💻 computer science

The Orchestration Gap: Why Process Automation Stalls in Operationally Complex Industries

This paper argues that the stagnation of agentic automation in operationally complex industries stems from a missing orchestration abstraction, proposing a staged framework where the critical architectural focus shifts between constraint enforcement and explainability depending on whether a sector's primary friction is regulatory or liability-driven.

Original authors: Jiechao Gao, Yuandong Pan. Yuangang Li, Jie Wang, Kincho Law, Michael Lepech

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiechao Gao, Yuandong Pan. Yuangang Li, Jie Wang, Kincho Law, Michael Lepech

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart assistant (an AI agent) who can write a perfect email, solve a math problem, or draft a contract in seconds. In the digital world of software and finance, this assistant is a star. But try to put that same assistant to work in a construction site, a hospital, or a shipping warehouse, and it often stalls. It gets confused, makes mistakes, or simply refuses to move forward.

Why? According to this paper, it's not because the AI isn't smart enough. It's because we are missing the conductor.

The Missing Conductor: What is "Orchestration"?

Think of a complex industry like a massive, chaotic orchestra where every musician is playing a different instrument, reading from a different sheet of music, and speaking a different language.

The AI is the virtuoso soloist who can play a perfect solo.
The Orchestration is the conductor who tells the soloist when to start, ensures they don't clash with the drummer, checks that they are following the rules of the song, and makes sure the human manager gets to approve the final performance.

The paper argues that in messy, real-world industries, the value isn't in the soloist's skill; it's in the conductor's ability to coordinate the whole mess. This "conductor" layer is what the paper calls Orchestration. Currently, most AI tools are great soloists but terrible conductors.

How Do You Know If You Need a Conductor?

The authors created a simple "test" to see if a job is too messy for a solo AI and needs a conductor. If a workflow has at least four of these five traits, it is "orchestration-bound":

Too many systems: The work jumps between 3+ different software programs that don't talk to each other (like a shipping app, a warehouse app, and a customer email).
Too many people: It requires handing off tasks between 3+ different roles (e.g., a doctor, a billing specialist, and an insurance agent).
Hidden rules: Decisions rely on "gut feelings" or unwritten experience that no one has typed into a manual.
Hard rules: There are strict "do not cross" lines (like safety laws or legal contracts) where a mistake is a disaster, not just a small error.
Old tech: The workers are stuck using ancient, clunky software that they refuse to throw away.

If your job looks like this, throwing a smarter AI at it won't fix it. You need a system to manage the chaos.

The Three Big Problems: Logistics, Hospitals, and Construction

The paper looks at three industries to show how this works:

Logistics (Shipping): When a package gets lost, it triggers a chain reaction of phone calls and system updates. The "friction" here is low (few strict laws), so the main problem is just the sheer number of people and systems talking to each other.
Healthcare (Insurance Approvals): Here, the "friction" is regulatory. You can't just approve a surgery; you must follow strict insurance rules. If the AI breaks a rule, the whole system fails. The value is in following the rules perfectly, not just being fast.
Construction: Here, the "friction" is liability. If a building collapses or a design is wrong, someone gets sued. Because many different companies (architects, builders, owners) are involved, no one will trust an AI unless it can provide a perfect, unchangeable "paper trail" explaining exactly why it made a decision.

The "Staged" Solution: Don't Run Before You Walk

The paper suggests we shouldn't try to build a fully autonomous robot that does everything at once. Instead, we should climb a ladder in four stages:

Observability: The AI just watches and reports what's happening (like a traffic camera).
Recommendation: The AI suggests what to do, but a human hits the "Go" button.
Bounded Automation: The AI handles small, safe, repetitive tasks on its own (like filing a form), but stops if things get weird.
Supervised Autonomy: The AI runs the whole show, but a human is watching closely to jump in if needed.

The Golden Rule: Friction Dictates the Plan

The most important idea in the paper is the Friction Principle. It says that the "best" AI system depends on what kind of trouble you are trying to avoid:

If your main trouble is Regulations (like in hospitals), the most important feature for your AI is Constraint Enforcement. It must be built to never break the rules, even if it means being slower.
If your main trouble is Lawsuits (like in construction), the most important feature is Explainability. The AI must be able to say, "I did this because of X, Y, and Z," so a judge or lawyer can verify it.
If your main trouble is just Chaos (like in shipping), the most important feature is Coordination. The AI just needs to be good at juggling many tasks at once.

The Bottom Line

We aren't failing to automate industries because our AI isn't smart enough. We are failing because we are trying to use a "smart soloist" in a "messy orchestra."

To fix this, we need to stop building better soloists and start building better conductors. We need to build systems that can handle the handoffs, respect the old software, enforce the hard rules, and keep a perfect record of who did what. Once we build that "conductor" layer, the smart AI can finally do its job.

Technical Summary: The Orchestration Gap

Problem Statement
While agentic AI systems have achieved significant progress in digitally native domains (e.g., software engineering, financial trading), their deployment in operationally complex industries—such as logistics, healthcare operations, and construction—remains stalled. Industry surveys indicate that while experimentation with agents is common, scaled, integrated deployments are rare, and a significant fraction of projects face cancellation. The paper argues that this "deployment gap" is not caused by a lack of model capability or accuracy. Instead, it stems from a missing abstraction: orchestration. In complex environments, value is not derived from a single model invocation but from the runtime coordination of multi-step workflows, the enforcement of hard domain constraints, the management of human approvals, and the integration of disparate legacy systems. Current multi-agent frameworks lack the architectural guarantees necessary to handle these operational realities.

Methodology and Conceptual Framework
The authors develop a conceptual framework to analyze and address this gap, moving orchestration from a slogan to a designable system component. The methodology involves three primary analytical steps:

Defining "Orchestration-Bound" Workflows: The paper proposes a five-property test to identify workflows requiring an orchestration layer. A workflow is considered orchestration-bound if it satisfies at least four of the following:
- System Fragmentation: Spans $\ge$ 3 systems of record without a unifying platform.
- Stakeholder Fragmentation: Requires handoffs among $\ge$ 3 roles or organizations.
- Tacit Decision Points: Relies on uncodified judgment not fully specified by written procedures.
- Constraint Criticality: Contains hard constraints (regulatory, safety, contractual) where violation is unacceptable.
- Brownfield Pressure: Operates on entrenched legacy tools that cannot be replaced wholesale.
Decomposing Orchestration Value: The authors separate the problem into three distinct dimensions to clarify where automation adds value:
- Fragmentation ( $F$ ): A measure of structural entanglement calculated as $F = \log_{10}(s \cdot r \cdot h)$ , where $s$ is the number of systems, $r$ is the number of stakeholder roles, and $h$ is the number of cross-boundary handoffs.
- Coordination Overhead ( $C$ ): The proportion of total effort spent on coordination rather than substantive work.
- Recoverable Value ( $V$ ): The potential value gain, calculated as $V = n \cdot c \cdot \rho$ , where $n$ is workflow frequency, $c$ is coordination cost, and $\rho$ is the achievable reduction in cost.
- Friction: A qualitative axis representing regulatory, liability, and trust barriers that govern the speed of deployment and the specific architectural guarantees required.
Industry Analysis: The framework is applied to three focal industries (logistics, healthcare operations, construction) and five additional sectors. The analysis reveals that while these industries share a similar "fragmentation band" (structurally tangled workflows), they diverge significantly in their dominant sources of friction (regulatory vs. liability vs. low friction).

Key Contributions

The Orchestration Gap Thesis: The paper posits that the primary barrier to AI deployment in complex industries is the absence of an orchestration runtime that treats coordination, constraint enforcement, and legacy integration as first-class architectural requirements rather than optional implementation details.
Five Requirements for Orchestration Runtimes: The authors define five specific requirements that current frameworks fail to meet comprehensively:
- R1: Multi-step coordination with human gates.
- R2: Domain-knowledge grounding (consulting structured rules beyond general model knowledge).
- R3: Constraint enforcement before delivery (blocking infeasible outputs structurally).
- R4: Explainability (emitting structured audit traces).
- R5: Brownfield integration (interoperating with existing systems).
- Analysis of current tools (e.g., AutoGen, MetaGPT, LangGraph) shows they handle R1 well but treat R2, R3, and R5 as optional or engineering challenges rather than enforced guarantees.
The Friction Principle and Staged Automation: The central claim is that automation should proceed in stages, and the "load-bearing" architectural guarantee depends on the sector's dominant friction:
- Regulatory Friction (e.g., Healthcare): The load-bearing requirement is Constraint Enforcement (R3). Value is destroyed if rules are violated; thus, the constraint gate is critical.
- Liability Friction (e.g., Construction): The load-bearing requirement is Explainability (R4). Distributed liability requires verifiable audit trails for automated actions to be accepted.
- Low Friction (e.g., Logistics): The load-bearing requirement is Coordination (R1). Automation can proceed based on coordination efficiency alone.
Staged Automation Path: The paper outlines a four-stage progression for autonomy:
1. Observability: Unifying workflow state without decision-making.
2. Recommendation: Proposing actions with human decision-making.
3. Bounded Automation: Executing well-defined, low-risk sub-tasks autonomously.
4. Supervised Autonomy: Executing multi-step workflows under human oversight.
  The "friction principle" dictates that different sectors will hit a "ceiling" at different stages based on their dominant friction.

Results and Evidence
The paper relies on public operational evidence and industry surveys rather than new experimental deployments, given the current lack of orchestration systems at the proposed granularity.

Deployment Statistics: Citing McKinsey, BCG, and Gartner, the paper highlights that while 72% of workers use generative tools and 39% experiment with agents, only 13% have agents integrated into broader workflows, and over 40% of agentic projects are projected to be canceled due to weak integration.
Quantitative Decomposition: The analysis of the three focal industries shows they all fall within a narrow fragmentation band ( $F \approx 5.1$ to $5.9$), confirming that structural complexity is a common baseline. However, their coordination overheads differ (e.g., 53% of shipping cost in logistics, 34% of health spending in administration, 20% of construction cost in rework).
Validated Effects: The paper uses existing data points to anchor the potential value of automation. For example, in healthcare, the transition from manual to electronic prior authorization reduces cost from $3.41 to $0.05 (a 98% reduction), yet only 20% of such workflows are fully electronic, illustrating that the gap is orchestration, not digitization.

Significance and Research Program
The paper concludes that the "orchestration gap" is the reason for the current plateau in industrial AI adoption. Its significance lies in shifting the research focus from improving model capabilities (benchmarks) to building infrastructure for orchestration. The authors propose a research program focused on:

Workflow-level Benchmarks: Measuring coordination cost, constraint-violation rates, and audit completeness rather than single-step accuracy.
Deployment Studies: Empirically measuring the reduction ( $\rho$ ) per stage on real workflows to calibrate the staged automation curve.
Runtime Evaluation: Conducting head-to-head evaluations of runtimes against the five defined requirements to test the "architectural guarantee" claim.

The authors maintain a modest stance, acknowledging that their framework is built on public evidence rather than deployed systems and that orchestration is necessary but not sufficient for success (data quality, incentives, and trust remain critical). They argue that addressing the orchestration gap is the most promising path forward for realizing the operational value of AI in complex industries.