AI wrapper: when your product becomes a button
Hugo Chamberland
7 min
Every AI wrapper founder should read Chegg's 2025 annual report. Revenue fell 39%, and the filing gives the cause plainly: reduced traffic, fewer subscribers. No competitor copied Chegg's catalogue. Students started asking a free model instead. Absorption rarely looks like competition.
Tanya Puram summed up the pattern in When Your Startup Becomes a Button: what was a startup in 2019 and an AI startup in 2023 is a button in 2026. This article looks at what makes a product absorbable, three cases with verified numbers, and a test to run before your next build.
An AI wrapper doesn't get copied. It gets absorbed.
What is an AI wrapper? An AI wrapper is a product whose core value comes from a general-purpose model it does not own, packaged with an interface, templates or prompts. It works as long as the model provider does not ship the same capability. When it does, the wrapper loses its reason to exist overnight, often without a single competitor involved.
The classic startup threat was a rival doing the same thing cheaper. The threat here is different: the platform you build on adds your product as a feature. Meeting summaries, PDF chat, resume generators and basic chatbots have all gone this way, as model providers shipped transcription, file uploads and custom assistants natively.
What the model absorbs first
Three traits make a product easy to absorb: little differentiation beyond the model's output, low switching costs, and a value proposition the model already delivers on its own. A product with all three is not a company yet. It is a feature waiting for someone else's release notes.
The Belgian market shows where that line runs. According to Statbel's 2025 ICT survey, 34.5% of Belgian companies use at least one AI technology, and the most common one is automatic analysis of written language, used by 23% of firms. That is exactly the capability general models now bundle for free. A product that sells only that is selling what its buyers already have.
💡 The one-line test: if a better model, or a feature the platform ships for free, would replace most of what you sell, you built a feature. If your product gets better when the models improve, you are building a company.
What survives absorption: three advantages the model can't ship
Not every AI product gets absorbed. The ones that last own something the model provider either cannot replicate or has no reason to build in depth.
1. Data that compounds with use
Proprietary data generated by real usage, which makes the product better over time in ways a public dataset never could. We covered why this is the most durable moat in Why some SaaS products become indispensable while others disappear.
2. Workflow the customer runs on
Deep integration into the customer's systems and multi-step processes, so the product becomes how the work gets done rather than a helper next to it. A model provider ships horizontal features. It does not map the approval chain of a Belgian logistics company.
3. Outcomes, not tokens
Selling a measurable result (hours saved, errors avoided, revenue recovered) instead of access to generation. When the price is tied to an outcome, falling API costs widen your margin instead of eroding your reason to charge.
⚠️ The trap: adding AI features to stay current. Most AI features startups race to ship are already on the roadmap of the platforms they build on. How to qualify one before building it is covered in Building on AI without asking if it will still matter in a year.
Three companies, three outcomes: absorbed, pivoted, compounding
1. Chegg: absorbed
Chegg sold curated homework answers. A model that explains each step and adapts to the student's question does the same job for free, and AI answers now appear in search before students reach Chegg's site.
Results observed:
- Total revenue down 39% in 2025, to $376.9 million; academic services down 43% (10-K)
- Paying subscribers down 43% year on year in Q3 2025 (10-Q)
- Strategy turned toward workforce skilling, as a survival plan rather than a growth one
2. Jasper: forced to pivot
Jasper sold templates and tone controls for marketing copy. ChatGPT, then native writing features in office suites, made that generation a commodity.
Results observed:
- 2023 revenue forecast cut by at least 30%, internal valuation lowered to about $1.2 billion, CEO replaced (Maginative)
- Pivot to enterprise marketing teams, with brand governance and workflow features: enterprise ARR multiplied by four (Contrary Research)
3. Cursor: compounding
Cursor started, like many others, on top of third-party models. Instead of stopping at autocomplete, it built a daily workflow that development teams find painful to leave.
Results observed:
- $2 billion in annualised revenue by February 2026 (TechCrunch)
- Over $4 billion by early June 2026, about 75% from enterprise customers (Dealroom)
👉 Same kind of models, opposite outcomes. All three built on models they did not own. What separated them is what they built around the model.
Feature or company? Five questions before your next build
Your buyers already hold the model. The Conseil supérieur de l'emploi found that one Belgian worker in three used generative AI at work in 2025. Whatever you sell gets compared, daily, with what they can already do in a chat window.
| Question | Feature signal | Company signal |
|---|---|---|
| What happens if the model gets twice as good? | Your product becomes unnecessary | Your product gets better |
| What does usage leave behind? | Nothing you own | Data that improves the next result |
| What does switching cost the customer? | An afternoon | Retraining a team, rewiring a process |
| What do you charge for? | Access, seats, credits | A measured outcome |
| Whose process does it run? | A generic one | The customer's own |
Three company signals or more is a product worth building. One or none is a feature, and features belong inside someone else's roadmap.
The honest limits of this test
Surviving today proves little about next year. Cursor now competes with coding tools from the very model providers it relies on, as TechCrunch notes. And some widely shared figures on wrapper churn have no traceable source, so we left them out. The test sorts risk; it does not remove it.
🔍 Run it on your roadmap, not just your product. Score each planned AI feature separately. The ones that land in the left column are candidates to drop before a single sprint is spent on them.
Four mistakes that turn an AI product into a button
1. Pricing per generation
When the unit you sell is a token or a credit, every drop in model prices becomes a drop in your perceived value.
2. Treating the prompt as the product
A prompt is copied in minutes; the data, integrations and process around it are what take a competitor months.
3. Building horizontal when the model is horizontal
A generic assistant competes head-on with the platform, while a vertical tool for one trade competes with nobody's roadmap.
4. Skipping distribution until the product is ready
A wrapper with a thousand paying users has time to deepen its product; one with none has only the next model release to fear.
💡 Pattern behind all four: each one makes the model the centre of the product, when it should be one component inside a system of work that is harder to copy.
How we qualify an AI product before building it at Nightborn
Step 1: score the idea, feature by feature
In a four-hour scoping workshop, we run each planned AI feature through the five questions above, against the model providers' published roadmaps. Deliverable: a short list of what to build, what to drop, and what to wait for. The format is on our Product Discovery page.
Step 2: design the part the model can't ship
We decide before the first line of code where the proprietary data will come from, which customer workflow the product plugs into, and what outcome it will be priced on. The model integration itself is the easy part (see our AI Integration page).
Step 3: build with distribution wired in
A fixed-price build of two to four weeks, with a Friday demo every week. Landing page, tested pricing, payment and analytics ship with the first version, so the MVP proves a model you can repeat, not a hunch. More on our MVP Design & Development page.
📌 Checklist before you ship an AI feature: it survives a twice-better model, it leaves data you own behind, it sits inside the customer's process, and its price is tied to an outcome.
AI wrapper or company: the model is a component, not the product
What to remember:
- An AI wrapper gets absorbed, not copied: the platform ships your product as a feature.
- Chegg and Jasper show the two exits left once that happens: decline, or a costly pivot toward workflow.
- Cursor shows the alternative: a workflow deep enough that better models make it stronger.
- Five questions separate a feature from a company; three company signals or more is worth building.
The implication is practical. Before the next sprint, score every AI feature on your roadmap. Drop the ones a better model would replace, and put the saved weeks into the data, the workflow and the pricing that make the rest hard to copy. For the structural reasons most AI startups fail, read Most AI startups are building on sand. If you want a second opinion on your own roadmap, that is exactly what a scoping workshop is for.
