The Agent Is the Cheap Part
Owning your agent IP and owning the platform underneath it are separate commitments. Most build business cases only price the first.
The best industrial operators I work with are building their own agents, and they are good at it. They have process engineers who know exactly how their plant fails, and they want that knowledge in their own IP, not in a vendor’s template. None of them want to own the platform they build it on.
McKinsey’s August survey puts a number on the first half of that. [The State of AI in 2026: On the Road to ROI](https://www.mckinsey.com/~/media/mckinsey/business functions/quantumblack/our insights/the state of ai/the-state-of-ai-in-2026-on-the-road-to-roi.pdf) was published by Dan Tinkoff and colleagues at McKinsey, from 1,719 responses collected between 4 May and 8 June. Thirty-two percent of organizations decided against buying at least one software product or feature because they could build the functionality in-house using agentic coding tools. In energy and materials it is 38 percent. Among the companies the survey classifies as AI high performers, meaning they attribute 5 percent or more of EBIT to their use of AI and report significant value from it, it is close to half. For everyone else it is 31 percent.
One word is doing two jobs here. A coding agent writes software for your developers. The agent I am talking about runs on your operation, watches a unit, weighs a trade-off against an objective and asks to act, and it is configured rather than coded. The first is how the second now gets built in a fortnight.
They are right to build. In sales, finance and HR, preconfigured capability is the point. You want your quote-to-cash to work the way everyone else’s quote-to-cash works, because the process is not where you compete and standardization is worth more than fit. Operations does not work that way. A copper mine in Chile does not fail the way a copper mine in Arizona fails, and those two run different ore bodies, different water chemistry, different equipment ages and different crews. The specificity that would be a liability in a CRM is the whole asset in a plant. Preconfigured operational capability makes you look like your competitors in the one place you are genuinely different from them.
None of that is an argument for building what they run on.
The line that survey drew sits at the application layer. Its question asks about software products and features. Nobody in it reports building their own agent substrate, because the substrate is not what got cheap. Coding agents collapsed the cost of writing an application. They did nothing to the cost of owning operational context across a brownfield estate, an authority model that says what an agent may do and who signs for it, coordination between several agents working the same asset, and a decision record that still makes sense to an auditor in 2031. They did nothing about needing somewhere to prove an action is safe before it runs.
The cost arrives after you ship
Eighteen percent of high performers say AI operating costs have constrained their use of software coding agents. For everyone else that figure is 6 percent. Three times the rate, among the organizations doing the most building. The companies furthest into build are the ones already feeling what build costs to run. A build gets approved on the price of construction and then bills every month for as long as the thing operates.
Construction was never the expensive part of industrial software. Ask anyone who has carried a historian integration through two control system migrations, or kept an asset hierarchy true after a debottlenecking project quietly changed the equipment it describes. Your engineers can stand up a working agent in a fortnight now, and I have watched them do it. Keeping trusted context current through every management of change, every model swap, every retirement and every new starter is a different kind of work, and it does not finish.
An unintended action in a plant is a physical event
Thirty-three percent of the organizations the survey did not classify as high performers are working to mitigate unauthorized or unintended action. Twenty-four percent are working on explainability. Physical safety sits at 7 percent for both groups, which says the sample is dominated by offices, and says something about how little of the current agent conversation concerns equipment that can hurt someone.
A wrong answer from a marketing agent is a bad paragraph. A wrong answer from an agent with a write path into a control system is an incident report, a management of change nobody raised, and a conversation with a regulator who will ask to see the decision.
Guardrails are the part teams find out they needed after they have already shipped. That order is expensive in software and unacceptable in a plant.
What separates the high performers is mostly platform
Look at the practices the survey attaches to its high performers. They have worked out where a human has to be in the loop and why, which is a design decision and not a policy statement. They actively manage what their AI solutions cost to run. They have defined processes for quantifying impact, and senior leaders own the initiatives instead of sponsoring them. Every one of those is a property of the platform, not a feature of any agent running on it. A team that decides to build its own substrate has signed up to build these as well, usually without writing them into the business case.
Scaled agent use in manufacturing sits at 3 percent, the lowest function in the survey, with supply chain and inventory management at 5 percent. Set that against the 38 percent of energy and materials companies who have decided they can build their own software, and industrial comes out carrying close to the highest intent to build alongside close to the fewest agents actually running at scale. Some of that gap is caution, and caution in a plant is usually right. Some of it is teams finding out what the substrate costs about three months into building one.
Two questions, asked separately
Stop asking whether to build or buy. Ask two ownership questions instead, and answer them one at a time.
Recommended by LinkedIn
Which parts of this do you want to still own in five years? The decision logic your process engineers encode, the failure signatures particular to your units, the trade-off rules that apply when throughput and reliability pull against each other. Those should be yours, exportable, and not locked inside a vendor’s proprietary format.
Which parts do you want to be responsible for maintaining in five years? The identity and semantics of every asset across the estate, the permission model, the simulation environment, the versioning of every prompt and policy, the audit trail. Ask whether your team wants to own the roadmap for those, and whether owning that roadmap makes the plant run better.
Owning your agent IP and owning the platform underneath it are separate commitments. Most build-or-buy advice treats them as one, which is how a sound instinct about IP turns into a ten-year maintenance obligation nobody costed.
Codified knowledge is the part an auditor can value
Salim Ismail described a case on the Moonshots podcast that bears on the first of those questions. A large business separated its data assets into a subsidiary, spent time cleaning them up, and had its accounting firm value the subsidiary for the balance sheet. The parent was worth around $10 billion. The subsidiary came in at $32 billion. He declined to name the company, so treat the figure as an anecdote and look at the mechanism instead.
An accounting firm cannot put a number on knowledge that lives in prompts on an engineer’s laptop, because there is nothing to inspect. Codified, versioned, explainable and auditable is what moves operational knowledge from something your engineers carry to something an auditor can sign. That is the same property that lets you change a decision policy without breaking the plant, and the same property a regulator asks about after an incident. One discipline pays three times.
Alex Wissner-Gross pushed back on the same episode, and he is right about the limit. Raw enterprise data has a shelf life, which Bloomberg discovered when the next generation of frontier models overtook Bloomberg GPT on its own financial benchmark. A decision record behaves differently, because your operation keeps producing it, every shift, in a form that says what was decided, under which constraints, and who approved it.
We built XMPro around that split, and it sits in the architecture instead of in a slide about shared responsibility. Your engineers define the agents themselves through configuration: the objective each one optimizes, the decision rules it follows, the constraints it works under, the team it belongs to. When an agent needs a capability the platform does not already have, a link into a proprietary optimizer or a physics model or a historian nobody else runs, your developers build that as an integration against our published SDK, version it, and deploy it onto a running system without the agent runtime changing. An agent’s tools are those integrations, called over the message bus, which means the agent process never holds a credential to your control system and every action it asks for crosses the same gate. The authority model, the coordination between agents and the decision record stay in the platform.
An asset nobody can inspect is not an asset, which is why the definitions of your agents and teams live as data in your own databases, in a form you can read without us. Agent cards are emitted in open formats, Asset Administration Shell submodels among them. Every agent carries an SPDX bill of materials naming its profile version, its prompt versions and its tools, so what was running when a decision was made has an answer an auditor can check.
**That boundary has a name: software-defined operations.**The decision and coordination logic is separated from the systems that observe conditions and carry out the response, so it can be changed, tested and governed on its own. Networking went through this twenty years ago and nobody would now buy a switch that made you rewire the building to change a policy. (Software-Defined Networking)
The survey asked a buying question. For industrial operations that was the wrong unit of analysis. The thing to own is the knowledge your engineers put into the agents, and they should own it outright, in a form they can export and take with them. The platform that keeps those agents inside the boundaries the operation set is something else, and the operators getting the best results have worked out that they can have the first without taking on the second.
Pieter van Schalkwyk is the CEO of XMPro, specializing in industrial AI agent orchestration and governance. Drawing on 30+ years of experience in industrial automation, he helps organizations implement practical AI solutions that deliver measurable business outcomes while ensuring responsible AI deployment at scale.
He authored the Industrial AI Agent Manifesto, published by the Digital Twin Consortium in February 2026, a governance framework for trustworthy autonomous operations
Our GitHub Repo has more technical information. You can also contact myself or Gavin Green for more information.
Read more on MAGS at The Digital Engineer
How this was made: I worked with Claude to check sources, draft from my direction and edit. I do not publish anything I have not rewritten myself.
This article originally appeared on XMPro CEO’s LinkedIn blog, The Digital Engineer, on 16 September 2026.
