The technology press has spent this week dissecting SAP’s decision to block external AI agents from accessing its systems, requiring everything to pass through Joule. It is a genuinely significant architectural bet, and the debate around it is worth following. Salesforce, meanwhile, is moving in the opposite direction, opening its platform to agents and doubling down on monetised digital labour through Agentforce. Two of the world’s largest enterprise software companies, making opposite calls on where the value in AI sits. 

It is an interesting fight. But a few days ago, I was in a conversation with a managing director at a leading European food processor, and I can tell you it is not the question his business is asking. 

His question was more fundamental. And I think it points to something most of the current enterprise AI conversation is missing entirely. 

A More Contested Market Than the Headlines Suggest 

Before getting to that conversation, it is worth stepping back from the SAP-Salesforce frame, because the market facing industrial businesses is considerably more complex than any two-vendor narrative captures. 

Large industrial organisations are being approached from multiple directions simultaneously. The ERP incumbents embedding AI into process workflows. The CRM-led platforms building agentic layers on top of customer engagement. Data and analytics specialists with strong track records in defence and government intelligence, now pushing into broader industrial markets. And the build-your-own path, where a business takes a frontier model from one of the hyperscalers, wraps it with consultancy support from a large professional services firm, and attempts to construct something bespoke. 

Each of these approaches has genuine merit in the right context. Each will find markets. 

For the industrial businesses that run the physical world, though, most of them start from the wrong place. And the reason why matters more than which vendor wins the agent access debate. 

Two Credible Paths, One Shared Blind Spot 

SAP is doing what SAP does well. Embedding AI into enterprise process, finance, procurement, HR, supply chain orchestration, layering intelligence onto decades of transactional data and business process context. For organisations where SAP is the operational backbone, this is a coherent strategy. The data is there, as is the process context and integration logic. Regardless of the current debate around Joule and agent access, that foundation is substantial. 

Salesforce is building something different but with the same outcome in mind. Their bet is on AI agents as monetised digital labour, operating across customer engagement, sales execution, and service workflows. Agentforce executed well is a serious move, building a new commercial layer on top of CRM where agents become billable units of work. It will resonate strongly with many customer-facing organisations, and the early adoption numbers suggest the market is responding. 

Both paths are credible, both will build revenue-streams but neither was designed for the operational realities of asset-intensive industry. More importantly, neither starts by asking what the right problem actually is. 

The Conversation That Reframed This for Me 

Back to the food processor. On paper his business looks like a textbook candidate for industrial AI: complex assets, high production volumes, significant maintenance costs. We worked through the numbers together. In his specific context, better predictive maintenance and uptime would free up somewhere in the low hundreds of thousands in additional cash flow. Real money, and the kind of outcome most vendors lead with… but nowhere near his actual problem. 

His real pressure is absorbing input cost inflation he cannot yet pass on to customers, across supply chain relationships he cannot afford to lose. Feed costs and fertiliser costs are up significantly, with genuine shortages in both right now. Labour costs in this very physical industry are rising while availability is falling. The price of food has not caught up with any of this yet, and he is wearing that gap himself. 

He put it plainly: the unit economics of any individual decision matter less right now than whether the relationships exist to persist through difficult times. 

Most vendors would have walked out of that meeting having attempted to sell him something useful but peripheral. They walked in with the answer before they understood the question. 

So the conversation shifted to something more commercially urgent. Where is the optimal place to buy stock? What does optimal actually mean? When and where to sell? How do you use better intelligence to strengthen supply chain relationships rather than just transact through them? How do you build the kind of trusted network that gives you predictability when the market is this volatile? 

Those are hard, commercially critical problems, and they are all AI-addressable. You only arrive at them if you understand enough about how this industry works to know that two-way supply chain reliability sits above everything else for this business right now. 

Getting to the right problem requires understanding the business, not just the data. 

Why Depth Beats Abstraction 

The second conversation that sharpened this was with a long-term contact in offshore drilling. His operational costs are being squeezed directly by global fuel prices and the pressure on every decision is real. 

Like many operators in his sector, his organisation has outsourced a significant portion of operations, including managing asset records, maintenance planning, and associated complexity to vendors, OEM’s and suppliers over the years. Sensible at the time. But with the capabilities now available, he is asking whether it makes more sense to insource some of that intelligence and own it more directly. 

That is only a viable path if the technology vendors supporting them understands the operational and commercial logic behind the data, not just the data itself. What matters in offshore asset management is specific to that environment. The sub-industry context differs from utility maintenance, from aerospace MRO, from food production. And the business-level context on top of that is more specific still. 

His view, which I share, is that technology vendors risk losing context at two levels simultaneously: the sub-industry layer and the reality of how this particular organisation actually operates. A vendor at too high a level of abstraction fails at both and every business is racing toward the need to act. 

The same challenge applies to the build-your-own path. Taking a frontier model and layering domain knowledge on top sounds straightforward. The domain knowledge required to act intelligently in these environments takes years to accumulate. A model advising on aircraft maintenance needs to understand airworthiness directives, component life tracking, and the scheduling constraints of a live flight operation. A model supporting a utility’s field workforce needs to understand network topology, safety isolation procedures, and how weather events cascade through asset risk. A model supporting a defence contractor needs to understand platform readiness, through-life support obligations, and what sovereign compliance actually requires in practice. 

That is not knowledge you can prompt-engineer your way to. It has to be built. 

Explainability as a Baseline 

There is a dimension to all of this that does not get enough attention in the current vendor debate. 

In a manufacturing plant, a field operation, or a critical infrastructure environment, AI recommendations exist within regulatory frameworks, safety protocols, and operational accountability structures. When an AI system recommends taking an asset offline for early maintenance, rescheduling a field crew, or adjusting a supply chain position, someone with operational accountability needs to understand the reasoning behind it. 

“The model recommended it” does not work in these environments. 

Industrial AI needs to be explainable as a baseline. The maintenance engineers, service dispatchers, and operations managers relying on it daily need to trust what it tells them. That trust comes from transparency, from context, and from AI that operates within the real constraints of the business. Whether an agent accesses a system through Joule or an open API is a secondary concern compared to whether the engineer on the shop floor or the offshore platform trusts what it tells them. 

What This Looks Like in Practice 

This is where I think the conversation needs to move from argument to evidence. 

At William Grant and Sons, the world’s largest independent distiller, IFS Nexus Black deployed Resolve, an AI system built with IFS domain models, the operational context, the asset-specific training, with underlying model capability provided by Anthropic, directly into their maintenance operations. When the team went on site for the initial assessment, 38 percent of all maintenance work was emergency or corrective repair. Faults were only being caught after alarms fired. The team was running to fix things that could have been anticipated. Resolve reads video, audio, pressure and temperature data alongside complex plant schematics to identify failures before they happen. The distillery now projects savings of £8.4 million a year at a single site. Their Chief Technology and Business Officer Badri Narasimhan put the philosophy clearly: wherever they deploy technology, it has to be led by a business problem, not technology for its own sake. 

At Kitron, the global electronics manufacturing services company operating across defence, aerospace, medical devices and offshore marine, the challenge was different. Supply chain workflows including inventory replenishment and supplier coordination were consuming significant operational bandwidth. IFS Loops Digital Workers now automate those workflows, with early shortage prediction protecting production schedules before problems occur. Jonatan Gustafsson, their Business Application Manager, described it simply: with structured operational data already in place, they can apply AI where it matters most. 

Two different industries. Two different problems. Both requiring genuine understanding of the operational environment before a line of AI was written. 

Commercial Model as Strategic Signal 

There is one more dimension worth raising, because I think it goes to the heart of what this moment in enterprise AI actually requires. 

In April we announced a fundamental change to how IFS prices its industrial AI. We moved away from per-user licensing entirely, to a model based on the assets a business operates. An energy company managing 400 offshore assets now pays based on those assets, not on the 12,000 people and machines that need to access the data. As our CEO Mark Moffat put it: we are not pricing the workers. We are pricing the work. 

That shift is more than a commercial decision. It is a signal about what we believe industrial AI should actually do. Per-user pricing made sense when software was a productivity tool for individuals. When AI is embedded in operational workflows, driving outcomes across maintenance planners, control-room operators and field technicians simultaneously, tying cost to headcount is the wrong measure. Cost should align with operational reality, and with the value the system actually creates. It is the same logic that runs through everything else in this article. Start with how the business actually operates and build from there. This also mitigates many of the very real concerns about scaling costs of AI native technologies.  

The Debate Worth Having 

The SAP versus Salesforce agent access story will run for months. It is strategically important for the vendors involved and for the architects evaluating platform lock-in. That debate is worth having. 

The industrial businesses running the physical world, keeping planes flying, networks live, energy flowing, and critical assets operational, are asking a different set of questions. Which AI actually understands my industry? Which vendor knows enough about how my business works to tell me where the real value is? Which technology partner will still be relevant to my operational problems in ten years, not just my platform architecture decisions today? 

The next era of enterprise software will not be defined by which vendor wins the agent access debate. It will be defined by which vendors genuinely understand the operational realities of specific industries and have built AI capable of acting intelligently within them and framing the commercial model appropriately with a shared value framework. 

Deep domain expertise built over years cannot be generated overnight. That is where IFS is investing. That is the category we are committed to leading.