.png){kind=logo}
A CNC machine on a factory floor and a SaaS project management tool are both targets for AI automation. But they are not the same kind of target, and treating them as such leads to poor decisions, both for teams building software and for those making strategic bets about where AI creates real competitive advantage.
AI agents crossed a threshold that the software market noticed fast: they can now execute multi-step business workflows end to end, without a human in the loop and without holding a seat license. Software investors reacted and coined the term SaaSpocalypse to describe the sell-off that followed. The reaction made sense for a narrow category of software. For a manufacturer running tight-tolerance production lines, or an energy company managing grid operations under regulatory oversight, the picture looks very different.
This article builds a framework for evaluating which workflows and industries AI genuinely disrupts, which it does not, and what the difference means for engineers and technical leaders making decisions today.
SaaS was built on a simple premise: take a workflow someone was doing manually, wrap it in a database and a UI, and charge per seat. For two decades that worked because the bottleneck was human time and attention. It starts to break when an AI agent can execute the same workflow end to end without holding a license.
Three structural weaknesses made horizontal SaaS the first category to feel this:
Per-seat pricing had no floor. One agent can handle what ten users were doing before: creating tasks, updating records, generating reports, sending sequences. The license count drops but the vendor's cost base does not.
Most horizontal SaaS held little proprietary data. A generic project management tool knows when tasks opened and closed. It does not know your industry, your compliance requirements, or how your team makes decisions. An agent connecting directly to your own systems can often replicate the core workflow without the vendor's product in the loop.
The UI was the product for many vendors. SaaS companies competed hard on interface quality. When an agent becomes the interface, that investment counts for nothing. The agent reads a schema or calls an API.
Those three factors combined to make generic horizontal SaaS fragile at exactly the moment AI agents became capable of multi-step autonomous execution.
Not all work is equally automatable, and not all industries sit in the same position. Five questions help clarify the picture: How available is the data, and what does it cost to collect it? Does the work require exact outputs or judgment calls? How much regulatory oversight constrains what can be automated? How much physical infrastructure sits between an AI output and a real-world outcome? And how deeply embedded is the current system?
In software, the data needed to automate most workflows already sits in structured form inside the tools themselves. In manufacturing, it lives in sensor outputs tied to proprietary industrial protocols, or inside legacy systems never designed to expose data externally. High availability speeds things up. Fragmented or expensive-to-collect data slows it down.
Some work requires one correct answer: a payroll figure, a circuit breaker state, a parts tolerance on a production line. Other work involves judgment: drafting a status update, triaging a support queue, ranking leads. AI agents handle judgment-based work well when the cost of an imperfect output is low. Deploying them on high-stakes exact work is a categorically different decision.
In regulated environments, specific decisions must come from licensed practitioners, get logged in particular ways, and pass review before taking effect. AI in a regulated industry looks like decision support and documentation assistance, not autonomous execution.
Changing a software workflow means changing code. Changing a physical workflow means changing equipment, facilities, and supply chains. The capital needed to physically integrate AI into a manufacturing plant or a power grid far exceeds the cost of routing an API call through an agent. Physical industries face longer timelines because capital is the constraint, not willingness.
An ERP customized over a decade, integrated with dozens of systems, and embedded in the habits of an operations team does not get replaced without significant cost. A standalone task manager with shallow integrations is a different situation entirely.
Generic horizontal SaaS sits at the worst position across all five factors: data easy to read and replicate, probabilistic workflows, light regulation, no physical assets, and low switching costs for anything without deep integrations.
The most exposed products are those whose value is organizing information that already exists somewhere else: basic task trackers, standalone email marketing platforms, generic support queues, simple dashboards. The most insulated are those where the data is the product: security platforms trained on years of environment-specific signals, HR systems with multi-jurisdiction compliance logic, ERPs with a decade of customization.
Salesforce is a useful case. Rather than defending its dashboard layer, the company rebuilt around agents that operate on CRM data. Agentforce ARR reached $800 million by end of FY2026, up 169% year on year, with 29,000 deals closed. By Q1 FY27 that figure crossed $1.2 billion, up 205% year on year. The bet: the data and execution layer is worth more than the interface.
AI exposure in retail concentrates where outputs are digital: demand forecasting, search ranking, personalized recommendations, customer service routing. Real applications with measurable returns. But retail does not share software's absence of physical infrastructure. A demand model only creates value if the logistics network can act on its outputs. The bottleneck is physical execution — warehouse throughput, supplier relationships, last-mile capacity — not analytical intelligence. AI optimizes the existing operation. It does not replace it.
Power grids generate large volumes of sensor data from substations, transformers, and generation assets. Most of it sits in operational technology systems built for isolation, not integration. Connecting IT and OT in a safe, auditable way takes years of engineering work. Grid switching, demand response, and load balancing operate within safety protocols that regulators will not accept AI executing autonomously without extensive validation. The near-term value is in decision support: predictive maintenance on equipment, anomaly detection, surfacing better information to engineers before they act.
A CNC machine cutting a component to tolerance has no acceptable error margin. A welding path that deviates two millimeters produces a defect. These are not workflows where approximate outputs work. AI returns in manufacturing show up in adjacent workflows: computer vision for quality inspection, predictive maintenance, supply chain risk modeling. Genuine applications, but not a transformation of production itself. That requires physical reconfiguration, and physical reconfiguration requires capital and time.
Map your workflows against the five factors. For each significant workflow, ask where it sits on data availability, work type, regulation, physical assets, and switching costs. The answer tells you whether AI is a near-term risk, a near-term opportunity, or a longer-horizon consideration.
Separate task automation from workflow transformation. Automating a task inside a workflow does not transform the workflow. Real transformation changes who is responsible for an outcome, not just who performs a sub-task. MIT Sloan research on this distinction is worth reading before drawing conclusions about headcount.
Do not copy strategies from structurally different industries. The response that makes sense for a generic SaaS company facing agent-driven competition does not transfer to an energy utility or a discrete manufacturer.
Build on advantages that take time or capital to replicate. Proprietary data, deep integrations, regulatory expertise embedded in tooling, physical infrastructure built over years. AI amplifies these advantages because better data produces better outputs from the same models.
Start where AI already has a track record. Predictive maintenance in industrial settings. Demand forecasting in retail. Documentation in regulated industries. Starting there builds capability and generates real evidence about what works in your specific environment.
Treating "AI can do this task" as the same as "AI will replace this workflow." The task and the workflow are different units. Removing one task from a role that contains many decisions and relationships is not the same as replacing the role. This MIT Sloan piece explains the distinction clearly.
Prioritizing technical possibility over operational context. Technical feasibility and operational viability are different claims. Regulatory and safety constraints do not move fast, and they exist for good reasons.
Building AI features any competitor can replicate quickly. If an implementation relies entirely on a public model and public data, it is a feature, not a moat. Sustainable positions come from applying AI to data and processes specific to the organization.
Responding to urgency without a clear thesis. Urgency is not a strategy. Deploying AI wherever it fits, without a clear view of where value comes from, produces technical debt rather than competitive advantage.
The SaaSpocalypse identified a real structural problem in software: per-seat pricing, thin proprietary data, and shallow integrations make generic horizontal SaaS a direct target for agent-based automation. That problem is not going away.
But the structure that makes software vulnerable does not exist in most other industries. Retail, energy, and manufacturing each carry physical infrastructure, regulatory constraints, and data complexity that slow AI-driven change in ways that have nothing to do with willingness to adopt.
The right question is not whether AI will transform your industry. It is whether the structural properties of your industry and your specific workflows make that transformation near-term, medium-term, or much further out — and what to do differently in each case.
