Confession: I overestimated how much AI would simplify the software stack.
AI arrived with a familiar promise. Work would become cheaper, faster and simpler. For a while, it felt true. One tool could draft copy, answer questions, summarize meetings or take repetitive admin tasks off someone’s desk. Then came a second model for analysis, a meeting bot, an agent layer, a connector platform and an AI upgrade inside more and more of the software already on the bill. The race to reduce complexity is creating the very inefficiency it was meant to remove. The chat box looks clean. The invoice does not.
3 Rules for Effective AI Integration
Instead of simplifying workflows, AI integration often leads to software fragmentation and hidden expenses. Adding new AI tools without retiring old systems creates subscription bloat, unexpected consumption charges, and added complexity around training, oversight and integration.
- Audit before adding: Require every new AI tool to eliminate an existing task or retire a legacy system.
- Calculate full costs: Account for token usage, connectors, training, security and human error correction, not just the demo price.
- Automate tasks, not accountability: Ensure a named human leader owns any AI output affecting compliance, customers, employees or reputation.
Amplifying Cost, Not Work
Humans are remarkably consistent. A shiny new tool appears, we’re told it will make everything easier, and we reorganize our work around it before understanding the consequences. This is why a human-first approach matters. Human-first doesn’t mean anti-AI. It means the technology serves the work rather than the work being redesigned to justify the renewal. The busywork AI removes was never the whole job. It was the task nobody had a better way to do.
To be clear, this argument isn’t me turning against AI. It’s me revising my view of the challenges related to integrating AI. The models will improve and access to more user-friendly chatbots, tools and plug-ins will become ubiquitous for companies of all sizes, but those things won’t automatically fix a badly designed workflow, unclear ownership or a collection of systems that do not speak properly to one another. AI reveals those problems, but it doesn’t magically solve them.
AI’s Cable-to-Streaming Moment
For an example from another industry, take streaming, which was supposed to kill cable. Instead, we rebuilt cable, one subscription at a time. One service has sports, another has the show everyone is talking about. A third stays because cancelling it would trigger a minor household revolt. What looked like choice became fragmentation, higher bills and 10 minutes spent searching for something you already pay to watch.
AI risks doing the same inside businesses. It’s not replacing software. Rather, it’s becoming software’s cable TV problem. The old technology stack hasn’t disappeared. It has been hidden behind a nicer front door and given another monthly charge.
This isn’t a theoretical cost. Zylo’s 2026 SaaS Management Index, based on more than 40 million licences and $75 billion in managed spending, found that spending on AI-native applications had risen 108 percent year-over-year. In its survey of 218 IT leaders, 78 percent had faced unexpected charges linked to consumption-based or AI pricing, while 61 percent had cut other projects because of unplanned SaaS cost increases.
The pilot may be cheap. The operating model rarely is.
Tokens, licences, integrations, data preparation, security reviews, training, human oversight, error correction and escalation all add up. So does the work of deciding which model should handle which task, keeping the connectors alive and explaining to finance why a tool used heavily last month has produced a completely different bill this month. Businesses are discovering that a lot of “AI strategy” is becoming subscription accumulation with better branding.
Be Careful What You Add
Adding AI without removing anything is not transformation. It is a surcharge.
We’ve seen this pattern before. The dot-com crash didn’t kill the internet. It killed the idea that adding “.com” made a business valuable. An AI correction could do something similar by separating the tools that remove real work from those that merely add another charge and a launch announcement. Gartner predicts that more than 40 percent of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. That wouldn’t prove AI has failed. It would prove that buying the future is easier than redesigning the present.
What remains is judgement, taste, creativity, expertise and the ability to connect with people. AI should increase human capacity, not create dependency on a chain of systems nobody fully understands. For leaders, the practical implications are straightforward:
Audit Before You Add
For every new AI tool, ask which task will disappear, which system can be retired and who owns the outcome. If nothing leaves the stack, you’ve added a layer, not removed one.
Count the Full Cost, Not the Demo Price
Include tokens, usage charges, connectors, data preparation, security, training, verification, exceptions, outages and the work required when the system fails.
Automate Tasks, Not Accountability
Once an AI-supported output affects a customer, employee, pay, promotion, compliance or reputation, a named human leader owns the final decision.
Avoiding the Streaming Mistake
AI doesn’t have to become software’s cable bundle. But we’re well on our way if every efficiency claim creates another licence, connector and verification step. The arms race is loud. The subscription problem is quieter and, frankly, may be more expensive.
If AI doesn’t remove work, cost or complexity, it’s not transformation. It’s just another subscription.
