Why AI Productivity Gains Are Stalling Now

Fexingo (0:01)

Look, if you have been tracking the software space over the last six months, there is a quiet but very real stagnation happening. It is not about the technology failing, it is about the economics of using it failing.

Lucas (0:15)

You mean the gap between the hype cycle and the actual bottom line?

Fexingo (0:19)

Exactly. We spent two years talking about how AI would slash headcount and double output. But now, with September ending, the data is showing something else. Enterprise spending on LLM infrastructure is still climbing, but the reported productivity metrics are flatlining or even dipping for many mid-cap tech firms.

And I think we need to talk about why your coffee is so good today.

Seriously, if this conversation helps you see where the market is actually going, rather than where it was yesterday, it might be worth keeping the show running ad-free. A couple of dollars a month is genuinely what keeps these going. So, if you want to support that, check out by me at coffee.com/fixingo.

Lucas (1:02)

I appreciate that and I know our listeners do too. It allows us to dig into the messy details without worrying about ad slots.

Fexingo (1:11)

So let us get back to the meat of it. The core issue is what I call the integration tax.

Companies are buying the tools, but they are not changing the workflows. They are just adding an AI layer on top of broken processes.

Lucas (1:26)

So instead of streamlining, they are creating more steps for employees to verify AI outputs.

Fexingo (1:32)

Precisely. And this is where the numbers start to look ugly for some of these valuations. Take a look at the recent movement in software stocks.

Oracle dropped nearly 6% recently while Snowflake barely budged. Meanwhile, Salesforce and ServiceNow are up significantly. Why? Because Salesforce has deeply integrated AI into their core CRM workflow, whereas Oracle is still trying to bolt it onto legacy database structures.

Lucas (2:01)

That makes sense. Integration depth determines whether AI is a cost center or a profit driver.

Fexingo (2:09)

It does. And the deeper point here is about the context window.

As models get larger, the cost of processing every single token in a long document skyrockets. We are seeing companies hit a wall where the compute cost of reading a 50-page contract outweighs the value of the summary AI provides.

Lucas (2:28)

So the bigger model isn't always the better model if the use case doesn't require deep reasoning.

Fexingo (2:34)

Correct. You don't need a 10 billion parameter model to summarize a meeting transcript. You need a small, fast, cheap model.

But enterprises are often locked into contracts with the big providers because of trust and security compliance. They are paying premium prices for suboptimal utility.

Lucas (2:54)

Which creates a margin squeeze. If they can't pass those costs onto clients, their own profitability takes a hit.

Fexingo (3:02)

Exactly. And this brings us to the second part of the stall, the human factor. There is a phenomenon called prompt fatigue.

Employees are spending more time crafting precise inputs to get usable outputs than they did doing the task manually.

Lucas (3:19)

It's the digital equivalent of filling out a form that asks for three different versions of the same answer.

Fexingo (3:25)

Spot on.

And when you combine prompt fatigue with the verification burden, you get a net negative productivity gain. I spoke with a CTO last week who said their team is actually working longer hours because they are managing AI errors alongside regular tasks.

Lucas (3:42)

That sounds exhausting. So the promise of freeing up human creativity is being replaced by administrative drudgery.

Fexingo (3:49)

It is. And the market is starting to punish this.

Look at Palantir. They are up 5% recently, and it is not just because of government contracts. It is because their platform enforces structured data entry before AI can touch it. They force discipline on the workflow.

Lucas (4:09)

So structure comes before intelligence. That is a key lesson for any company trying to adopt AI right now.

Fexingo (4:16)

Absolutely. You cannot automate chaos.

If your internal data is messy, your AI will just produce messy results faster. This is why companies with clean data architectures are outperforming those with flashy AI demos, but dirty backends.

Lucas (4:32)

It feels like we are moving from the wow phase to the how phase, and the how is much harder and less glamorous.

Fexingo (4:40)

That is the perfect way to put it. The wow phase is over.

Now we are in the boring phase of implementation. And boring usually means expensive and difficult. Investors are starting to realize that the easy wins are gone.

Lucas (4:56)

So where should businesses focus their limited budgets if they aren't chasing the biggest model?