From Asset Management to Asset Governance 

It’s 04:12 on a Tuesday and vibration signature on a boiler feed pump crosses a threshold that a physics model and a machine learning classifier both agree is not noise. The system does exactly what it was built to do. It raises an advisory, ranks it against the 200 other open advisories, writes a tidy summary in plain English, and puts it in front of a reliability engineer who logs on at 07:30. 

They read it in ninety seconds and agree with it. Then the real work starts. 

They take it to planning. Planning asks whether it can wait for the shutdown in six weeks. Production says absolutely not, that unit is carrying the quarter. Maintenance says the spare is in a warehouse 400km away and the last one took eleven days. Safety wants a hot work assessment because of the adjacent line. Finance wants to know why we are spending unbudgeted money in month nine. Somebody books a meeting for Thursday. 

By the time a work order has a date, a crew and a permit attached, it can already be the following Monday. The initial detection may have happened almost instantly, but getting the organisation into a position to act took six days. I’ve watched versions of this play out at more sites than I care to count, and detection was rarely the bottleneck. 

Wait, don’t we have systems for that? 

Let me be fair to the industry first, because the last fifteen years have delivered real engineering. Physics-based degradation models that understand how a bearing fails rather than just noticing that it has. Unsupervised anomaly detection for the very common case where you have no failure history because the thing has never broken. Hybrid approaches that constrain a neural network with thermodynamics so it stops proposing physically impossible answers. A generative layer that turns a wall of spectra into three paragraphs a planner can read between meetings. All of that is genuine. 

But look at where it is pointed. Every one of those capabilities makes the same thing happen faster. They help a competent human get to an accurate, well-argued problem statement sooner. 

That is not an accident of product design, it is written into the standard. ISO 13374 defines six functional blocks for condition monitoring and diagnostics, and the last one is called Advisory Generation. Its job is to “provide recommended actions and alternatives and the implications of each.” There is no seventh block. There is no Act. The architecture we have all been building against for twenty-odd years formally terminates at the point of giving advice. 

So when a vendor says they have added AI to asset performance management, what they usually mean is better algorithms in blocks three, four and five. Sharper prognostics, prettier advisories. I have helped build some of this and I am not going to pretend it has no value. I just do not think it is transformative. We made the first half of the loop extremely clever and left the second half exactly as we found it: a chain of humans forwarding a PDF to each other while waiting for diaries to align. 

The reframe 

The moment that advisory lands in a human inbox, it exits the digital world. Once it has exited, everything downstream runs at the speed of calendars. 

What if it didn’t have to leave? 

What I am proposing is not another algorithm bolted onto detection. It is a change to what happens in the ninety seconds after it. 

Picture a standing panel of specialist agents. Not one general purpose assistant with a big prompt, but a group of them, each trained on how one part of your organisation actually thinks. Safety, which knows your permit regime, your isolation standards and your last three incident investigations. Reliability, with your failure modes and criticality matrix. Maintenance, with crew skills, backlog and contractor rates. Production, holding the plan and the customer commitments behind it. Environmental, with your permit limits. Energy, because a degraded pump costs you money long before it costs you an outage. Finance. Cybersecurity, because someone has to ask whether that firmware update opens a door. 

When the pump advisory fires, these agents get it first. Not the human. And they argue about it. 

They argue the way Thursday’s meeting would have argued, except they already hold all the procedures, constraints and precedents, and they can run it in under a minute. Safety objects to the proposed window. Production counters with the shipment. Finance points out that the deferred option costs more across the fiscal year than acting now. Maintenance says the spare can be air-freighted for £4,000 and asks who wants to justify that. They iterate until they reach a quorum, or until they establish that they cannot. 

Then they bring a plan to the operator. Not an alert. A plan, with the reasoning attached, the trade-offs made explicit and the dissent recorded. 

Depending on how much trust you place in that panel, either the operator approves it, or the operator is simply informed and the panel gets on with it. Work orders raised, materials reserved, production forecast adjusted, financials updated, permit request queued, shift handover annotated. 

I’m calling this an AI Tribunal, mostly because the alternatives were worse. A committee sounds like it will still be arguing on Thursday. A swarm sounds like something you call pest control about. Tribunal at least implies a duty to reach a finding and to record why. 

The disagreement is the point 

Most organisations will instinctively want the agents to agree, because consensus feels like confidence. I would argue the opposite. The dissent is the most valuable thing the tribunal produces, and the consensus is the least interesting part of it. 

There is decent research support for the general shape. The 2023 multi-agent debate work from MIT and Google DeepMind found three agents debating over two rounds beat a single agent substantially across arithmetic, maths, chess and MMLU. The more useful finding sits in the same paper: single-agent self-reflection made factual accuracy worse on two benchmarks while debate improved it. Being asked to check your own work is not the same as being challenged by somebody with a different mandate, we can all personally relate to that in real life, so why not in the digital world. 

Now the honest half, because the evidence against naive multi-agent architectures is substantial and growing. The Berkeley paper “Why Do Multi-Agent LLM Systems Fail?” measured production frameworks failing real tasks at rates between 15% and 87% and catalogued fourteen distinct failure modes. A run of 2025 and 2026 studies have documented sycophancy propagation, where models abandon correct answers because a peer disagreed, and consensus collapse, where the right answer was in the pool and got voted out. One 2026 study measured peer-answer adoption at up to 85.5% while burning two to three times the tokens for equal or worse accuracy. 

If you build a tribunal where every agent is the same model with a different job title in the prompt, you will not get deliberation. You will get an echo chamber with a governance veneer, which is worse than no tribunal at all, because it launders one model’s opinion through something that looks like a process. 

So the design constraint is adversarial by necessity. The safety agent has to be structurally incapable of being talked round by the production agent, with objections grounded in a documented standard rather than a persuasive argument. Disagreement needs to be a valid and common output, and an unresolved conflict has to escalate rather than get averaged away. Bluntly, I want a system that says “we cannot agree, and here is exactly why” more often than one that always produces a plan. 

What we do not have yet 

Some of the gaps are technical. The harder ones are not. 

Your procedures are not machine readable. We have spent a decade building digital twins of equipment and almost nobody has built one of their own decision rules. The criticality matrix lives in a spreadsheet, the approval thresholds live in a policy PDF, and the real deferral logic lives in the head of a planner who has been there nineteen years. A tribunal is only as good as the operating context each agent holds, and most organisations cannot export that context in any usable form. That, far more than the model layer, is the actual work. 

Write-back also crosses a boundary that exists for good reasons. NIST SP 800-82r3 is explicit that traffic between the enterprise level and the operations management level should pass through services in a DMZ, and under IEC 62443 any such flow has to be modelled as a conduit with its own risk assessment and target security level. An enterprise-layer agent autonomously writing downward into operational systems inverts the normal direction of trust. Solvable, but it is a security design exercise, not a feature flag. 

There is no ratified standard for agent identity, delegated authority or tamper-evident audit trail. There is real activity: OAuth work on on-behalf-of authorisation, SCIM work on an AI agent resource type, a SCITT profile draft for agent execution records. Every one of those is still an Internet-Draft, none is an RFC. Meanwhile OWASP’s agentic threat taxonomy, now at seventeen threats, lists Repudiation and Untraceability as a first-class concern. If you cannot prove after the fact which agent said what, on what evidence, under whose authority, you do not have a governance system. You have a very expensive opinion generator. 

Functional safety has no route to certify a probabilistic component. IEC 61508 treats software failure as systematic rather than random, so you cannot claim a failure rate for it, only a systematic capability earned through a complete and unambiguous specification and graded fault-avoidance techniques. A language model has no bounded output space and no deterministic mapping from input to output. ISO/IEC TR 5469:2024 covers AI and functional safety and is worth anyone’s time, but it is a Technical Report. Guidance, with no certifiable requirements in it. That is a structural incompatibility rather than a maturity gap, and it is why the tribunal has to sit firmly outside the protection layers. 

And nobody has decided who is accountable. ISO 55001:2024 builds its accountability model on natural persons holding assigned authority and demonstrable competence, and offers no route for a non-human actor to hold authority. OSHA’s management of change rule at 29 CFR 1910.119(l)(2)(v) is more interesting than people expect: it requires your written procedure to address authorisation requirements, but names no person, title or role as the authoriser. It is silent on automation because in 1992 the question was absurd. So the blocker is not the regulation. It is that no MoC procedure written anywhere contemplates a non-human signatory, and no agent can discharge the separate duty to inform and train affected workers before start-up. 

On liability, the EU has moved. Directive (EU) 2024/2853 must be transposed by 9 December 2026 and brings software squarely into product liability. Article 7(2)(c) makes a product’s “ability to continue to learn or acquire new features after it is placed on the market” a criterion of defectiveness, and Article 10(4) requires a court to presume defectiveness or causation where a claimant faces excessive difficulty due to technical complexity. Read those together while picturing an opaque multi-agent deliberation nobody can reconstruct. The proposed AI Liability Directive, meanwhile, was formally withdrawn in October 2025 and nothing has replaced it. 

For what it is worth, the EU AI Act is probably not your constraint. Annex III point 2 catches AI used as a safety component in critical infrastructure, and maintenance optimisation generally sits outside it. Regulation (EU) 2026/1744 pushed the high-risk obligations out to December 2027 and August 2028 anyway. That is not a free pass. It means the binding constraints are your own asset management system, your own MoC procedure and your own security architecture, which is harder than compliance because nobody hands you the checklist. 

Is industry ready? No. But that is not the interesting question. 

The adoption data is not flattering. Deloitte’s 2026 enterprise survey found close to three-quarters of companies planning to deploy agentic AI within two years and only 21% reporting a mature model for agent governance. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and separately predicts 40% of enterprises will demote or decommission autonomous agents by 2027 because governance gaps get found only after a production incident. 

So no, most organisations are not ready, and I would not encourage anyone to jump to autonomous execution on safety-relevant assets. 

But the “we would never let software decide” objection does not survive contact with our own history. A safety instrumented system under IEC 61511 takes irreversible action to bring a process to a safe state with no human in the loop, precisely because human reaction time is not reliable enough. Closed-loop multivariable control has run refinery units for thirty years. Market-wide circuit breakers halt an entire equity market automatically at defined thresholds. Autonomous haulage has been moving ore in the Pilbara since 2008. 

My favourite example is fault location, isolation and service restoration on the distribution grid. The US Department of Energy studied 266 FLISR events across ten operating companies and found a 55% reduction in customers interrupted per event. The finding that should give every human-in-the-loop advocate pause is this one: fully automated switching and validation typically produced greater reliability improvements than operator-initiated remote switching with manual validation. Automated actions took under a minute, manually validated ones five or more. The price of that autonomy was extensive up-front validation and calibration, which is exactly the right price. 

We are not squeamish about machine authority in general. We are squeamish about machine authority we cannot bound, cannot audit and cannot switch off. 

There is also a trap in the middle that I think is badly under-appreciated. Gartner’s autonomy framework describes a level where an agent executes only after explicit human approval for every action, and warns that “approvals can degrade under time pressure or approval fatigue, creating a false sense of safety while expanding the attack surface.” OWASP calls the same failure Overwhelming Human-in-the-Loop and lists it as a named threat. Put a human approval gate on 4,000 agent decisions a month and you have not built oversight, you have built a rubber stamp with a job title. You have also made the audit trail worse, because now there is a name against a decision nobody actually read. 

The choice is not autonomy versus oversight. It is meaningful oversight of a bounded system versus theatrical oversight of an unbounded one. 

What actually changes 

If you accept the argument, the deliverable is not a better Asset Management System. It is something I would rather call an Asset Governance System, and the difference is not cosmetic. 

An asset management system records what happened and what is planned. Work orders, notifications, history, costs. Its central asset is the transaction. 

An asset governance system records what was decided and why. Its central asset is the deliberation. Every intervention carries the reasoning behind it, the constraints considered, the objections raised and by whom, the trade-off accepted, and the authority it was executed under. Ten years later, when someone asks why that pump was deferred in the third quarter, the answer is not an archaeological dig through three inboxes and a departed planner’s memory. It is right there, in structured form, with the dissent intact. 

That is the part that makes me think this is worth doing even if you never switch on autonomous execution. Most organisations cannot reconstruct why a decision was made six months ago, and we have collectively decided that is normal. 

You would also need things we do not build today. Guardrails as code rather than as policy, circuit breakers that halt the tribunal on threshold violation, rapid rollback, clear ownership of agent behaviour, and a defined autonomy level per asset class, so a critical safety-related asset never leaves advisory mode while a low-criticality utility pump might reasonably operate autonomously inside a spend limit. 

Over to you 

I am fairly convinced the detect-faster race has run its course, and that the next real step is moving the argument inside the machine rather than getting the alert to the front of the queue faster. I am much less certain how you get there safely, and I would rather be argued with than agreed with. So, three things I would like other views on. 

Would you let a panel of agents raise a work order without your sign-off? On which asset class, at what spend limit, and what would you need to see first? 

Is the deliberation record enough of a prize on its own to justify the build, even with execution left switched off? 

And who signs it? Not in the abstract, but on your site, on your MoC form, next to the box marked authorised by. Whose name goes there when nine agents deliberated and one of them objected? 

I do not think that last one has a clean answer yet. I think we should be discussing it now, rather than after the first incident report makes it urgent.