The AI Surcharge Won’t Last

For the last two years, software vendors have been presented with an extraordinary commercial opportunity. Take an existing application, add AI, create a new licence tier and charge more. In some cases, charge considerably more. The logic has been understandable. Artificial intelligence was new, scarce and expensive. Customers wanted access to it. Vendors had invested heavily in development. AI therefore became something that could be sold as a premium capability.

But that period may prove surprisingly short. The market is already beginning to change. AI is moving from something customers regard as an additional feature to something they simply expect modern software to contain. And once that transition happens, the economics of enterprise software change dramatically.

AI is becoming part of the product.

There is already evidence of this shift. Oracle now provides AI agents embedded within its Fusion Supply Chain & Manufacturing applications at no additional cost to existing Fusion customers. SAP still operates a Premium AI model using AI Units for advanced capabilities and agentic execution, but at the same time includes Base AI within its standard cloud subscriptions and has announced that the majority of its generative AI capabilities are moving into Base AI during 2026. Kinaxis has gone in another direction, introducing Maestro Activity Units as part of its next-generation pricing structure, with usage increasing according to AI tasks, automations, scenarios, calculations and other activity.

So the market is experimenting. Some vendors bundle AI. Some meter it. Some create premium tiers. Some combine all three. But the customer expectation is becoming clearer.

Buyers are beginning to question the AI premium.

AI is no longer enough, on its own, to justify a price increase. G2’s 2026 research found that 81% of buyers expect core AI functionality to be included within the base product, and a 20% AI premium was described as difficult to justify by 58% of respondents unless the vendor could demonstrate clear value. That is an important change. Customers are not necessarily unwilling to pay for value created by AI. They are increasingly unwilling to pay simply because something has AI written on the label. There is an enormous difference.

If AI removes £1 million of inventory, prevents lost sales or eliminates thousands of hours of manual planning work, customers can quite reasonably pay for some of that value. But if the existing planning application suddenly becomes 30% more expensive because a forecasting algorithm or chatbot has been added, the conversation becomes considerably more difficult. AI is rapidly becoming table stakes.

We have seen this before.

Technology almost always follows the same pattern. Initially, a new capability attracts a premium. Eventually, it becomes expected. Nobody now accepts paying a separate surcharge because a business application has a web interface, mobile access, API connectivity, cloud hosting, dashboards, workflow or search. These were once differentiators. Now they are simply software. AI is heading in the same direction. In a few years, describing a planning system as “AI-enabled” may sound as unusual as describing it today as “internet-enabled”. Customers will simply assume that intelligent software contains intelligence.

That creates a problem for traditional software economics.

For large software companies, AI creates an uncomfortable equation. Customers increasingly expect AI, but AI is not free to operate. Traditional SaaS economics were attractive because the marginal cost of an additional software transaction was tiny. AI changes that. Every model call consumes compute. Every agentic workflow potentially generates multiple calls. Every simulation, reasoning loop, optimisation cycle or background agent consumes resources. And the more sophisticated the AI becomes, the greater that consumption can become.

Gartner recently predicted that inference costs per agentic workflow could rise more than fivefold through 2028, despite falling underlying model prices, because agents will perform increasingly complex multi-step work. That sounds counterintuitive. Models are becoming cheaper, yet applications can become more expensive to run. Why? Because developers use more AI.

Cheap tokens do not necessarily mean cheap software.

A poorly designed AI application can consume enormous amounts of compute unnecessarily. Imagine an agent that repeatedly reads a large dataset, calls a language model, asks another agent for validation, retrieves additional context, runs another model, rechecks the answer and repeats the process. The cost of the individual calls may be falling. The number of calls may be exploding. This is becoming one of the most important architectural questions in enterprise AI: how much intelligence does it cost to reach the decision? The winning software companies will not simply have the cleverest AI. They will have the most economically efficient AI.

Supply chain makes this particularly important.

Supply chain systems can process enormous volumes of information. Millions of transactions. Thousands of products. Multiple locations. Long histories. Repeated simulations. Continuous replenishment decisions. Production sequences. Shipment combinations. Inventory calculations. If every decision requires a large, general-purpose model to reason from first principles, the economics can become ugly very quickly.

That is why the architecture matters. Machine learning, optimisation, pattern recognition and deterministic logic should each be used where they are best suited. Not every problem needs a large language model. Not every calculation needs an agent. Not every decision needs thousands of tokens. The cleverest AI architecture may sometimes be the one that uses less AI.

Poor AI architecture could become the new technical debt.

For the last twenty years, software companies accumulated technical debt through old code, legacy architecture and excessive customisation. The next generation may accumulate AI cost debt. A model that appears inexpensive during a pilot can become very expensive when 100 users become 10,000, 100 transactions become 10 million, one scenario becomes hundreds, or an occasional agent becomes a continuous autonomous process.

This matters because customers will increasingly demand predictable pricing. Research into software buying behaviour already shows increasing scrutiny around token consumption, variable pricing and unpredictable AI costs. The problem for the vendor becomes obvious. Customers want AI included. But badly engineered AI may be expensive to provide. Someone has to absorb that cost.

The AI surcharge therefore has a limited life.

There will still be premium AI products. There should be. A system that autonomously performs high-value work can create enormous economic value. But the simple idea that AI equals an additional licence fee is unlikely to survive indefinitely. The market will increasingly separate two things. Basic AI capability will become part of the software. Exceptional business outcomes may command premium pricing. That is a much healthier model. Customers should pay for value, not technology fashion.

This could be uncomfortable for the largest software vendors.

Large enterprise software companies carry large organisations. Thousands of developers. Large sales teams. Large consulting ecosystems. Large product-management structures. Complex legacy codebases. Multiple acquired technologies. Layers of management. Historically, scale was an enormous advantage. Developing enterprise software required armies of people. AI is changing that equation.

Modern development tools allow much smaller teams to build sophisticated applications. And increasingly, the strongest development model may not be hundreds of programmers receiving requirements through layers of product management. It may be a subject matter expert, experienced backend engineers and AI development tools. That combination can be extraordinarily productive. The supply chain expert understands the decision. The engineer understands architecture, security and scalability. AI accelerates the translation between the two. Suddenly, a team of ten can potentially accomplish work that once required fifty or one hundred people.

That does not mean developers disappear.

This is an important distinction. AI does not eliminate the need for engineering. It changes what good engineering looks like. Someone still needs to understand architecture, integration, security, scalability, databases, testing, performance, observability, governance, deployment and supportability. AI-generated software without those disciplines can become dangerous very quickly. But the number of people required to turn an idea into production software may fall dramatically. That creates a structural challenge for software companies built around high development headcounts and expensive delivery models.

Small can suddenly become powerful.

Historically, customers often selected large software vendors partly because smaller suppliers simply could not match their development resources. That assumption is becoming less reliable. A smaller specialist vendor can now combine deep domain expertise, modern architecture, AI-assisted software development, specialised machine learning and a small number of experienced engineers. The result can be faster development, lower overhead and much closer contact between the customer problem and the person building the solution.

That doesn’t guarantee success. Large vendors retain major advantages in distribution, brand, global support, security and enterprise relationships. But the technological barrier protecting them has fallen significantly.

The realignment has already begun.

The next competition in enterprise software will not simply be: who has AI? Almost everybody will. The competition will be: who can deliver useful AI economically? That introduces an entirely different set of competitive advantages. How quickly can the vendor develop? How close are the developers to the subject matter? How efficiently does the AI run? How much compute is required per decision? Can the product be deployed modularly? Can customers start small? Can the AI work with the existing system landscape? And critically: can the vendor provide AI without turning every intelligent feature into another expensive licence? Those questions favour a different type of software company.

The O8 perspective.

At O8, we believe AI should make supply chain software better and more accessible, not simply more expensive. Our approach is therefore modular. Forecast Intelligence can improve the demand signal. Organic Planner can improve ordering and replenishment decisions. AI Shipment Builder can optimise shipment construction. Production Sequencer can improve production scheduling. Network Visualisation and the Organic Planner Simulator can support disruption and scenario planning. Self Serve and KPI & Reporting provide analytical and management visibility around those decisions. Customers should be able to adopt the capability they need without replacing their entire technology landscape.

Equally importantly, AI should be applied appropriately. Machine learning where patterns need to be learned. Optimisation where alternatives need to be evaluated. Pattern recognition where data needs to be repaired. Deterministic algorithms where deterministic logic is the most efficient solution. Generative or agentic AI where those technologies genuinely add value. The objective is not to maximise AI consumption. It is to maximise the quality of the decision.

Value per decision will matter more than price per token.

For a while, the AI industry has obsessed over token prices. That is probably the wrong metric. A cheap model that takes fifty steps to reach an answer may be considerably more expensive than a specialised algorithm that reaches the answer in one. The better commercial metric will ultimately be: what did it cost to make the decision, and what value did that decision create?

That is particularly important in supply chain. If an AI planning engine saves £10 million in inventory while costing £50,000 to operate, nobody cares how many tokens it consumed. If it saves £50,000 while generating £500,000 of inference costs, the architecture is broken.

The next software winners may look very different.

AI will not destroy enterprise software. It will reshape it. Customers will expect more intelligence for less money. They will become less tolerant of expensive development cycles. They will question AI surcharges that cannot be linked to measurable value. And they will increasingly compare large traditional platforms with smaller, highly specialised alternatives that can develop faster and operate with lower overhead.

That does not guarantee that every small vendor wins. It does mean that size alone is becoming less of a defence. The winners will be companies that combine domain expertise with strong engineering, efficient AI and low organisational friction. Those companies may be much smaller than the software giants they compete against.

AI is not just changing software. It is changing the economics of the software industry. The first phase of enterprise AI was about capability. Can AI forecast? Can it create content? Can it analyse? Can it make recommendations? Can it act autonomously?

The next phase will be about economics. How much does it cost? Who pays? Is the cost predictable? Does the AI create enough value to justify it? And why should the customer pay more simply because the vendor has finally modernised its software? Those questions will become increasingly uncomfortable.

Because AI is rapidly moving from premium feature to expected capability. And once that happens, the companies carrying the highest costs may have the hardest problem of all. They will need to deliver more intelligence. At lower cost. With less friction.

That is not simply an AI transformation. It is a software market realignment.