Murali Swaminathan is the Chief Technology Officer at Freshworks.
getty
Every major technology wave lands on top of whatever was already there. The internet landed on client-server architectures that weren't built for it. SaaS landed on on-premise procurement cycles that didn't know what to do with it. Cloud landed on security models designed for data centers. AI is landing on all of it at once, and the foundation can't absorb the weight.
I've been building enterprise software through every one of those transitions. The pattern is remarkably consistent. The constraint has never been the technology itself, but the complexity underneath it.
This is called the complexity tax: the invisible cost organizations pay, cycle after cycle, to keep systems stitched together well enough to function. AI doesn't create the tax, but it makes the bill impossible to ignore.
Complexity tax rarely traces back to a single bad decision. It builds through a series of reasonable ones.
A support team picks a ticketing tool that solves a real problem. IT adopts a different platform for internal requests. Finance runs a third system for approvals. Each system works fine on its own. But eventually the business expects them to work together, so teams start stitching integrations across them: API calls, middleware, manual handoffs and the CSV export someone still runs every Tuesday morning.
The weight grows so gradually that teams stop noticing it. But if an AI agent is introduced and needs to operate across those systems—verifying identity in one, checking policy in another, executing an action in a third and logging the result in a fourth—every gap becomes a place where the agent slows down, hallucinates context it doesn't have or fails silently. The AI isn't broken. The architecture underneath it is carrying decades of layered decisions.
Let's say an employee forgets their password. They open a chat, type "I need to reset my password," and an AI agent picks it up.
Except the identity system lives in one platform, the access policy engine sits in another and the ticketing system that logs and closes the request lives somewhere else entirely. None of them were designed to share context in real time. So the agent has to move across systems step by step, interpreting different response formats and piecing together enough information to complete the task—all while the employee sits there watching a loading spinner.
In a clean environment, that interaction takes seconds. In the environments most enterprises actually operate in, the agent stalls. It can't verify the employee's role because the identity system returns data that the policy engine doesn't recognize. It can't auto-resolve because the ticketing system requires a field that the agent doesn't have. At that point, it escalates to a human, which is exactly what the AI was supposed to prevent.
Now multiply that by every routine request across IT, HR and customer support. That's the complexity tax: thousands of small frictions, every day, absorbing the time and budget that was supposed to go toward building something better. It works in pilots because it stays inside a boundary. But the moment it crosses systems, it inherits every integration gap that was already there.
Every wave brings a version of the same temptation: skip the fundamentals because the new technology is powerful enough to compensate for the gaps. It rarely does.
With AI, the fundamentals are the same ones that have always mattered. Define what success looks like in operational terms before you build—not "implement AI" but "reduce Tier 1 resolution time from 14 minutes to three." Start in a contained environment before you scale. Build in control points: confidence thresholds that determine when the AI acts autonomously and when it escalates to a human. Create a clean audit trail so that when compliance asks how a decision was made, you can answer.
None of this is new. AI just raises the cost of skipping it, because failures surface faster and harder to unwind.
Ask the architectural questions first. Where does the data actually live? Which systems need to share context for a workflow to function end-to-end? What does the human-to-AI handoff look like when the agent's confidence drops below the threshold? Those aren't glamorous questions, but they're the difference between a pilot that generates a press release and a deployment that changes how the organization operates.
The instinct when AI stalls is to treat it as a technology problem—upgrade the model, add a tool and bring in a systems integrator. More often, that's treating a symptom.
The underlying issue is structural. When supporting software requires a dedicated team just to keep it running, that's the complexity tax in action. And the solution isn't more AI on top of the complexity. It's reducing the weight AI has to carry before it ever gets deployed. Streamline the toolchain, connect the data and consolidate the systems that need to share context.
McKinsey found that nearly two-thirds of organizations have not yet begun scaling AI across the enterprise, even as 64% say AI is enabling innovation and only 39% report EBIT impact. AI is visible everywhere, but the operating model underneath still isn't built to carry it.
Most organizations aren't debating whether to invest in AI. That decision is settled. The harder question is one most leadership teams haven't confronted directly: Of everything you're spending on AI right now, how much is actually advancing capability, and how much is just servicing the complexity tax you've been accumulating for the last two decades?
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。