Human-AI-Lean Workflow

• Human-AI-Lean Workflow is a family of designs that combine human judgment, AI assistance, and lean processes to reduce friction in iterative tasks.
• It emphasizes multi-role collaboration, asynchronous checkpoints, and continuous semantic review to balance automation with human authority.
• The approach leverages structured orchestration and provenance-aware interfaces to enhance efficiency and transparency across diverse applications.

Searching arXiv for the cited workflow papers to ground the article and verify metadata. arxiv_search(query="(Rahman et al., 2023) OR (Mendoza et al., 5 May 2026) OR (Khoshkonesh et al., 23 Nov 2025) OR (Fogliato et al., 2022) OR (He et al., 19 Jan 2025) OR (Gu et al., 2020) OR (Spillner et al., 5 Mar 2026) OR (Yanahama et al., 16 Mar 2026) OR (Zhu et al., 30 Jan 2026) OR (Tang et al., 13 Feb 2026) OR (Zhang et al., 3 Apr 2025)", max_results=10) Searching again with a narrower query. arxiv_search(query="(Rahman et al., 2023)", max_results=5) Human-AI-Lean Workflow denotes a family of workflow designs in which human judgment, AI assistance, and workflow structure are jointly organized to reduce friction across iterative work rather than to maximize isolated automation. Across the literature, it is associated with multi-role, iterative, tool-spanning, and domain-dependent workflows; asynchronous human checkpoints in compute-intensive settings; predictive and human-centered planning loops; transparent multi-step decision procedures; and Lean-based formalization pipelines in which AI generates code, Lean checks formal correctness, and humans retain semantic authority (Rahman et al., 2023, Mendoza et al., 5 May 2026, Khoshkonesh et al., 23 Nov 2025, Garg, 11 Jun 2026). The common premise is that efficiency emerges from disciplined allocation of responsibilities, explicit workflow state, selective escalation, and built-in review mechanisms, not from removing humans from consequential decisions.

1. Conceptual scope and intellectual framing

The broadest workflow-level formulation is “Multifaceted Human-centered AI” (M-HCAI), which argues that real AI work is not a single user interacting with a single model, but a multi-role, iterative, tool-spanning, domain-dependent workflow involving humans and automated agents together (Rahman et al., 2023). In this framing, human-AI support should not be designed as isolated point solutions for one task, one interface, or one user type. Instead, workflow support must account for multiple stakeholders, diverse tasks across phases, different tools and interfaces, synchronous and asynchronous collaboration, and interactions among stakeholders and automated agents.

Several later papers generalize this claim in adjacent directions. “Workflow as Medium” treats workflow itself as a “dynamic and transformative medium” that shapes the artifact, the human practitioner, the AI collaborators, and the workflow’s own evolution (Ackerman, 22 Nov 2025). “Human Tool” moves the coordination center from human-led orchestration toward AI-led orchestration, exposing humans as callable tools through structured schemas of capabilities, information, and authority (Tang et al., 13 Feb 2026). The hybrid-allocation theory of idempotent equilibrium formalizes long-run delegation as a fixed point $A^\star = F(A^\star)$, with a mixed equilibrium $x^\star = \alpha / (\alpha + \beta)$ whenever $\beta > 0$, thereby excluding full automation under persistent human-centric task creation (Alpay et al., 2 Aug 2025). This suggests that “Human-AI-Lean Workflow” is less a single method than a family of designs for stable human-AI complementarity.

A second major strand concerns Lean-based formalization. Here the “lean” component is partly workflow efficiency and partly the Lean proof assistant itself. EconCSLib, Lean Atlas, LeanArchitect, CSLib, and the SLT formalization project all treat workflow design as the key mechanism for combining AI generation, machine-checked proof, and human semantic review (Garg, 11 Jun 2026, Yanahama et al., 16 Mar 2026, Zhu et al., 30 Jan 2026, Barrett et al., 4 Feb 2026, Zhang et al., 2 Feb 2026). In these systems, the central problem is not only logical correctness, but also semantic fidelity: whether a formal statement actually expresses the intended mathematics.

2. Workflow structures and lifecycle models

A defining feature of the literature is rejection of linear stage models. M-HCAI emphasizes that workflows move back and forth among phases rather than proceeding linearly, and that iteration also occurs within phases through repeated cycles of task execution, checking, deliberation, and revision (Rahman et al., 2023). This micro-iterative view is echoed in FlexMind, where ideation is described as non-linear movement among searching, creating, and evaluating, supported by opt-in aids that can be invoked “at any moment” (Yang et al., 15 Sep 2025).

Several papers make the lifecycle explicit. Lean 5.0 operationalizes a human-AI-lean workflow through the Predictive Lean Flow framework, a dual-layer socio-technical model with the five-phase cycle Plan -> Sense -> Predict -> Collaborate -> Learn, linking a Digital Analytics Layer to a Human Collaboration Layer (Khoshkonesh et al., 23 Nov 2025). CIF, designed for asynchronous Human-in-the-Loop AI in hybrid HPC environments, summarizes its control flow as

$$\text{Parse} \rightarrow \text{Schedule} \rightarrow \text{Execute} \rightarrow \text{HITL Review} \rightarrow \text{Adapt} \rightarrow \text{Continue}.$$

Its distinctive property is that workflow branches can pause logically for human review without halting unrelated jobs or holding scarce compute resources idle (Mendoza et al., 5 May 2026).

Decision-support workflows often use staged interaction rather than lifecycle phases. In clinical imaging, the critical distinction is between a one-step workflow, where AI is shown from the start, and a two-step workflow, where a clinician first enters a provisional decision and only then sees AI output (Fogliato et al., 2022). In composite fact-checking, the Multi-Step Transparent workflow uses a two-stage setup: participants first work through decomposed sub-facts and evidence, then receive final AI advice together with intermediate steps and answers (He et al., 19 Jan 2025). In Human Tool, the agent reasons over a task hierarchy and invokes a human only when capability complementarity, information exchange, or authority control is required (Tang et al., 13 Feb 2026).

Formalization projects instantiate yet another workflow family. EconCSLib starts from a paper PDF or $\LaTeX$ source, extracts definitions and theorems, builds a dependency DAG, generates Lean code, exposes a compact PaperInterface.lean for review, and records translation judgments in a validation dashboard (Garg, 11 Jun 2026). Lean Atlas computes a pruned semantic-dependency set for selected target theorems, so humans inspect only the declarations whose semantic correctness can affect those theorems (Yanahama et al., 16 Mar 2026). LeanArchitect embeds blueprint metadata directly in Lean declarations and synchronizes blueprint state from code rather than manually maintaining a parallel $\LaTeX$ artifact (Zhu et al., 30 Jan 2026).

3. Division of labor, authority, and human responsibility

The literature consistently rejects simple replacement narratives. In M-HCAI, example stakeholders include product managers, subject matter experts, and data scientists, while automated agents are treated as first-class actors in the workflow (Rahman et al., 2023). Lean 5.0 assigns forecasting, integration of fragmented data, and probability-based alerting to AI and predictive analytics, but reserves contextual interpretation, collaborative prioritization, trust-based judgment, ethical oversight, and final decision-making for foremen, planners, managers, analysts, and governance bodies such as an AI Governance Committee (Khoshkonesh et al., 23 Nov 2025).

Human Tool offers the most explicit schema for role allocation. Humans are modeled through Capabilities, Information, and Authority. Capability complementarity covers novel reasoning, creativity, complex judgment, or interaction with the external world; information exchange covers domain expertise, private constraints, and preferences inaccessible to AI; authority control covers decisions requiring explicit human responsibility or authorization because of safety, privacy, or ethics (Tang et al., 13 Feb 2026). The framework is explicit that “tool” is a coordination abstraction, not a reduction of human authority or responsibility.

The creative co-creation literature uses a different vocabulary but preserves the same asymmetry. The Creative Intelligence Loop assigns AI roles such as Blue Team constructive specialists and Red Team critical specialists, while the human remains the “final arbiter for ethical alignment and creative integrity” (Ackerman, 22 Nov 2025). FlexMind likewise treats AI as an on-demand thinking partner whose interventions are opt-in and contextual, while the human chooses which idea to inspect, when to request variants or risk analysis, when to ask questions, and when to save or revisit ideas (Yang et al., 15 Sep 2025).

Formalization workflows make human authority even sharper. EconCSLib states the tri-partite contract directly: “an LLM writes Lean code, Lean checks formal statements and proofs, and humans (assisted by an LLM) verify the translation boundary from paper claims to formal statements” (Garg, 11 Jun 2026). Lean Atlas calls the remaining human obligation “semantic verification” and defines aligned Lean code as code whose propositions and definitions have undergone human semantic verification (Yanahama et al., 16 Mar 2026). The SLT formalization project describes the same allocation operationally: humans designed proof strategies and decomposed mathematics into manageable lemmas, AI agents executed tactical proof construction, and Lean 4 acted as the final verifier and debugging oracle, producing a roughly 30,000-line codebase over about 500 hours of supervised development with no sorry or axiom (Zhang et al., 2 Feb 2026).

4. Interfaces, transparency, and workflow memory

Interface design is a central determinant of whether a workflow is actually lean. M-HCAI proposes plastic interfaces as boundary objects that are “both plastic enough to adapt to local needs and the constraints of the several parties employing them, yet robust enough to maintain a common identity across site” (Rahman et al., 2023). The associated system-level claim is that provenance-aware interfaces, integrated workflows, and a graph-based backend should be designed together, because provenance cannot be bolted on later.

Clinical and decision-support papers show that interface sequencing changes reliance patterns. In radiology, one-step AI-first display increased agreement with AI, but also strengthened anchoring around the AI’s 0.6 flag threshold; the two-step condition reduced agreement with AI but induced very little revision after AI was shown, with only 70 revisions out of 10,560 finding-level decisions (Fogliato et al., 2022). A later trust study found no evidence that a 2-step setup reduces overreliance; instead, the 2-step workflow significantly increased overreliance, while explanation effects crossed over by workflow and domain knowledge (Spillner et al., 5 Mar 2026). These results directly challenge the view that simply forcing an initial human judgment is a universal safeguard.

Transparency itself is not uniformly beneficial. In composite fact-checking, global transparency increased reliance, but the Multi-Step Transparent workflow did not generally improve appropriate reliance or team performance across all settings (He et al., 19 Jan 2025). Its main benefit appeared in the three tasks where AI advice was wrong, and only when users had relatively high AR-Intermediate, defined as accurate intermediate sub-fact decisions. The annotation-heavy MSTworkflow+ condition increased Mental Demand and Frustration and degraded outcomes, showing that additional explicit processing can backfire if it adds cognitive overhead without improving intermediate reasoning (He et al., 19 Jan 2025).

Workflow memory and navigation are equally important. xPath organizes pathology AI around “joint-analyses of multiple criteria” and “explanation by hierarchically traceable evidence,” so a pathologist can move from suggested grade to criterion-level evidence to original slide context (Gu et al., 2020). FlexMind uses a canvas, sidebar, and explored-canvas view to preserve a revisitable ideation landscape rather than a linear transcript (Yang et al., 15 Sep 2025). CIF preserves workflow state through WorkflowExecution, JobRegistry, and Watcher, allowing branches to wait for human review without collapsing the rest of the pipeline (Mendoza et al., 5 May 2026). This suggests that a lean workflow depends not only on accurate models, but also on persistent shared state that reduces reconstruction effort across iterations, tools, and roles.

5. Infrastructure, orchestration backends, and Lean-based formalization

At the system level, the literature converges on explicit workflow representations. M-HCAI proposes that “knowledge graphs (KGs) may serve as the core data model for the support systems managing M-HCAI workflows,” encoding people, roles, agents, tasks, interactions, workflow context, and provenance relationships (Rahman et al., 2023). CIF uses a declarative TOML workflow file as the organizing artifact in a workflow-oriented architecture. Its runtime is composed of CIF-CLI, TaskScheduler, TaskExecutor, WorkflowExecution, Watcher, and JobRegistry, with asynchronous monitoring of HPC jobs through .done markers because workflow-native scheduler notifications are unavailable (Mendoza et al., 5 May 2026).

A different but related infrastructure strand appears in formalization and verification systems. LeanArchitect introduces the @[blueprint] annotation and automatically infers dependency information, proof status, and synchronized $\LaTeX$ blueprint content directly from Lean code, eliminating duplication between formal and informal representations (Zhu et al., 30 Jan 2026). Lean Atlas visualizes Lean dependency graphs and uses Lean Compass to prune theorem-value proof dependencies, reducing semantic review sets by 94–99% on proof-heavy projects such as PrimeNumberTheoremAnd, Carleson, and Brownian Motion, by 59.8% on a six-theorem FLT milestone subset, by 69.0% on PhysLib, and by 27.3% on the definition-heavy XMSS project (Yanahama et al., 16 Mar 2026).

EconCSLib operationalizes author-facing paper formalization with 11 fully formalized papers and 3 partially formalized papers in its public repository as of June 11, 2026, together with review dashboards, result dependency graphs, and FINAL_VALIDATION_REPORT.md files (Garg, 11 Jun 2026). CSLib extends the same ambition to computer science, aiming to provide for CS what Mathlib provides for mathematics and to support manual and AI-aided engineering of large-scale formally verified systems (Barrett et al., 4 Feb 2026). Pocketflow, by contrast, addresses compound AI systems outside theorem proving. It uses a minimal architecture based on Node, Flow, Flow-as-Node nesting, and action-string conditionals, with benchmarked cold starts of 47 ms for a RAG pipeline, 53 ms for an agent loop, and 41 ms for a multi-step flow, while keeping the framework package at 56 KB (Zhang et al., 3 Apr 2025). Across these systems, lean-ness is achieved by explicit orchestration, decomposition into reusable units, and machine-readable workflow artifacts.

6. Empirical findings, tradeoffs, and recurring controversies

The empirical record does not support a single optimal workflow. In construction management, Lean 5.0 reports a 13.1% improvement in PPC, a 22.0% reduction in Rework Ratio, a 23.4% reduction in Waiting Waste, a 13.8% improvement in Coordination Efficiency, and a 42.1% improvement in Forecast Accuracy (± Days), with all reported effect sizes $d > 0.8$ in a 12-week single-case field study (Khoshkonesh et al., 23 Nov 2025). In deductive coding, a selective workflow using MacBERT for all samples, routing cases with probability $< 0.6$ or assigned to code categories occurring in $< 5\%$ of samples, and then using GPT-4o plus expert adjudication improved reliability from $x^\star = \alpha / (\alpha + \beta)$0 to $x^\star = \alpha / (\alpha + \beta)$1; only 44 of 500 dialogue turns were escalated, requiring about 45 minutes of expert correction (Li et al., 24 Dec 2025). In Human Tool, AI-led orchestration improved Travel Planning accuracy from 72.66 to 86.72 and Story Writing human-preference win rate from 0.371 to 0.611, while reducing Story Writing mental effort from 87.875 to 70.625 (Tang et al., 13 Feb 2026).

Yet several common intuitions are contradicted by the evidence. First, human-first does not universally reduce miscalibrated reliance. In veterinary radiology, pre-commitment reduced agreement with AI but also reduced revision and perceived usefulness (Fogliato et al., 2022). In a student-support decision task, the 2-step workflow increased overreliance rather than reducing it (Spillner et al., 5 Mar 2026). Second, more transparency is not automatically better. Global transparency can increase reliance without guaranteeing appropriate reliance, and cognitive-forcing additions can worsen outcomes by increasing workload (He et al., 19 Jan 2025). Third, better workflow integration can increase performance while also creating new risks. xPath improved diagnostic correctness from 7 of 12 to 17 of 20 decisions and sharply improved perceived comprehensiveness and integrability, but some incorrect cases resulted from over-reliance on AI-highlighted evidence and insufficient broad slide review (Gu et al., 2020).

A final controversy concerns the meaning of “lean.” In some papers it primarily means minimizing context-switching, idle compute, or documentation loss across a multi-role workflow (Rahman et al., 2023, Mendoza et al., 5 May 2026). In others it refers to predictive control and waste reduction in construction management (Khoshkonesh et al., 23 Nov 2025), or to selective escalation and workload concentration in qualitative coding (Li et al., 24 Dec 2025). In Lean formalization papers it partly denotes the Lean proof assistant itself and workflows for semantic assurance (Garg, 11 Jun 2026, Yanahama et al., 16 Mar 2026). This suggests that “Human-AI-Lean Workflow” is best understood as a cross-domain design orientation rather than a single canonical framework: assign responsibilities by comparative advantage, surface critical state and provenance, preserve human authority where meaning or accountability is at stake, and reduce waste by structuring iteration rather than by assuming more automation is inherently superior.