In a borrowed conference room, a founder clicks run. The demo hums along on a frontier model she doesn’t own. The room nods. Then the inevitable question: is this just a wrapper?
The wrapper myth meets the real market
Dismissals come easy in a boom. In today’s AI pile-on, wrapper has become shorthand for shallow: a thin interface atop GPT, Claude or Gemini. The sneer skips the hard part. Turning a general-purpose model into something customers trust every day is not a veneer. It is engineering, data plumbing, workflow design, guardrails, latency work, cost control, and a hundred tiny product decisions that only show when the system is under stress.
Plenty of the popular apps do rely on third-party models. Many will stay that way. Winning does not require building a foundational model any more than building a great operating system required fabricating your own silicon. The real question is where value accrues: the model, the infrastructure that feeds it, or the software that wraps painful work in a reliable experience.
Where value hides: orchestration, memory, and fit
Three areas separate toys from tools:
- Orchestration. Real work spans multiple services, models, and tools. Good orchestration makes them talk to each other predictably, with retries, observability, and policy baked in.
- Context storage and retrieval. Memory is the difference between one-off parlor tricks and compounding utility. It demands careful data design, security, and speed.
- Owners-only data. Fine-tuning or adapting on proprietary workflows, logs, and outcomes can create an advantage that isn’t instantly replicable by a platform toggle.
That’s the unglamorous layer where buyers live. A customer doesn’t purchase a model; they buy fewer headaches. They buy fewer tickets in the queue, faster close rates, safer drafts, cleaner handoffs. When AI fits the grain of a workflow, the model is an ingredient. The meal is what counts.
The platform squeeze is real
None of this gives application builders a free pass. History says platforms will reach up the stack. When they do, they often cut prices and replicate the obvious. That leaves many AI applications competing against their own suppliers. In an analysis of the code-assistant market, writer Ethan Ding argued that some breakout tools grew by subsidizing expensive API calls, creating brittle economics. As he put it, they were “selling dollar bills for quarters.”
The lesson is straightforward: if your differentiator is indistinguishable from a provider’s default setting, you are one press release away from a margin crisis. The antidote is to own something the platforms don’t want to own: messy integrations, industry-specific compliance, post-sales deployment, change management, or the ongoing data loops that make the product better only for your customers.
Hype, limits, and the shovel sellers
Even strong products run into the wall of realism. Early adopters who tried AI for complex work often found the features ahead of the reliability curve.
“The output was initially impressive, but our enthusiasm for it waned as we began to see the gaps,” said Andy Gillin, managing partner at GJEL Accident Attorneys, describing a legal research tool he tested. “After several months, it became quite clear that while the tool had potential, its execution fell short of our firm’s requirements.”
“Trends and market conditions can shift quickly,” said Jon Morgan, CEO of consulting firm Venture Smarter, about his experience with predictive analytics. “The insights provided by the platform didn’t offer the level of accuracy and reliability we needed to make informed decisions.”
There are structural reasons for that unevenness.
“LLMs are inherently limited by the data they’ve been trained on and this hasn’t really been acknowledged,” said Dr Ruairi O’Reilly, a computer science lecturer at Munster Technological University. “As these models get larger and larger, they need more computing power. So, at a certain point, the efficiency of larger models will be outpaced by the costs associated with training them.”
Those costs have minted kings in the hardware aisle. Graphics chips, memory, networking, power, cooling, and data centers have been the quiet winners of the AI rush. O’Reilly notes that Nvidia controls a dominant share of the GPU market and has been buoyed by the demand for training and serving these models. The shovel-seller analogy endures because it is true: whether a thousand prospectors strike gold or wash out, someone sells the picks, racks, and diesel.
What durability looks like now
So what separates a lasting AI product from a fleeting wrapper? A few patterns keep showing up in companies that stick:
- They pick a narrow, high-friction problem and own the entire outcome, not just a step in the middle. That means stitching together models, retrieval, rules, and humans-in-the-loop.
- They adapt to customer data without exfiltrating it, turning usage into better performance for that customer rather than a generic improvement for everyone else.
- They make deployment painless. Hosting, auth, billing, audit logs, failure modes, SLAs, rollback plans. Boring is a moat.
- They price against value delivered, not tokens consumed. When the buyer’s spreadsheet shows fewer tickets, shorter cycle times, or higher conversion, budget appears.
- They build with, not against, the platforms. If a foundation model closes the gap on a feature, they move up a layer, not into denial.
Evidence suggests this approach can work. Customer service is one of the first places automation has stuck.
Programs will “allow companies to use automated workflows that keep humans in the loop,” O’Reilly said, pointing to early successes handling customer queries without human intervention.
That framing matters. Keep humans in the loop. Close loops with data. Pick loops that customers pay to close.
Owning the model is not the point
Building a frontier model is a heroic, capital-intensive pursuit wrapped in legal and logistical thorns. It is also unnecessary for most software companies. The long arc of computing is full of firms that built enduring businesses atop shared compute platforms: databases on commodity servers, cloud apps on standardized chips, mobile apps on iOS and Android. The same pattern is unfolding with LLMs.
Will some application layers get steamrolled when platform providers roll out native features? Certainly. That only sharpens the strategic question every AI startup must answer: what do you own that a model vendor cannot copy overnight, that customers will miss if you disappear, and that gets stronger as they use it?
Sometimes the answer is a foundational model. More often, it is the stack around it: orchestration that never flakes, memory that never leaks, domain data that never leaves the building, and a product that dissolves a painful workflow so thoroughly the buyer forgets there was ever plumbing under the floor.
