If you’re skeptical about AI, you’re paying attention.
Valuations look stretched. OpenAI lost $7.8 billion in the first half of 2025 despite $4.3 billion in revenue. Anthropic burned $5.3 billion in 2024. One analyst report from MacroStrategy Partnership estimates the current AI capital misallocation is 17 times larger than the dot-com bubble. Central banks—the IMF, the Bank of England—are issuing warnings about speculative excess.
Over 75% of surveyed economists describe current AI investment levels as “bubbly.” At Yale’s CEO Summit in June 2025, 40% of executives and VCs foresaw an imminent correction.
So yes, the skepticism is warranted. The question is whether it’s the right question.
Bubbles Burst Speculation, Not Value
Here’s what the dot-com bust taught us: the bubble bursting didn’t kill e-commerce. It killed Pets.com. Amazon survived. PayPal survived. The companies solving real problems at sustainable economics survived.
When the hype settled, the infrastructure remained. The capability remained. And the organizations that had integrated those capabilities into their operations had a compounding advantage over those who’d waited for “clarity.”
AI will follow the same pattern. Some of what’s being funded today is pure speculation—chat wrappers with $100 million valuations, companies with no clear path to revenue, solutions looking for problems. That speculation will get corrected.
But some of what AI enables is simply capability that didn’t exist before. Capability that solves problems organizations have had for decades. And that won’t go away when investor sentiment shifts.
What’s Clearly Speculation
Let’s be honest about what probably won’t survive the correction:
“AI will replace all jobs in five years.” It won’t. The tasks AI handles well are specific: pattern recognition, language processing, classification, extraction. These are narrow capabilities, not general intelligence.
AGI timelines measured in years. Experts disagree wildly on whether artificial general intelligence is five years away or fifty. Building strategy around imminent AGI is speculation, not planning.
Consumer chatbots solving problems nobody had. Most consumer AI applications are features, not products. The novelty wears off, and the retention isn’t there.
$100 billion valuations for API wrappers. Companies that add a chat interface to existing LLMs aren’t building defensible businesses. When the APIs improve and costs drop, the margin disappears.
If your AI strategy depends on any of these being true, you’re speculating.
What’s Clearly Real
Now for what’s different. There’s a category of AI application where the value is boring, measurable, and permanent: batch data processing.
This isn’t about chatbots or disrupting industries. It’s about taking work that existed before AI—work that was expensive, manual, and often impossible to automate—and actually automating it.
Taking dirty data and making it clean. Every organization has decades of data sitting in formats that don’t talk to modern systems. PDFs. Handwritten notes digitized to images. Spreadsheets with inconsistent formatting. Legacy databases with cryptic field names. Documents that require human interpretation to understand.
Traditional automation failed here. Regex can’t understand context. Rules engines require every edge case to be anticipated. Pattern matching breaks on variations.
LLMs can read a document and understand what it means, not just what pattern it matches. They can take a maintenance log written in shorthand by a technician in 1987 and extract structured data that feeds into a modern system.
This isn’t speculation. Organizations are doing it today.
Processing documents at scale. McKinsey reported that their internal AI tool reduced document classification time from 20 seconds per document to 3.6 seconds—while improving accuracy from ~50% to ~80%. That’s 676 hours of manual work saved per analyst per year.
Insurance companies are processing claims documents with 70% accuracy in near real-time, freeing agents from repetitive extraction tasks.
Legal teams are achieving 92% accuracy in extracting structured data from contracts and agreements.
These aren’t proof-of-concept demos. They’re production systems with measured ROI.
Connecting systems that couldn’t talk before. AI becomes the preprocessing layer before your ETL. It takes unstructured inputs and produces structured outputs that flow into databases, data warehouses, APIs. It bridges the gap between what humans write and what machines need.
A shipping company can take bills of lading—documents that arrive in inconsistent formats from thousands of partners—and normalize them into a unified schema. An energy company can take decades of equipment maintenance logs and make them queryable. A financial services firm can process contracts and extract key terms without reading each one manually.
The Economics Actually Work
“AI is expensive” is true for real-time chatbots handling millions of queries. Batch processing has different economics.
When you don’t need instant responses, you can use batch APIs that offer 50% discounts for 24-hour turnaround. When you’re processing thousands of similar documents, you can optimize prompts and cache results. When the task is classification or extraction rather than open-ended generation, you can use smaller, cheaper models.
Most firms recover costs within 12-18 months. The crossover point—where AI costs less than human labor for the same task—is often surprisingly low volume.
And the economics only get better. Model costs have dropped dramatically year over year. They’ll continue dropping. The work you automate today becomes cheaper to run tomorrow.
The Risk of Sleeping On It
Here’s the uncomfortable truth: while you’re waiting for the bubble debate to settle, your competitors aren’t.
The organizations integrating AI into their data pipelines today are building compounding advantages. They’re processing documents faster, with fewer errors, at lower cost. They’re connecting systems that couldn’t talk before. They’re making decisions with data that was previously locked in unstructured formats.
By the time “is it a bubble?” has a clear answer, the gap will have widened. It takes time to build these capabilities. To identify the right use cases. To train your team. To integrate with existing workflows.
The winners won’t be the organizations that called the bubble correctly. They’ll be the ones that distinguished speculation from value and acted on the value.
What Comes Next
Maybe AI is a bubble. Maybe it isn’t. Either way, there’s value being created right now that isn’t going anywhere.
The question isn’t “is AI overhyped?” It is.
The question is what you’ll do while everyone else argues about it.
This is the first post in a series on AI for batch data processing. Coming up:
- The Dirty Data Problem AI Was Made For — Why legacy data is the universal pain point, and how LLMs solve problems traditional automation couldn’t.
- The Economics of AI Batch Processing — Cost models, batch API strategies, and when the ROI actually works.
- Human-AI Workflows That Actually Work — Why 100% automation isn’t the goal, and how to build hybrid systems that scale.
- Getting Started: Your First AI Batch Pipeline — Practical patterns for IT teams ready to experiment.
The capability is real. The question is whether you’ll use it.
InFocus Data helps organizations build AI-powered data pipelines that turn messy, unstructured data into clean, usable information. If you’re ready to move from theory to implementation, let’s talk.