The tech industry has a long history of overpromising and underdelivering, but AI has taken this to new heights. We’re bombarded daily with headlines about AI writing novels, diagnosing diseases, and even replacing entire job functions. Yet, when you peel back the layers, you find a landscape littered with half-baked implementations, inflated claims, and solutions that work only in the most controlled environments.
That means AI must either perform a critical task more efficiently than existing methods or enable something previously impossible at a justifiable cost. It’s the AI dichotomy between promise and practicality.
Right now, only a handful of AI applications meet this bar. The rest? They’re stuck in the land of PowerPoint promises and Silicon Valley echo chambers. Or, as Financial Times journalist John Thornhill aptly put it: “How people use a product in the real world may diverge wildly from the designers’ intentions” (“We Are the New Gremlins in the AI Machine”; ft.com/content/aaa57d4b-fee6-4109-87ac-9222d706fe07).
Where AI Actually Delivers (And Where It Doesn’t)
1. Intelligent Document Processing (IDP): A Rare Bright Spot
Let’s start with the clearest success story—Intelligent Document Processing (IDP). Traditional document capture was a mess, with error-prone OCR, manual data entry, and endless exceptions. AI is changing the game by dramatically improving accuracy and introducing real-time decision making at scale.
Why does IDP work? Because:
♦ The rules are well-defined (an invoice is an invoice).
♦ The data is structured (fields, labels, formats).
♦ The outcomes are measurable (fewer errors, faster processing).
This isn’t speculative—it’s provable ROI. And that’s why IDP is one of the few AI use cases that consistently gets budget approval. Companies such as UiPath and Hyperscience have built entire businesses around this because it solves a real, painful problem—not because it sounds cool in a press release.
2. AI-Powered Customer Service: Less Terrible, But Still Limited
Next up: AI-powered customer service. Let’s be honest—most chatbots are still terrible. But the latest generation of agentic AI is making them less terrible, reducing frustration by a whopping 30%– 50% in well-scoped scenarios.
So, what’s the catch? These systems only work when:
♦ The interactions are scripted (think password resets, tracking updates).
♦ The fallback is human (escalation paths matter).
♦ The training data is clean (garbage in, garbage out).
This isn’t artificial intelligence— it’s artificial competence. And that’s fine. Not every problem requires AGI. (Maybe none require AGI, but that’s an argument for another day.)