Why "In One Week" Is Realistic
When most people hear "AI assistant for business", they imagine months of development and hundreds of thousands of dollars. But in 2026 the reality is radically different. Claude API, LangChain, Streamlit and cloud vector databases let you build a working prototype in 5–7 working days — one developer, real data, real users.
Below is the exact day-by-day plan we use ourselves and recommend to clients.
Day 1: Define Use Case and Success Metrics
The most important day. A bad use case kills the project before the first line of code is written.
A good first AI assistant: solves a real, frequent problem (daily, not once a year), has a clear input and output, is easy to measure, and has a data source (documents, FAQ, previous chat logs).
Ideal first use cases: customer FAQ bot (website or messenger), internal HR assistant for new employees, sales helper (product questions, proposal generation), summarizer for emails/reports/meeting notes, classifier for routing support tickets by category and priority.
Day 1 deliverable: one page — problem, use case, target users, success metrics, scope (what the assistant does NOT do in v1).
Day 2: Prepare the Knowledge Base
Collect, clean and structure documents: FAQ pages (HTML → plain text via BeautifulSoup), PDFs (pypdf or pdfplumber), Notion/Confluence (official APIs), Google Docs, previous support tickets.
Clean the data: remove duplicates and outdated information, unify formatting, ensure each document has metadata (source, date, category), remove or anonymize personal data.
Then chunk and embed using LangChain's RecursiveCharacterTextSplitter (500 tokens, 50 overlap) and OpenAI text-embedding-3-small into Chroma for the prototype.
Day 2 deliverable: vector database with all documents, verified on several test queries.
Days 3–4: Choose Model and Build MVP
Model recommendation: Claude 3.5 Sonnet for business assistants where quality matters, Claude 3 Haiku or GPT-4o mini for high-volume FAQ bots.
Build a LangChain RetrievalQA chain with a system prompt that instructs the model to answer only from provided documents and say "I don't know" if the answer isn't found — with a fallback to human support contact. Add a Streamlit chat interface that shows source documents for every answer.
Days 3–4 deliverable: locally running MVP that answers questions and shows sources. Ready for internal testing.
Days 5–6: Test With Real Users
Recruit 5–10 real potential users (not the developer or project manager). Give them tasks. Observe: where they get stuck, what they rephrase, what frustrates them.
Test: happy path (direct questions with answers in docs), edge cases (vague questions, synonyms, typos), out-of-scope questions (does it refuse correctly?), basic prompt injection attempts.
Add 👍/👎 buttons after each response. Log: query, response, rating, found sources. This is a gold mine for iteration.
Common fixes: not finding relevant docs → improve chunking or metadata, try hybrid search; too generic answers → lower temperature, add specificity to system prompt; hallucinations → strengthen "answer only from documents" instruction, add reranking.
Day 7: Deploy and Scaling Plan
Deployment options: Streamlit Community Cloud (free, good for internal use), Render or Railway ($7/month), Docker + Hetzner VPS (€4/month, full control), Telegram bot (ideal if team is already on Telegram).
Monitoring: log all queries and responses, alert if API is unavailable or response takes >15 seconds, weekly review of negative 👎 ratings.
Scaling plan: weeks 2–4 — expand knowledge base, improve prompts from feedback; month 2 — integrate into main channel (site, CRM, Slack), set up human escalation; month 3 — analyze metrics, A/B test prompts, possibly switch model based on cost/quality analysis; month 4+ — expand use cases based on real query data.
First Month Costs
| Cost Item | Amount |
|---|---|
| MVP development (1 dev, 1 week) | $800–1,500 |
| Claude API (Sonnet, ~500 queries/day) | $30–80/month |
| OpenAI Embeddings (one-time) | $2–10 |
| Hosting (Render or VPS) | $7–20/month |
| Vector DB (Pinecone serverless) | $0–25/month |
| Total first month | $840–1,635 |
| Monthly after | $40–130 |
Conclusion
One week and ~$1,000–$1,500 is a completely realistic budget for a first working AI assistant. The key: correctly chosen use case and clearly defined success metrics before starting. Everything else is technical details solved along the way.
Want to take this journey with us? We run AI Discovery sessions where in 2 hours we identify the optimal use case for your business and estimate the MVP cost. Contact us.