Your company might actually run smoother if you deleted every AI
I’ve been looking at the sheer amount of capital being diverted toward "token burning" lately. We see startups raising millions to build wrappers around existing models, and enterprise-level giants spending fortunes on seat licenses for Copilots. It feels like a massive experiment is running in real-time, and the hypothesis is: "AI increases productivity." But there is a glaring lack of empirical data to back this up for the average mid-sized business.
The Hidden Cost of AI Integration
When we talk about the impact of AI, we usually focus on the potential upside—the speed of code generation or the instant drafting of emails. However, we rarely talk about the "hidden friction" that these tools introduce into a professional AI workflow.
- The Verification Tax: Instead of writing a document from scratch, an employee spends thirty minutes prompting an LLM and another forty-five minutes fact-checking the hallucinations. That's nearly an hour of high-cognitive labor spent just babysitting a machine.
- Context Fragmentation: Every time a team uses an AI to summarize a meeting or a long thread, they lose the nuance and the "vibe" of the original conversation. You get the bullet points, but you lose the subtext that actually drives human decision-making.
- Technical Debt: Implementing an LLM agent into a production pipeline isn't a one-and-done task. You now have to manage model drift, prompt versioning, and the inevitable edge cases where the model decides to go off the rails.
Would we actually be better off?
There is a strong argument to be made that a "pre-AI" workflow might actually be more profitable for many sectors. If you stripped away the AI, you'd immediately reclaim those massive subscription costs and API credits. You would also likely see a return to more intentional communication. Without the ability to generate a "good enough" email in three seconds, people might actually think twice before hitting send.
For developers, the question is even more pointed. We see a lot of hype around how much faster a junior dev can ship code with an AI assistant. But if that code requires a senior engineer to spend double the time debugging because the AI used an outdated library or a non-optimal pattern, the net productivity of the company has actually decreased.
I'm curious if anyone here has actually measured their team's output or error rate before and after a major AI deployment. Did the "speed" gained in the first stage actually translate to a better product, or did it just create a mountain of low-quality work that someone else had to clean up later?