A production AI chatbot grounded in your own data typically takes 4–6 weeks to ship. A scoped prototype you can click through takes 1–2 weeks. A multi-system assistant that takes actions across your tools — with human handoff, guardrails, and evaluation — runs 8–10+ weeks. The spread comes almost entirely from three things: the state of your data, how many systems the bot must touch, and how much proof you need that it won't say something wrong to a customer.
This is the timeline companion to our cost breakdown — if you're budgeting rather than scheduling, read how much an AI chatbot costs alongside this.
What are the realistic timeline tiers?
| Tier | Timeline | What you get | Right for |
|---|---|---|---|
| Scoped prototype | 1–2 weeks | A working chat UI over one model with a slice of your real docs; enough to test answer quality on your actual questions | Validating the idea before committing budget |
| Production RAG assistant | 4–6 weeks | Retrieval over your full knowledge base with citations, brand-safe prompting, analytics, one channel (web widget or Slack), and an evaluation set | Customer support, internal Q&A, sales assist |
| Multi-system agent | 8–10+ weeks | Everything above plus actions (create tickets, update CRM, trigger workflows), human handoff, role-based access, and audit logging | Automation where the bot does work, not just answers |
Where do the weeks actually go?
Here's the week-by-week shape of a typical 5-week production build — the middle tier above. The model integration is the fast part; the work that makes it trustworthy is what takes time.
- Week 1 — discovery and data audit. Define the questions the bot must answer, the ones it must refuse, and where the truth lives (help docs, PDFs, Notion, tickets). Messy or contradictory source content found here is the single most common cause of delay.
- Week 2 — retrieval pipeline. Ingest, chunk, and index the content; wire retrieval-augmented generation (RAG) so answers come from your data with citations rather than the model's imagination. Our RAG pipeline guide covers this architecture in detail.
- Week 3 — answer quality. Build an evaluation set of real questions with known-correct answers, then iterate on chunking, retrieval, and prompting until accuracy holds. This is the step teams skip when a chatbot embarrasses them later.
- Week 4 — integration and guardrails. Embed the widget or connect the channel, add refusal behavior for out-of-scope and sensitive topics, rate limiting, and abuse handling; wire analytics so you can see what users actually ask.
- Week 5 — hardening and launch. Load testing, fallback behavior when the model API degrades, human-handoff path, and a soft launch to a subset of users before full rollout.
What makes a chatbot build take longer?
- Source content that disagrees with itself. If your pricing page and your help docs give different answers, the bot will too. Cleanup adds 1–2 weeks and is worth every day.
- Waiting on API access. Every system the bot reads from or writes to needs credentials, and enterprise approval loops for those regularly cost more calendar time than the engineering.
- Compliance review. Legal or security sign-off for what the bot may say and log — common in fintech and health — can add weeks that have nothing to do with code.
- Fine-tuning when you don't need it. Most business chatbots need retrieval, not fine-tuning; training custom models adds weeks and rarely improves grounded accuracy. Our prompting vs fine-tuning vs RAG comparison explains when each applies.
How do you compress the timeline?
The builds that ship fastest share three decisions: they start with one channel (a web widget, not web + WhatsApp + Slack at once), one job (support Q&A, not support plus sales plus onboarding), and existing content (docs that already answer the questions, so week one isn't spent writing them). Widening scope later is cheap; starting wide is how six weeks becomes twelve.
Model choice barely moves the schedule — swapping between OpenAI and Anthropic models is a configuration change in a well-built pipeline, so you can defer that decision without cost. Pick per task later; see our OpenAI vs Claude comparison for how we choose.
From our own builds
We shipped a conversational AI business consultant and RAG systems that analyze 500–1,000-page legal contracts (case study). The pattern held both times: the model was never the bottleneck — data quality and integration approvals were.
When should you not build a custom chatbot?
If you handle under a few hundred support conversations a month, an off-the-shelf tool or a well-organized FAQ page usually beats a custom build on cost — revisit custom when volume, integration needs, or answer quality outgrow it. An honest agency should tell you this in the first call; it's one of the questions in our guide to vetting development agencies.