There is a familiar rhythm to new technology cycles. First comes the breakthrough. Then comes the conviction that everything will change immediately. And then, usually, reality steps in.
AI feels very much in that middle phase right now. The capabilities are real. In some cases, they are extraordinary. But the diffusion of those capabilities into everyday business operations is likely to take longer than the loudest voices in tech would have us believe.
From where we sit, that is not a sign of failure. It is simply what happens when a powerful new capability meets the messy structure of the real world.
A lot of the public conversation about AI still treats adoption as if it were mostly a matter of model quality. If the models get smarter, the thinking goes, businesses will naturally reorganize around them. But in practice, intelligence is only one layer of the problem. The harder layer is integration. The slowest layer is trust.
That is especially true in enterprise environments.
A startup can often decide, almost overnight, to rebuild a workflow around AI. It may have a simpler architecture, fewer compliance obligations, and less institutional memory to wrestle with. It can tolerate some ambiguity because speed is part of its advantage.
A large enterprise operates differently. It has systems that were designed over decades, not quarters. It has processes shaped by audit requirements, approval chains, procurement rules, internal politics, data boundaries, and a long history of “temporary” technical decisions that became permanent. In that setting, inserting an AI agent into a workflow is not like adding a helpful assistant to the room. It is more like introducing a new actor into a tightly coupled system where every connection matters.
That is one reason AI diffusion will be uneven. The limiting factor will often be less about whether a model can perform a task and more about whether an organization can let it do so safely, reliably, and at a cost that makes sense.
The cost question is one of the least glamorous and most important parts of this story.
For the next few years, engineering compute budgets are going to become a serious operational topic. Not just for AI-native companies, but for almost any business that wants to use AI at scale. There is a tendency to think of software as something that becomes cheaper as it scales. Traditional software often rewards that assumption. But AI systems can introduce a different cost structure. Every agent action, every inference, every orchestration step, every retrieval call, every loop through a workflow can carry marginal cost.
That changes the conversation.
Teams will need to think not just about headcount budgets, but about token budgets, inference budgets, and the economics of autonomy. At a small scale, these costs can feel negligible. At enterprise scale, or at the scale of hundreds or thousands of agents operating continuously, they become architectural concerns.
This is where the discussion gets especially interesting. We may need to design software with the assumption that there could one day be far more AI agents than human users. Maybe a hundred times more. Maybe a thousand. If that turns out to be true, many of our current interface assumptions begin to look oddly human-centric.
Most software today is built for people sitting in front of screens. It assumes visual navigation, manual review, and relatively low-frequency interaction. But agents do not need beautiful dashboards in the same way people do. They need clear affordances. They need stable APIs, structured outputs, transparent permissions, machine-readable documentation, sensible failure states, and environments where they can act without improvising themselves into trouble.
That may sound technical, but it has real business consequences.
When a system is easy for a human to use but difficult for an agent to navigate, the organization will not get the full benefit of automation. The quality of the interface becomes part of the economic value of the software. In that sense, APIs, command-line tools, event systems, and other structured interfaces are no longer just implementation details. They are becoming strategic surfaces.
This is also where older enterprise systems create tension. A company may have invested heavily in platforms like SAP, Salesforce, Oracle, ServiceNow, or dozens of internal tools stitched together over time. These systems often contain the logic of the business. The workflows are there. The approvals are there. The data is there. The domain expertise is embedded there, even if imperfectly.
So when people ask whether AI will replace software categories, a more immediate question is often whether AI can work with the software categories that already exist.
In many cases, the first wave of meaningful adoption will come not from replacing core systems, but from finding ways for agents to operate across them. That sounds straightforward until you confront the details. Permissions are fragmented. Data models are inconsistent. Documentation is incomplete. Critical business rules live in tribal knowledge rather than code. Security teams are understandably cautious. And the people who know how everything really works are often busy keeping the current system alive.
This is why domain expertise matters more than generic enthusiasm. Building useful AI systems in business settings is not just a model problem. It is a systems problem. It requires understanding the operational shape of the work, the edge cases, the human checkpoints, and the consequences of being wrong.
Human oversight, in this context, is not just a temporary safety blanket. It is part of the design.
There is a recurring fantasy in AI discussions that oversight is evidence of immaturity, and that the end state is full autonomy everywhere. Maybe in some domains. But in many business environments, the more realistic future is selective autonomy. Agents will do a great deal independently, but within carefully defined boundaries. They will escalate when confidence is low, when risk is high, or when a decision crosses a threshold that still carries legal, financial, or reputational weight.
That sort of human-in-the-loop architecture is not a compromise. It is often the difference between a demo and a deployable system.
Security adds another layer. The more authority an agent has, the more carefully that authority has to be managed. A read-only research assistant is one thing. An agent that can move money, alter records, trigger procurement events, modify supply chain data, or execute internal workflows is something else entirely. Once AI becomes operational rather than merely advisory, the control model matters enormously.
This is why standards and controls are going to be such an important part of the next phase. Not because they slow innovation, but because they make it sustainable. Identity, permissions, observability, auditability, rollback paths, environment isolation, and policy enforcement all become central once agents are acting inside real systems. Businesses are not resisting progress when they ask these questions. They are doing the work required to make progress durable.
There is also a broader economic shift hidden inside all this.
If AI agents can perform more work at lower marginal cost, then software companies may need to rethink not only their product design but also their pricing and business models. Charging per human seat starts to feel less natural in a world where value may come from fleets of non-human operators. Some categories may move toward usage-based pricing, outcome-based pricing, or hybrid models that reflect the fact that the “user” is no longer always a person.
This may seem like a commercial footnote, but it is actually central to adoption. Technology spreads more smoothly when the economic model aligns with how it is used. When it does not, friction builds quickly.
There are historical echoes here. Many major computing transitions looked awkward and uneconomical at first. Mainframes to client-server. On-premise to cloud. Early internet infrastructure. Even semiconductors, in their time, were not judged solely by their eventual significance but by whether they made immediate economic sense within old frameworks. New technologies often appear too expensive, too awkward, or too narrow right before they become foundational.
AI may follow a similar path. The current debates about token costs and agent economics may one day look as temporary as older debates about server utilization or bandwidth pricing. But temporary does not mean irrelevant. These constraints shape what gets built now, which in turn shapes what becomes feasible later.
That is why the next few years may feel contradictory. AI will keep getting better, sometimes dramatically so, and yet adoption may still feel slower than expected in many sectors. Not because the technology stalled, but because diffusion depends on much more than invention. It depends on redesigning interfaces, rebuilding trust, reshaping operating models, and integrating with systems that were never built for this kind of actor.
And perhaps that is the most useful way to think about the moment. We are not just adding a new feature to software. We are introducing a new participant into digital work.
That participant needs tools, boundaries, interfaces, economics, and governance. It needs to be legible to the organization around it. And the organization, in turn, needs time to adapt.
Silicon Valley often underestimates how long that takes. Not because it lacks imagination, but because it tends to confuse technical possibility with institutional readiness.
The gap between those two things is where most of the real work lives.
And in a quieter way, that is encouraging. It means the future of AI in business will not be decided only by whoever builds the best model. It will also be shaped by whoever can integrate intelligence into real workflows with care, clarity, and respect for how organizations actually function.
That is a slower story than the headlines suggest. But it is probably the more important one.