AI SDK by Vercel: What It Does and Where Businesses Use It
AI SDK is Vercel's open-source toolkit for adding AI features to software. Here is how MILabs sets it up and where it fits.
Updated 2026-04-22
What AI SDK is
AI SDK is an open-source toolkit from Vercel, written in TypeScript, for building software that uses large language models. It gives developers one interface for calling several model providers, so application code keeps working when the model changes. It covers common tasks such as generating text, streaming a response while it is produced, letting the model call functions the application defines, and returning structured data a program can read. It also includes helpers for chat-style screens in popular web frameworks. The toolkit is a library rather than a chat product. A developer builds an application with it, and that application decides what the model can see, which tools it can use and what happens to each answer. Because the code runs in the business's own environment, the owner can choose where it is deployed and which model providers it uses. For a small business, it usually sits underneath a custom assistant, a drafting tool or an internal lookup that reads approved documents. It suits work that has to live inside existing software.
AI SDK in plain language
Large language models are reached through an API. The application sends a prompt and receives text back. Each provider has its own request format, authentication method and response shape, so code written directly against one supplier must be reworked when the business changes. A layer that smooths those differences keeps the application logic in one place.
A feature built on a model rarely sends a single question. It keeps a running conversation, supplies fixed instructions, attaches reference material and decides when a result needs another step. Each of these parts has to be written and tested by someone who understands both the software and the model's habits.
Tool calling lets the model request a function the application has defined, such as one that looks up an order. The model proposes the call with specific inputs, but the application runs it and decides what comes back. Permissions therefore live in ordinary code that can be reviewed.
Structured output asks the model for data in a fixed shape, so a CRM or spreadsheet receives fields it can store. Streaming displays text as it is generated, which keeps a chat window responsive. Model choice stays in configuration, so changing provider becomes a contained task.
What businesses use it for
- Answering customer questions on a website using the business's own policies and product information.
- Drafting replies to enquiries, quotes and reviews for staff to check before anything is sent.
- Condensing long documents such as contracts, tenders or specifications into a summary for the team.
- Pulling fields from invoices, forms or emails into a structured record.
- Giving staff a chat assistant inside an internal tool so they can ask about company procedures.
Who it suits
- Teams with a developer or contractor who can maintain the application
- Owners who want an AI feature inside their own website or software
- Businesses that may switch between AI model providers over time
How MILabs sets it up
The first step is to say hello through the enquiry form, and the owner or team lead then agrees which job the AI feature should do, what information it may read and what it must never do. The model provider and plan are chosen together, and the business confirms the terms with that provider. Next we connect AI SDK to the systems the business already uses, such as the website, the booking tool or the shared drive. Permissions are set so the feature reads only what it needs and changes nothing without agreement. Before handover we write down who looks after the feature, how errors are reported and who reviews changes, so the business is not reliant on one person's memory.
Builds it fits
Industries it fits
AI SDK in practice
In a working business, the feature usually appears as a small panel inside a tool staff already open, such as the enquiry inbox or the website chat. The person using it sees a suggested answer, the source passages it drew on, and a control to accept, edit or discard. That design keeps the human decision visible.
Behind that panel, the team keeps example questions taken from real past cases. Whenever the instructions or model version change, those examples run and the results are compared with answers a senior person has approved, and differences are reviewed before anything reaches staff.
Well-built applications also define what happens when the model cannot help. A question outside the approved material receives a plain message saying so, together with a route to a person. Logs record which documents were used, which tool calls ran and how long each step took, so the owner can answer a complaint by reading the record.
Costs tend to follow usage. The owner receives a regular summary of request volumes by feature, which shows whether each function justifies its running cost. Where a feature is rarely used, the team may simplify or retire it.
Frequently asked questions
- What is AI SDK?
- AI SDK is an open-source TypeScript toolkit from Vercel for building applications that use large language models. It provides one interface for several model providers, covering text generation, streaming, tool calls and structured data. Developers add it to their own software, so the result is a custom feature that the business owns. Staff use the feature built on it, not the SDK itself.
- How does AI SDK compare with other tools for the same job?
- Ready-made assistants such as ChatGPT, Claude and Gemini give staff a finished chat window, but it sits outside the business's own systems. AI SDK builds that capability into your own software, so you decide the interface, the data sources and the permissions. Developer frameworks such as LangChain and Mastra cover overlapping ground with different structures. The right choice depends on whether the business needs a product to use or an application to build.
- How do I get my team to use AI SDK?
- Staff rarely use a developer library directly. They use the feature built on it. The practical approach is to pick one task that already causes repeated effort, build the feature around that task, and place it inside the tools people already open. We agree with the team lead what the feature should and should not handle, then gather feedback from the people who use it and adjust the instructions to match real work.
- Can I self-host AI SDK?
- Yes. AI SDK is a library that runs inside your own application, so it can be deployed on servers the business controls. Self-hosting decides where the application runs. The SDK sends requests to whichever model provider is configured, so the model runs with that provider. If the model itself must run on your own hardware, that is a separate decision, made with the model server in mind and then connected to the application.
- Which builds work well with AI SDK?
- Builds that answer questions from a defined body of information suit it well, such as a client FAQ assistant or a staff knowledge assistant that reads approved documents. Lead intake and quote request assistants also fit, because the SDK can return structured fields that a CRM or form can store. Email triage and review reply drafting work well when a person checks each output before it is sent.
