The wrapper problem
Most B2C AI apps are thin interfaces sitting on top of someone else's model. They take the OpenAI or Anthropic API, wrap it in a custom UI, add a system prompt, and charge $20 a month. The product is not the technology. The product is the packaging.
This worked in 2023 and 2024 because the foundation model providers had not yet built consumer-facing products that covered every use case. There were gaps. AI writing assistants filled the gap before ChatGPT added long-form editing. AI coding tools filled the gap before Claude and ChatGPT added code interpreters and artifact generation. Each gap was a temporary window, and wrapper companies rushed through it.
The window is closing. When OpenAI added Code Interpreter, a wave of "upload your CSV and get analysis" startups became redundant overnight. When Anthropic shipped Artifacts, tools that let you "build things with Claude" lost their differentiator. When ChatGPT added native image generation, dozens of prompt-engineering wrappers for DALL-E lost their reason to exist. The foundation model providers are not just building models — they are building products. Their products will always have a structural advantage: they control the model, they have the deepest integration, and they can offer features at marginal cost that wrapper companies must charge for to survive.
The moat in AI is the model itself — the training data, the research team, the compute infrastructure. Wrapper companies rent this through an API and mark it up. When the landlord opens their own storefront on the ground floor, the subletter upstairs has a problem.
Not every B2C AI product will fail. Products with genuine proprietary value — unique datasets, specialized fine-tunes, hardware integration, network effects — will survive. But the majority of what shipped between 2023 and 2025 was a system prompt and a subscription page. Most of these will disappear within three years, absorbed by the platforms they depend on or abandoned by users who find the same capability built into their existing tools.
Why enterprise is different
Enterprise AI is a fundamentally different problem. The model is maybe 20% of the work. The other 80% is everything the model cannot do on its own.
Start with data. Every company of meaningful size has its information spread across systems that were never designed to talk to each other. A mid-size insurer might have policy data in an Oracle database from 2008, claims records in a proprietary system with a SOAP API, customer communications in a CRM customized beyond recognition, and compliance documents in SharePoint folders organized by someone who left four years ago. Connecting an AI agent to this data is not a prompt engineering problem. It is a data engineering problem that requires understanding legacy schemas, proprietary formats, access control models, and the unofficial workflows that fill the gaps between how the system was designed and how employees actually use it.
Then there is workflow integration. Enterprise AI that works in production does not replace existing processes — it augments them. A contract review agent does not eliminate the legal team. It reads the contract first, flags clauses that deviate from standard terms, and presents them to a lawyer who makes the final call. The agent must fit the existing approval chain, respect the existing permission model, and produce output in the format the downstream process expects.
Change management kills more AI implementations than any technical limitation. Employees who have spent years developing expertise in a process are not always enthusiastic when a machine arrives to do that process faster. Some are skeptical about accuracy. Some fear obsolescence. Some simply do not want to learn a new system when the old one works. Deploying AI without addressing these reactions is a reliable way to spend $200K on a tool nobody uses.
Finally, accuracy in production is a different standard than accuracy in a demo. An AI agent that extracts data from contracts correctly 95% of the time sounds impressive until you realize that at 500 contracts a month, 25 will contain errors a human must catch. In regulated industries — finance, insurance, healthcare, legal — those errors trigger compliance violations, misstate liabilities, or create legal exposure. Maintaining consistent accuracy means building evaluation pipelines, monitoring for drift, handling edge cases, and keeping up with regulatory changes. This is ongoing operational work, not a one-time build.
The integration tax
Every enterprise is different. This is the structural reason enterprise AI creates durable value.
A law firm reviewing commercial leases needs an agent that understands lease-specific terminology, recognizes non-standard clauses across dozens of landlord templates, and integrates with iManage or NetDocuments or whatever the firm actually uses. An insurance company processing motor claims needs an agent that reads damage assessments, cross-references policy terms, checks fraud indicators, and outputs recommendations in the format their claims system expects. An e-commerce company optimizing inventory needs an agent that ingests sales velocity data, accounts for seasonal patterns, and triggers purchase orders in their ERP.
These are not the same problem wearing different clothes. The domain knowledge, data structures, workflows, compliance requirements, and failure modes are different in each case. There is no generic "enterprise AI" product you install and it works.
This is why consulting firms charge $200K or more for AI implementations. It is also why many of those implementations fail — not because the model was not good enough, but because the integration was done by people who understood the technology and not the business. They built a technically impressive demo that did not survive contact with the client's actual data, actual workflows, and actual employees.
The gap between "AI can do this" and "AI is doing this reliably inside your organization" is where the real work lives. Companies that close that gap build something defensible because the knowledge required — understanding a specific client's systems, data, and processes — is not easily replicated.
What this means for builders
If you are building AI products, the defensible position is in the enterprise layer. Not in the model layer — you will not out-train OpenAI or Anthropic. Not in the consumer wrapper layer — the platform providers will absorb your features. The enterprise layer is where durable value accumulates because the problems are specific, complex, and resistant to commoditization.
Concretely, this means data connectors that handle the messy reality of enterprise systems. Workflow orchestration that fits AI into existing processes rather than demanding the organization redesign around it. Domain-specific agents that understand the terminology, rules, and edge cases of a particular industry. Compliance tooling that maintains audit trails and adapts to regulatory changes. And the organizational skills to manage change — introducing AI in a way that actually gets adopted.
The model will keep getting better on its own. Every six months, foundation model providers release something more capable, faster, and cheaper. You do not need to solve the model problem. You need to solve the integration problem, and that problem is not getting easier.
Where we see this playing out
We say this from experience. At cmdev, our work building Bedrock pipelines and OpenClaw managed agents follows the same pattern across every engagement. The technical build — model selection, prompt engineering, API integration — is the part that goes smoothly.
The hard part is everything around it. Understanding a client's workflow well enough to know where an agent adds value and where it creates friction. Mapping data flows across systems built independently and never intended to share information. Designing the handoff between AI output and human decision-making so the AI accelerates the process without introducing new failure modes.
That understanding — of specific industries, specific systems, specific organizational dynamics — is the moat. It cannot be replicated by a better model or a prettier UI. It is earned through the work of deploying AI in real environments where failure has consequences. The next decade of AI value creation will be dominated not by the companies building models or wrapping APIs, but by the companies doing the hard, unglamorous work of making AI function inside the organizations that need it most.