A good AI chatbot answers the question your customer or employee actually asked, from your own content, and admits when it doesn't know. A bad one sounds confident and makes things up. The difference is almost never the AI model — it's how the chatbot is scoped, grounded and maintained.
This guide walks through how we plan and build custom AI chatbots for clients, and what we've learned running AI in our own products. If you're comparing quotes, pair it with what drives AI chatbot cost.
Why a custom chatbot instead of an off-the-shelf bot
Subscription chatbot tools are quick to switch on and fine for simple FAQs. Businesses usually move to a custom build when they hit one of these limits:
- Your content is the product. Answers must come from your catalogue, policies, contracts or manuals — with sources — not general internet knowledge.
- Data rules. Documents can't leave your cloud account or your servers.
- Integrations. The bot needs to read or write to your CRM, helpdesk, order system or internal tools.
- Per-seat or per-conversation fees that grow faster than the value.
- Ownership. You want the code and configuration, not a rented account.
Pick one use case to start
The fastest route to a chatbot people actually use is one clear job, done well. These are the use cases we build most often:
| Chatbot type | What it does | Content it learns from |
|---|---|---|
| Website AI support agent | Answers customer questions, captures and qualifies leads, books meetings | Website pages, FAQs, product documents |
| E-commerce AI chatbot | Product discovery, order and delivery questions, returns, cart recovery | Product catalogue, shipping and returns policies |
| Internal knowledge assistant | Lets staff find HR, IT and process answers in a portal, Slack or Teams | HR policies, SOPs, IT guides, onboarding material |
| Document AI assistant | Answers, summarises and compares long documents with citations | Contracts, reports, regulations, manuals |
| Compliance and legal assistant | Extracts clauses, deadlines and obligations; flags risks | Laws, circulars, contracts, policies |
Write down the ten questions the bot must answer perfectly on day one. If you can't, the use case isn't clear enough yet. For AI use cases beyond chatbots, see AI for business: 15 practical use cases.
How a custom AI chatbot is built
Almost every business chatbot we build uses retrieval-augmented generation (RAG): your content is indexed, the most relevant passages are retrieved for each question, and a language model writes the answer from those passages only. It's the reason the bot can cite a source and stay current without retraining. Our RAG development guide explains the pipeline in depth; in short, a chatbot has five parts:
- Content ingestion — PDFs, Word and Excel files, web pages, database records and product exports, cleaned and split into passages.
- Search index — embeddings in a vector database (plus keyword search where it helps), with metadata for filtering and access control.
- Language model — GPT-4o, Claude, Gemini, or an open-source model such as Mistral or Llama, chosen for quality, cost and data rules.
- Chat interface — a website widget, an internal portal, or Slack/Teams, branded to match you.
- Admin panel — upload documents, trigger re-indexing, read conversations, see unanswered questions and export leads.
The build process, week by week
Our typical custom chatbot project runs 6–8 weeks from kickoff to production:
| When | Stage | What happens |
|---|---|---|
| Week 1 | Discovery | Agree the use case, the must-answer questions, data sources and how success will be measured |
| Week 2 | Architecture | Choose the model, search index, hosting and integrations |
| Weeks 3–5 | Build | Document ingestion, AI backend, chat interface and admin panel |
| Week 6 | Testing and tuning | Accuracy review against real questions, edge cases, prompt improvements |
| Weeks 7–8 | Launch | Staging and user acceptance testing with your team, then production launch and monitoring |
From your side, we need the documents or data the chatbot should know, one person who understands the subject and can check answers during testing, a decision on which AI provider you're comfortable with, and brand guidelines if it's customer-facing.
Automated checks catch answers that are wrong. Only a subject expert catches answers that are technically correct but unhelpful, incomplete or worded in a way your customers would misread. Book that person's time before the project starts.
Integrations: website, CRM, Slack and Teams
- Website: a lightweight JavaScript widget added with one script tag — works on WordPress, Shopify, Webflow, plain HTML and custom React or Vue apps.
- CRM: push captured leads to HubSpot, Salesforce or Zoho; personalise answers for logged-in users; create helpdesk tickets (for example in Freshdesk or Zendesk) when the bot can't resolve an issue.
- Slack and Microsoft Teams: the usual home for internal knowledge assistants, with access limited by channel or team.
- API: a REST API so you can build your own interface — a mobile app, a portal or a voice front end.
More detail is in our integrations FAQ.
Hosting, privacy and ownership
There are three sensible hosting models: cloud-hosted with an LLM API, deployed inside your own AWS or Azure account, or fully on-premise with open-source models so nothing leaves your network. For confidential material, add role-based access so users only retrieve documents they're allowed to see, and audit logs of every query.
Ownership is worth settling in writing before you start. On our projects, you receive the full source code, index configuration, prompt templates and deployment scripts at handover — no lock-in.
After launch: keeping answers accurate
A chatbot is never "finished", because your content isn't. Plan for:
- Weekly review of unanswered and low-confidence questions — these are your content gaps.
- Re-indexing when documents change — from the admin panel, in minutes.
- Clear fallbacks — "I couldn't find that" plus a human or contact route, rather than a guess.
- Cost monitoring — API usage grows with traffic; caching and model routing keep it in check.
If the chatbot is part of a wider support operation — triage, agent assist, helpdesk tickets — read how to build an AI-powered customer support system.
Our projects include a 30-day post-launch support window for bug fixes and accuracy tuning based on real usage, after which you can self-manage or move to a support retainer.
Questions to ask any chatbot developer
- Will answers come only from our content, with citations users can click?
- What happens when the bot can't find an answer?
- How will you test accuracy before launch — with whose questions?
- Where will our documents and conversations be stored and processed?
- Do we own the code, prompts and index at the end?
- Can our team add or update documents without a developer?
- What will it cost per month to run at our expected volume?
- Have you run an AI system in production yourselves? Can we see it?
That last question is why we point people to our own products: our case studies show two AI platforms we built and run ourselves.