Adrien Guilmineau

The Future of SaaS lives in ChatGPT

How OpenAI might build the ultimate platform

The Future of SaaS lives in ChatGPT How OpenAI might build the ultimate platform I don’t want to leave ChatGPT. Not for payroll, not for editing media, not for building spreadsheets, not even to pay vendors. I want ChatGPT to do it. Yet, as soon as something more complicated than searching publicly available information is required, I’m stuck clicking on URLs, tabbing into a SaaS dashboard, creating an account, learning a new UI or connecting tools together. It feels like a relic of the past, and I don’t think I’m alone. So I want to outline how OpenAI might support this growing user demand and what it might mean for SaaS. Generative Interfaces OpenAI made the decision to design ChatGPT like a messaging app. This was a brilliant move that helped them introduce AI to billions of people for the first time. 3 years later people use ChatGPT for so many things, that it’s been thinking about improving certain responses. Document editor Earlier this year, ChatGPT introduced its first major UI change: the document editor, a side panel for editing long text documents. This new feature is contextual. It opens up when ChatGPT feels it could be useful. It cannot be summoned by pressing a button or navigated to. When the document editor is open, the chat interface stays open so that you can ask ChatGPT to change something, but it also allows the user to select, edit and format text like you would in a normal text editor. This is a great example of how OpenAI is thinking about additional interfaces that can make it easier, in certain contexts, for ChatGPT to help its users achieve their tasks. One window, infinite combinations It feels natural to imagine OpenAI generalizing this concept to a lightweight window which only appears when the task at hand requires it. Could ChatGPT spin up a basic spreadsheet? a simple dashboard? a map view with your trip itinerary? a step sequencer to write down a beat? a chess board to learn how to improve your game? a flash card deck to study for your upcoming test? Imagine asking: “Can you help me track my spending more efficiently?” ChatGPT would ask you if you would like it to build a simple budgeting app. You’d accept, and it would code up a tracker for you, right there in the chat. You’d be able to interact with your app both through the window and by continuing to chat - a pattern that’s becoming table stakes for all new AI-native software. I’m most excited by how modular and personal software could become. Your budgeting app could be asked to: “count everything in terms of hours of my time instead of dollars” or to “allow me to tag when transactions are joint with my wife and when they are personal” or even to “reflect in my yearly budget, that my parents pay for half the children’s school bill.” A budgeting app that perfectly fits your needs, built in real time. This is what I believe users want. This is what I want! Generative Database Once users can build infinite little apps directly inside ChatGPT, I think OpenAI will need proper long-term deterministic memory to keep track of booking IDs, tasks, dates, uploaded documents and more. Without this, users will find it hard to rely on their apps for anything meaningful. ChatGDB That’s why I believe ChatGPT will roll out its own database and file storage — a place where every generated app can safely store and retrieve information. The question for OpenAI is: how do you build a database without knowing what the app does or what it will do in the future? That’s what I call a “Generative Database”, a modular database technology that can infinitely adapt to what the user needs. Each app will have its own schema. There will be modular datatypes, secure inputs, read/write rules, record IDs and relationships between tables. It’ll have private files, limited to single chat, and shared files so multiple apps can reference them. It’ll seamlessly migrate data as the app evolves and the needs of the software change. A whole suite of tools enabling ChatGPT to host the millions of apps its users create. KitApps - the successor to SDKs Once we have generative interfaces and databases, ChatGPT will be far more useful than it is today. People will start building increasingly advanced software inside the chat: calendars, project trackers, budget tools, mini CRMs and thousands of other productivity-enhancing tools. But some things can’t be vibe-coded, and just like Lovable or Replit can be great at building a prototype HRIS, when domain logic becomes too important, you need expertise to set clear rules and processes which force the software to behave in a certain way. Business Logic What fields do I need to represent compensation accurately? Are performance bonuses and sales commissions stored in the same field? How are RSUs reflected in compensation? What about stock options? How do you include part-time workers in your headcount reporting? Should EOR employees be stored in the same entity as the full-time employees? Can an employee have multiple line managers? That’s what we refer to as “business logic”. Every domain has its own structure: CRMs need opportunities, contacts, and stages. Payroll has employees, salaries, and tax rules. Budgeting has items, categories, and transactions. These rules are built and validated by product managers, software engineers, lawyers and accountants who form a higher level vision of how software needs to work based on experience, regulation, best practices, interoperability, future-proofing etc… When you buy SaaS, you’re buying the business logic that’s mapped to the problem you are trying to solve. KitApps I imagine that while ChatGPT will technically be able to vibe code just about anything, just like you could build your own payment back-end, ChatGPT will want to use a SDK-like package that increases the reliability of its apps. I call those KitApps. Let’s take the example of a fictitious KitApp called “BudgetKit” which would help ChatGPT build great personal finance products. I think it would contain 4 main components: Schema: it tells ChatGPT what a budget item is, how budget categories work, how to define time periods, perhaps constants like end of tax years in certain countries, how to parse a mortgage correctly... Taste/Opinions: it encodes typical budgeting rules and best practices. Tags common categories with “investments“, “income from work“, “income from capital“, “day-to-day expenses“, “leisure“,“rent“, “utilities“, etc… It perhaps provides guidance to ChatGPT on how to build a good budgeting app UI. Audited operations: it sets out useful functions like estimate_end_of_year_income_tax_liability()— a battle tested function, reviewed by a tax advisor working for BudgetKit, that ChatGPT can rely on to provide tax guidance to the user. Link to resources: it might provide recommendations to other services like useful APIs (or MCPs) ChatGPT could use, other KitApps developed by the same company or even common integrations the KitApp works well alongside of. Much like an SDK, the makers of BudgetKit won’t know how their SDK is being used. This means that KitApps will be fully composable. One simple app might contain dozens of KitApps that all help ChatGPT build quickly and reliably. Connecting to the internet Generative Interfaces, Generative Databases and KitApps working together will provide insane value for consumers and small businesses. It’ll allow ChatGPT to build perfectly tailored software (infinitely flexible and adaptable), updating dynamically to always match what the user needs. Sometimes, certain apps will gain from connecting to the internet: pulling a stock price, retrieving data from your bank account, searching your email for a receipt, looking up flight options, booking a hotel room, sending a payment, updating the status of a task in your company’s task manager… There are signs that MCPs might be the technical solution to the problem - a standard way for ChatGPT to access millions of softwares directly from its interface, but I’m unsure whether this is what the final user experience will look like. LLMs are indeterministic, so I struggle to imagine a world where millions of consumers and small businesses pass sensitive data to external services that might be buggy or malicious. It feels like a disaster waiting to happen. Imagine ChatGPT accidentally sharing a blood test PDF with an insurer evaluating your mortgage application — that’s catastrophic. Or an MCP that quietly requests too much data, stores it, and later sells it for ad targeting. Not acceptable at scale. If OpenAI takes inspiration from Apple (and Sam has hinted they do), the safer and more scalable path will be to build a walled garden within ChatGPT itself. I like to think that accessing external services will require validation from the user. Perhaps there will be an agent managing data passing to and from external services, ensuring that data is anonymized or that the authentication token is passed correctly. One way to think about this is how iOS services like location, photo gallery or HealthKit prompt the user to approve each access. I think ChatGPT will be infinity more nuanced here, understanding intimately context and intent, but it feels like a good way to think about it. I think the right way to think about MCPs is websites on the iPhone — useful, but clearly not the native experience. Sometimes, an app might open a web view and we may have certain MCPs pop in and out of some experiences on ChatGPT. But eventually, native apps will take over, just as apps rule the iPhone. ChatGPT is the final platform There are 3 main assumptions here that I’m making: People want ChatGPT to take care of everything Linking to existing software is clunky and will never feel seamless ChatGPT will have unlimited resources to solve this because it’ll provide trillions in value to users For ChatGPT to achieve this it needs 4 things: Generative interfaces - a interactive window powered by codex Generative databases - a deterministic fully-featured data hosting suite KitApps - some sort of verified business logic A walled garden - something to manage permissions and access to user data The consequence of this cannot be understated. I believe there will be an incredible disruption to incumbent SaaS products and the emergence of a new business model: KitApps. It will change the way we consume software. First a KitApp that gives ChatGPT the verified tools it needs to help users build their own version of Duolingo will go viral on TikTok. “that’s cute!” - app developers. People will create a small tracker for their hiring pipeline inside of ChatGPT and only think about getting an ATS much later in their growth journey. “That’s a very simple version, but yes it works.“ - SaaS vendors. It won’t be sudden, but bit by bit, we’ll see bigger and bigger tools built directly in ChatGPT. “The next big thing will start out looking like a toy.” People thought SaaS was a toy and look at where we are now.

Source