Almost every client now asks us how AI can fit into their product. The honest answer: AI is powerful when it solves a specific problem, and disappointing when it is added for its own sake. This roadmap is the process we use to find the right AI feature and ship it safely.
1. Where AI creates real value
Large language models (LLMs) are very good at working with language and unstructured information. The highest-value features usually fall into a few patterns.
- Summarise: long documents, support tickets, meeting notes or customer feedback.
- Draft: emails, product descriptions, replies and reports that a human then reviews.
- Classify and extract: route tickets, tag content, or pull fields out of invoices and forms.
- Search and answer: let users ask questions about your docs or data in plain language (see our RAG guide).
- Act: AI agents that complete multi-step tasks using your tools (see our AI agents guide).
“Start with a painful, repetitive task your users already have. AI is the method, not the goal.”— Strategic Takeaway
2. Choosing a model provider
The main providers (OpenAI, Anthropic, Google and open-source models hosted by various platforms) all offer strong models in several sizes. Larger models reason better but cost more and respond slower. Smaller models are cheap and fast, ideal for classification and simple drafting.
Model lineups and prices change every few months. We therefore never hard-code a model name deep in the codebase. It lives in an environment variable, and we use a provider-agnostic SDK so switching takes minutes.
3. Your first AI feature with the Vercel AI SDK
The Vercel AI SDK is a free, open-source TypeScript library with one consistent API across providers. Here is a complete server action that summarises customer feedback.
npm install ai @ai-sdk/openai zodOPENAI_API_KEY=sk-...
AI_MODEL=your-chosen-model # pick from the provider's current model list"use server";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";
export async function summariseFeedback(feedback: string[]) {
const { text, usage } = await generateText({
model: openai(process.env.AI_MODEL!),
system:
"You are a product analyst. Summarise customer feedback into the top 3 themes, " +
"each with one sentence and an example quote. Be concise and neutral.",
prompt: feedback.join("\n---\n"),
});
console.log("Tokens used:", usage); // log usage to track cost
return text;
}4. Shipping safely
AI output is probabilistic, so quality and safety need deliberate design.
- Build a small evaluation set of 20–50 real examples and check outputs before every prompt or model change.
- Keep API keys on the server only; never call model APIs directly from the browser.
- Add rate limits per user to prevent abuse and runaway costs.
- Require human approval for anything that sends, deletes, pays or publishes.
- Tell users when content is AI-generated and give them an easy way to correct it.
- Don't send sensitive personal data to a model unless your provider agreement and privacy policy cover it.
5. Understanding the cost
Model APIs charge per token (roughly ¾ of a word), with separate prices for input and output. Output tokens are usually several times more expensive than input. Most providers also discount repeated (cached) input and offer cheaper batch processing for work that doesn't need an instant response.
A simple formula keeps budgets honest: requests per month × average tokens per request × price per token. We calculate this during discovery, so clients know the running cost before we build.
Official resources
Principles for deliberate execution
- 1The best first AI features remove repetitive work: summarising, drafting, classifying and searching.
- 2Use a provider-agnostic SDK so you can switch models as prices and quality change.
- 3Measure quality with real examples before launch, and keep a human in the loop for important actions.
- 4Track cost per request from day one; small prompt changes can multiply your bill.