How is a serious AI application different from an AI wrapper?
An AI wrapper is not defined by having a thin codebase or lacking RAG. The meaningful distinction is whether a product merely relays a model response, or takes responsibility for a defined task with trustworthy data, business controls, repeatable evaluation, and ongoing operation.
Putting a simple interface around a general model can be entirely appropriate for low-risk writing or translation. The problem begins when that interface is sold as if it understands a particular company and can reliably act for it, without showing where its facts come from, which actions it may take, how failures are handled, or how performance was measured. Conversely, a vector database, workflow engine, or private deployment does not by itself make a product mature.
When defining model, data, and production boundaries, also compare What happens if an AI application misses its accuracy target after launch? and What is a workflow-based AI agent, and how is it different from a chatbot?; the linked guidance adds context that should be considered in the same decision.
| Delivery level | What it actually does | Defensible promise | Main limitation |
|---|---|---|---|
| Model interface | Sends prompts and displays responses | Convenient general-purpose generation | No enterprise source of truth or task assurance |
| Task assistant | Adds templates, validation, and editing | Reduces effort for a narrow content task | Cannot represent company facts by default |
| Knowledge application | Retrieves approved, versioned sources and cites them | Answers within the indexed and authorized material | Retrieval and stale-source errors still propagate |
| Business application | Adds identity, permissions, APIs, approval, logs, and rollback | Queries or acts within explicit authority | Highest integration and operating burden |
RAG is one architectural technique for giving a model retrieved external context. The original NeurIPS 2020 RAG paper discusses external non-parametric memory, knowledge updates, and provenance; it does not prove that every corporate knowledge bot will be accurate. Poor parsing, incorrect permission filters, or a missed passage can still produce a wrong answer.
Before launch, ask for five concrete artefacts: a task specification describing users, inputs, allowed outputs, and exclusions; a data register naming the owner, version, access level, and refresh date of every source; a frozen evaluation set drawn from real work; operational traces covering model, prompt, retrieval, tool calls, human edits, and cost; and named ownership for updates, incidents, complaints, and rollback.
Metrics must match the product. A knowledge assistant needs retrieval recall, evidence-supported answer accuracy, citation consistency, appropriate refusal, severe-error rate, P95 latency, and full cost per task. An action-taking agent also needs tool-argument accuracy, duplicate-action counts, unauthorized attempts, approval rate, and rollback success. A hand-picked demo is not commercial evidence.
Google Research's ML Test Score frames production readiness across data, models, infrastructure, testing, and monitoring. ISO/IEC 25010:2023 adds broader software qualities such as reliability, security, and maintainability. Both are more useful maturity signals than a model name.
Wavesteam can deliver knowledge integration, system connections, permissions, evaluation, monitoring, and an exit plan. Those capabilities are not proof that a custom project will outperform an existing product. The same standard applies to us: scope, data flow, test design and results, interfaces, deployment assets, operational controls, and unresolved assumptions should be inspectable in the proposal or acceptance package. If we cannot produce repeatable evidence for a specific business outcome, calling the work “custom AI” does not make it more than concept packaging.