From Software to Outcomes: Why Industrial AI Is Moving Toward Agent-as-a-Service

For decades, enterprise technology has followed a familiar pattern. Buy software, configure it, train users, and hope adoption produces value.

Artificial intelligence initially followed the same route. Companies bought platforms, assembled pilots, connected models to data, and waited for measurable improvement. In industrial settings, where decisions carry high consequences, data is scattered, and operational knowledge is hard-won, this formula keeps breaking down.

Models are plentiful. The distance between a model and an operational decision is the real bottleneck.

A recent paper by Zhenfeng Cao, "The End of Software Engineering: How AI Agents Are Restructuring the Software Paradigm" (arXiv:2606.05608, June 2026), describes a transition from AI → Software → Result to Agent → Result. The paper argues that agents can generate and revise decision logic at runtime. In practice, agent systems may plan, use tools, retrieve context, execute bounded actions, and adapt their work around a defined objective. The enduring unit of value shifts from the software artifact to the outcome.

That idea has major implications for industrial organizations. In sectors like oil and gas, energy, manufacturing, logistics, and capital projects, the next era of AI is likely to be defined by who can turn years of operational data, expert judgment, and live workflows into repeatable, evidence-backed decisions.

This is the opening for Agent-as-a-Service.

The old model: buy technology, carry the complexity

Traditional enterprise software has always asked customers to absorb most of the burden. Define the process in advance. Configure a platform. Integrate disconnected systems. Clean and structure the data. Train teams. Maintain workflows as operations change. Prove value after the fact.

Software-as-a-Service improved infrastructure delivery. The vendor ran the servers, issued upgrades, and simplified access through subscriptions. The customer still had to translate a business problem into requirements, select the right use case, integrate data, redesign workflows, and make the system useful.

In industrial AI, that gap is especially visible. An operator may have pre-job risk registers, daily operational reports, end-of-project reviews, spreadsheets, maintenance histories, engineering documents, technical standards, and expert knowledge distributed across teams. Each source may be valuable. Each is also incomplete on its own.

The result is a familiar cycle. Teams rebuild plans from memory. The same risks are rediscovered. Lessons learned stay locked in documents. AI initiatives stall at the demonstration stage.

Industry spending on digital and AI is rising, yet fewer than one in five organizations track KPIs for their generative AI solutions (McKinsey, 2026 State of AI). The real problem runs deeper than unstructured data. Organizations lack a trusted operating layer that connects what was planned, what happened, what was learned, and what should happen next.

Human in the loop has become a leadership priority

Executives have already reached this conclusion in public. BCG's analysis of earnings calls found that mentions of "human in the loop" climbed sharply between the first and second quarters of 2026, with the largest reported increases in insurance and healthcare, followed by financial institutions and technology, media, and telecom (BCG Data Point Interactive and BCG Analysis, July 2026, n: 5,802).

The phrase matters less than the admission behind it. Leaders accountable for AI budgets are telling investors that the version of AI they intend to scale is supervised, governed, and expert-validated.

For high-consequence industrial use cases, human validation is a core design control. The organizations moving fastest are the ones treating it as structure and designing for it from the first workflow. Industrial operators have strong reasons to prioritize this early, because the cost of an unchecked recommendation is measured in safety, capital, and production.

The agentic shift: from tools to accountable systems

AI agents change the design premise entirely. Instead of asking a user to navigate multiple applications, find documents, reconcile facts, and interpret the answer, an agentic system can be assigned an objective with constraints. It can retrieve evidence, coordinate tasks, reason across sources, propose an action, and show the basis for its recommendation.

The human role does not disappear. It shifts to where judgment matters most: defining the desired outcome, providing operational constraints and acceptance criteria, reviewing high-consequence recommendations, validating quality and safety, and governing access, accountability, and responsible use.

Cao's description of "agentic engineering" is useful here. Humans move from being the authors of every decision rule toward becoming intent architects, coordinators, and outcome auditors. For industrial businesses, the goal is AI that can do the heavy work of searching, synthesizing, comparing, monitoring, and learning, all while keeping domain experts and accountable leaders in control.

A useful principle: agents can execute bounded tasks within defined boundaries, while people retain intent-setting, accountability, and approval authority for high-consequence decisions.

The harder question inside human in the loop: which human?

Once an organization accepts that experts belong inside the workflow, a sourcing problem appears almost immediately. Finding a specialist with the right expertise, available at the moment the workflow needs a decision, is a recurring operational challenge.

The expertise required is specific. Validating an agent's reading of casing wear indicators calls for a drilling engineer who has worked comparable formations. Confirming a proposed mitigation on a subsea tieback calls for someone who has lived through one. Reviewing a corrosion insight from twelve years of maintenance history calls for an integrity specialist who can tell a genuine signal from an artifact of the data.

Most programs discover this mid-deployment. The people with the right judgment are fully committed to live operations, distributed across contractors, partners, and alumni networks, or simply not on the payroll. Validation then becomes the slowest step in the system, and an agentic capability designed to compress cycle time starts waiting on calendars.

This is where ExpertHub plays a significant role. Built by SwarmLens with rp² and Drillers.com, ExpertHub connects a defined requirement to a proven oil and gas professional who has done that exact work. Your experience. Their requirement. One connection.

When validation capacity can be sourced on demand, human in the loop becomes something an organization can plan, schedule, and staff, and the loop closes at the speed the operation actually runs.

Why AaaS is different from SaaS with an AI feature

Agent-as-a-Service represents a genuine shift in service design. In a SaaS model, the customer buys access to a product. In an AaaS model, the provider takes greater responsibility for turning a specific operational objective into a working, managed capability.

That capability may include:

• A roadmap for where AI will create measurable value
• Agent workflows tailored to the client's operating environment
• Secure integration of existing documents and data sources
• Domain-specific context and terminology
• Human expert validation where the risk or ambiguity requires it
• Continuous monitoring and performance measurement
• A recurring service model with ongoing accountability for outcomes

This is where industrial AaaS differs from generic copilots. The valuable agent is grounded in the company's actual operating history. It uses the right evidence. It understands the workflow. It makes its reasoning traceable. And it is backed by named experts who can stand behind what it recommends.

Intelligence without provenance is interesting. Intelligence with provenance can be operational.

A practical roadmap: discover, deploy, learn

The most effective route to agentic AI is a clear sequence from intent to evidence to repeatable value.

1. Identify the decisions that matter. A stronger starting question than "where can we apply AI?" is this one: which decisions are repeated frequently? Which carry costly delays, uncertainty, or rework? Where do teams repeatedly search through documents to reconstruct prior experience? SwarmLens's AccioLens approach reflects this first stage - it evaluates and sequences potential AI use cases against an evidence-based library of more than 1,000 catalogued use cases, calibrated with input from domain experts and scored for fit, delivery duration, and potential return.

2. Deploy agents into the workflow. Once a priority use case is selected, industrial agents should be built around real workflows. In well delivery, a planning agent might connect historical operational reports, pre-well risk registers, after-action reviews, and subject-matter expertise - surfacing comparable events, flagging recurring risks, and showing the exact source material behind each conclusion. SwarmLens's AssureLens is designed around this model. In one Middle Eastern drilling-data pilot, 600 daily drilling reports produced 20,000 raw insights; after deduplication, categorization, and parallel expert validation, 447 entries were retained as validated, source-linked knowledge.

3. Make the system learn responsibly. The real promise of industrial agents is compounding learning. A completed project should improve the next one. A surprise during execution should update the risk picture for the future. That creates a closed loop: Historical evidence → Operational decision → Observed result → Validated learning.

The limits are real, and that is why design matters

The current agentic moment requires ambition, and it also requires honesty. Cao's paper notes that agents can perform strongly on bounded tasks while struggling with long-horizon work, context drift, error propagation, and verification across continuous change. Evidence from the EvoClaw benchmark shows agent performance falling from above 80 percent on isolated tasks to at most 38 percent in continuous settings.

Industrial environments carry analogous risks. An agent may retrieve the wrong context. A technically plausible answer may miss an operational nuance. That is precisely why the best industrial AI architectures should include:

• Source traceability. Every important recommendation links back to evidence.
• Role-based controls. Sensitive data and high-consequence actions require appropriate permissions.
• Human validation. Experts review decisions that affect safety, compliance, capital, or operations - and the program has a dependable way to reach those experts when reviews are due.
• Evaluation metrics. Measure usefulness, accuracy, cycle time, adoption, and realized value.
• Continuous feedback. Capture corrections so the system improves rather than repeats mistakes.
• Clear boundaries. Define what agents can recommend, what they can execute, and where human approval is mandatory.

Where SwarmLens fits

SwarmLens sees Agent-as-a-Service as a practical operating model for industrial AI adoption. The company's role is to help clients move through the full journey:

• AccioLens. Build an evidence-chained AI roadmap around the most relevant and achievable use cases.
• Industrial agents. Deploy agents tailored to the client's workflows and existing data.
• AssureLens. Turn completed projects and operational records into source-linked institutional knowledge.
• UXLens. Customize a unified interface through which teams can use their own enterprise knowledge.
• ExpertHub. Connect teams to the specialist expertise that validation depends on.
• Managed service. Operate and improve the capability as a monthly service.

Taken together, that is a roadmap, an agent layer, an evidence base, and a human layer, managed as one capability. Most companies have no reason to become AI product companies. The advantage comes from getting better at using their own information, expertise, and operating experience.

A leading category of industrial AI is likely to be defined by whether the system can answer a simple, demanding question: can this help our people make a better, faster, more defensible decision, using evidence they can trust? That is the transition from software as a tool to agents as an operating capability. And it is where the conversation about Agent-as-a-Service is only beginning.