AI Strategy
AI Is 80% Marketing and 20% Real Work: What Makes an AI Product Defensible
Updated Muhammad Taha6 min read
Most AI products on sale today are thin wrappers: a prompt template and a text box in front of someone else's model. They are easy to copy and have weak margins, so they are not defensible. The real value sits in the engineering around the model: orchestration, retrieval, validation and proprietary data.
Key takeaways
- A wrapper is a product whose main asset is a prompt. A competitor can rebuild it in a weekend.
- Per-use model costs squeeze wrapper margins, because every user action costs real money.
- Defensibility comes from three layers: controlled workflows, high-quality retrieval and proprietary data loops.
- If you are buying AI software, ask what happens when the model provider ships the same feature.
What is an AI wrapper?
An AI wrapper is software where the model does the valuable work and the product adds little more than an interface and a system prompt. Typical examples are a tool that summarises contracts, writes cold emails or drafts social captions by passing your input to a general-purpose model.
That is not automatically bad. Wrappers can be useful, cheap to build and a good way to test demand. The problem is the business model, because very little protects them. When your whole product can be duplicated with a single system prompt, you have a feature, not a company.
Two structural problems
1. The margin squeeze
Traditional software has high gross margins, because serving one more user costs almost nothing. AI products pay the model provider every time a user acts. Here is an illustrative calculation with made-up numbers:
| Item | Illustrative value |
|---|---|
| Price per user per month | $20 |
| Average requests per user per month | 400 |
| Average model cost per request | $0.03 |
| Model cost per user per month | $12 |
| Gross margin before hosting and support | 40% |
Industry commentary often puts classic SaaS gross margins far higher than that, but the exact figures vary widely, so work out your own: price per user minus model cost per user. If your heaviest users cost more than they pay, growth makes the problem worse. Our guide to LLM API cost control and AI SaaS pricing goes deeper.
2. Feature risk
If the core feature is a prompt that summarises documents, you depend on a model provider who can add the same feature, lower its price or expand its context window at any time. When that happens, the reason to pay for your product can disappear overnight.
Wrapper versus defensible product
[Thin API wrapper]
User input -> System prompt template -> Foundation model API -> Output
(Easy to copy, high churn, weak margin)
[Defensible pipeline]
User input -> Orchestration -> Retrieval / cache -> Model -> Validation code -> Approved output
(The value is the system around the model)| Question | Wrapper | Defensible product |
|---|---|---|
| What is the asset? | A prompt | Workflow, data and integrations |
| Can a competitor copy it in a weekend? | Yes | No |
| What if the model is swapped? | Behaviour changes unpredictably | Tests catch it and the pipeline continues |
| What happens on a bad output? | It reaches the user | Validation retries or hands off to a person |
| Where does the data come from? | Whatever the user pastes | Your own connected systems |
The three layers of a real moat
1. Deterministic control around the model
A business application cannot throw a prompt at a model and hope. Real automation uses workflow engines such as n8n or Temporal to treat the model as one conditional step. If an output fails a schema check, the system retries it or hands the task to a person. The value is in the system, not the text box. See human-in-the-loop design for the patterns.
2. High-quality retrieval
A model is only as informed as the context you give it. Serious systems invest in data pipelines: semantic caching to reuse good answers, hybrid search that combines keywords with vector search, and careful chunking so the model never reads irrelevant text. Our overview of retrieval-augmented generation explains the basics.
3. Proprietary data loops
The strongest moat is verified data that cannot be scraped from the public internet: your operational records, compliance changes or aggregated industry insights. The model becomes interchangeable, because you can swap it to cut costs, while the data pipeline stays yours.
A five-question test for your own product
- Could a developer rebuild the core feature from the prompt alone?
- Does the product connect to data or systems the customer cannot easily connect themselves?
- What happens to the product if the model provider adds the same feature next quarter?
- Do you measure accuracy on real cases, and is there a fallback when the model is wrong?
- Does every extra user make the product better, or only more expensive?
Questions to ask before you buy AI software
- Which models does it use, and can they be changed?
- What is connected to my systems, and with what permissions?
- How are errors detected, and who reviews them?
- Where does my data go, and is it used for training? See our AI data privacy checklist.
- What happens to my workflow if the vendor shuts down?
If you are deciding between a product and a custom system, read build vs buy for AI agents.
What to build instead
Stop rewriting your system prompt in search of magic words. Build the unglamorous parts: data ingestion, error handling, validation and integrations. Those are the parts that make automation reliable at scale, and they are what a customer will still pay for after the next model release.
Related guides
- Build vs Buy: Custom AI Agents or Off-the-Shelf Tools?
- AI SaaS Pricing Models and Unit Economics: Keeping Margins When AI Costs Money
- LLM API Cost Control for SaaS: Caching, Routing and Limits That Protect Your Margin
- What Does It Cost to Build an AI Agent? The Real Cost Drivers
- Service: AI Workflow Automation
Frequently asked questions
Is every AI wrapper a bad business?
No. Wrappers can be useful and profitable, especially in a narrow niche with distribution or brand. The risk is relying on the prompt as your only advantage.
How can I tell if an AI tool is just a wrapper?
Ask what it connects to, how it handles errors and what happens if the underlying model changes. Vague answers suggest a thin product.
What is an AI moat?
Something a competitor cannot easily copy, such as proprietary data, deep integrations, workflow ownership or trust built over time.
Do bigger models make wrappers obsolete?
They make generic features cheaper and easier to rebuild. Products grounded in your own data and workflows are less exposed.
Where should a small business start?
With one narrow, measurable task connected to your own systems, as described in how to find tasks worth automating.
Sources and further reading
Primary documentation and research behind this guide.