRN 4974044 (2024). doi:10.2139/ssrn.4974044

work page doi:10.2139/ssrn.4974044 2024

[6] [6]

Michael E. Gorman. 2002. Types of knowledge and their roles in technology transfer.J. Technology Transfer27, 3 (2002), 219–231. doi:10.1023/a: 1015672119590

doi:10.1023/a: 2002

[7] [7]

Steffen Holter and Mennatallah El-Assady. 2024. Deconstructing human-AI collaboration: Agency, interaction, and adaptation.Computer Graphics Forum43 (2024), e15107. doi:10.1111/cgf.15107

doi:10.1111/cgf.15107 2024

[8] [8]

Steffen Holter, Caterina Moruzzi, and Mennatallah El-Assady. 2026. Toward agency in human–AI collaboration.IEEE Computer Graphics and Applications46 (2026), 13–25. doi:10.1109/mcg.2025.3623892

work page doi:10.1109/mcg.2025.3623892 2026

[9] [9]

Daniel Kahneman. 2011. Thinking, fast and slow.Farrar, Straus and Giroux(2011)

2011

[10] [10]

Kephart and David M

Jeffrey O. Kephart and David M. Chess. 2003. The vision of autonomic computing.Computer36, 1 (2003), 41–50. doi:10.1109/mc.2003.1160055

Pith/arXiv arXiv doi:10.1109/mc.2003.1160055 2003

[11] [11]

2019.The Soar cognitive architecture

John E Laird. 2019.The Soar cognitive architecture. MIT press

2019

[12] [12]

Raja Parasuraman, Thomas B Sheridan, and Christopher D Wickens. 2000. A model for types and levels of human interaction with automation. IEEE Trans. on Systems, Man, and Cybernetics(2000), 286–297

2000

[13] [13]

2011.Normal accidents: Living with high risk technologies

Charles Perrow. 2011.Normal accidents: Living with high risk technologies. Princeton University Press

2011

[14] [14]

Jens Rasmussen. 2012. Skills, rules, and knowledge; signals, signs, and symbols, and other distinctions in human performance models.IEEE Trans. Systems, Man, and Cybernetics3 (2012), 257–266. doi:10.1109/tsmc.1983.6313160

doi:10.1109/tsmc.1983.6313160 2012

[15] [15]

Horst W. J. Rittel and Melvin M. Webber. 1973. Dilemmas in a general theory of planning.Policy sciences4, 2 (1973), 155–169

1973

[16] [16]

Lisa Soder, Julia Smakman, Connor Dunlop, Weiwei Pan, Siddharth Swaroop, and Noam Kolt. 2024. Levels of autonomy: Liability in the age of AI agents. InWorkshop on Socially Responsible Language Modelling Research

2024

[17] [17]

Ankit Srivastava. 2019. Sense-plan-act in robotic applications. InIntelligent Robotics Seminar. 1–8

2019

[18] [18]

Endel Tulving et al. 1972. Episodic and semantic memory.Organization of Memory1 (1972), 381–403

1972

[19] [19]

Ilkka Tuomi. 1999. Data is more than knowledge: Implications of the reversed knowledge hierarchy for knowledge management and organizational memory.J. of Management Information Systems16, 3 (1999), 103—-117. doi:10.1080/07421222.1999.11518258

work page doi:10.1080/07421222.1999.11518258 1999

[20] [20]

triggers

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InProc. Int. Conf. on Learning Representations. A Theoretical Foundations of Grounding Our taxonomy of knowledge grounding synthesizes concepts from cognitive science and organizational theory, ad...

2022