A custom AI chatbot takes 1 to 10 weeks to build. A website FAQ bot takes 1 to 2 weeks. A RAG chatbot that answers from your own documents takes 3 to 5 weeks. An assistant with sign-in, user roles, and connections to your CRM, HR system, or WhatsApp takes 6 to 10 weeks.
Those ranges come from the projects we scope and build, and they match the prices on our AI chatbot development page. But a range on its own doesn’t help you plan a launch date. So this guide breaks the timeline down the way an engineering team does: how many working days each part of the build takes, which decisions add weeks, and which delays have nothing to do with code.
Key numbers in 30 seconds
- Build time and calendar time are different. A 3-week build can take 6 weeks on the calendar if logins, documents, and approvals arrive late.
- Data is the biggest variable. Clean help-center pages take about a day to ingest. Scanned PDFs that need text recognition can take a week.
- Testing is a phase, not a step. Expect 3 to 4 working days to build a test set and fix weak answers on a typical RAG bot.
- Each extra system adds days. Each connection to a CRM, helpdesk, or order system usually adds 2 to 4 working days, plus whatever time your team takes to grant access.
- Some waiting is outside anyone’s control. Apple reviews about 90% of app submissions in under 48 hours. Meta can take up to 24 hours to approve each WhatsApp message template.
AI chatbot development timeline by type
The more a chatbot has to read, decide, and do, the longer it takes to build. Here is how the four common types compare.
| Chatbot type | Typical scope | Build time | Typical cost |
|---|---|---|---|
| Website FAQ bot | One source (help center or FAQ page), web widget, hand-off to email or live chat | 1 to 2 weeks | $1,500 to $3,000 |
| RAG knowledge-base bot | Many documents, answers with source links, analytics, basic sign-in | 3 to 5 weeks | $3,000 to $6,000 |
| Enterprise assistant | Sign-in, role-based access, CRM, HR or WhatsApp integrations, mobile app | 6 to 10 weeks | $6,000 to $12,000 |
| AI agent | Plans and completes multi-step tasks, writes to your systems with approval steps | Scoped after a discovery week | Scoped per project |
RAG stands for retrieval-augmented generation. The bot first finds the right passage in your files, then writes an answer based on it and links the source. Our guide on what a RAG chatbot is explains it in more depth. If you’re unsure whether you need a chatbot or something that takes actions, read AI chatbot vs AI agent first. Most teams should launch a chatbot, then add actions one at a time.
Where the time goes: a RAG chatbot, part by part
When we estimate a RAG chatbot, we don’t guess a number of weeks. We add up the parts. The table below shows the working days we plan for each part of a typical RAG knowledge-base bot.
| Part of the build | What happens | Working days |
|---|---|---|
| Discovery and test questions | Agree on the goal, the users, the sources, and 30 to 50 real questions with correct answers | 2 to 3 |
| Ingestion and parsing | Pull text out of web pages, PDFs, Word files or a database, and clean it | 1 to 4 |
| Chunking and indexing | Split text into passages, create embeddings, load them into a vector database | 1 to 2 |
| Retrieval tuning | Combine meaning-based and keyword search, add reranking, filter by user permissions | 2 to 3 |
| Prompts, guardrails and hand-off | Answer rules, citations, “I don’t know” behavior, hand-off to a person | 2 to 3 |
| Chat interface, sign-in and analytics | Web widget or in-app chat, user login, conversation logs and feedback buttons | 3 to 5 |
| Evaluation and fixes | Run the test questions, fix wrong answers, re-run after every change | 3 to 4 |
| Launch and monitoring | Deploy, set alerts, review the first real conversations | 1 |
| Total | 15 to 25 working days (3 to 5 weeks) |
Two parts of this table move the most: ingestion and evaluation. Both depend on your content, not on our code.
Ingestion: why the file format matters
Clean HTML from a help center can be indexed in about a day. PDFs are slower because tables, headers, footers, and two-column layouts get scrambled when the text is pulled out. Scanned documents are slowest. They need text recognition (OCR) first, and someone has to spot-check the output. A single misread number in a policy document becomes a wrong answer.
Retrieval tuning: the step cheap builds skip
Basic semantic search finds passages that mean something similar to the question. It often misses exact terms like product codes, policy numbers, or names. That’s why we combine it with keyword search (BM25) and a reranking step. Anthropic’s research on contextual retrieval found that adding context to chunks plus BM25 cut failed retrievals by 49%, and adding reranking cut them by 67%. Those 2 to 3 days are where a demo bot becomes a reliable one.
Evaluation: where “done” gets decided
Anthropic’s guide to building evaluations recommends testing with questions that look like real traffic, including edge cases, and favoring many test questions over a few perfect ones. In practice, the first run of a new bot usually exposes a handful of wrong or vague answers. Most come from gaps or conflicts in the source documents. Each fix means another run, so we budget 3 to 4 days for this loop.
7 factors that stretch the timeline
If your project has any of these, plan for the upper end of the range, or the next tier up.
| Factor | Typical extra time | Why |
|---|---|---|
| Scanned or messy documents | +2 to 5 working days | Text recognition, cleanup, removing duplicate or outdated versions |
| Each extra data source or system | +2 to 4 working days each | New connector, sync schedule, error handling, permissions |
| Role-based access | +3 to 5 working days | Every search must be filtered so users only see documents they’re allowed to see |
| WhatsApp channel | +3 to 5 working days, plus Meta approvals | Business verification, phone number setup, message templates reviewed by Meta |
| Native iOS or Android app | +1 to 2 weeks, plus store review | App screens, push notifications, App Store and Google Play review |
| More than one language | +2 to 4 working days per language | Separate test questions and answer checks for each language |
| Regulated data (health, finance, HR) | +1 to 2 weeks on the calendar | Security questionnaire, data processing agreement, audit logs, stricter testing |
Channels bring their own waiting time. Apple says its App Review team reviews at least 50% of submissions in under 24 hours and 90% in under 48 hours (Apple Developer). On WhatsApp, Meta says a message template decision can take up to 24 hours (Meta for Developers). Neither sounds like much, but a rejection restarts the clock, so we submit early.
Week-by-week: a 5-week RAG chatbot plan
Here is a realistic plan for an internal help bot at a 40-person company that answers from about 200 policy and process documents. Each week ends with something you can check.
| Week | What we build | What we need from you | Done when |
|---|---|---|---|
| 1. Data and design | Source review, conversation flows, hand-off rules, test questions | Documents, one content owner, 30 to 50 real questions | You sign off the question list and the answer rules |
| 2. Retrieval pipeline | Parsing, chunking, embeddings, vector database, hybrid search | Access to any shared drive, Notion or database | The right passage comes back for most test questions |
| 3. The chatbot | Prompts, citations, guardrails, chat screen, sign-in | Brand colors and where the bot will live | Your team can try it on a private link |
| 4. Testing | Full test run, fixes, tricky and off-topic inputs, permission checks | Feedback from 3 to 5 people within 2 days | Agreed accuracy on the test set, no permission leaks |
| 5. Launch and tune | Deploy to a first group, monitoring, content gap report | A go-live date and the first user group | Real conversations reviewed and gaps fixed |
A website FAQ bot compresses weeks 1 to 3 into about a week, because it has one clean source and no permissions. An enterprise assistant adds time for sign-in, roles, and each connected system, which is why it sits in the 6 to 10 week band.
3 example projects and how long each takes
These are typical scopes we see from startups and growing businesses. The timelines follow the parts and factors above.
Example 1: help-center bot for a SaaS product (about 2 weeks)
Scope: 80 help-center articles, a chat widget on the website and in the app, hand-off to the support inbox. Why it’s quick: the content is already clean HTML, there’s one source, and every user can see every article. Watch out for: articles that contradict each other after old feature changes. Fixing those is the customer’s job, and it’s usually the longest part.
Example 2: internal HR and policy assistant (4 to 5 weeks)
Scope: about 200 PDFs and Word files, Google sign-in, three roles (staff, managers, HR) that see different documents. Why it takes longer: PDF cleanup, role-based search filters, and a stricter test set, because a wrong answer about leave or pay causes real problems. Watch out for: documents with no clear owner. Someone has to decide which version of a policy is current.
Example 3: WhatsApp support bot that checks orders (6 to 8 weeks)
Scope: answers from the help center, checks order status in the store’s system, hands off to a human agent in the helpdesk. Why it takes longer: two system connections, WhatsApp setup and Meta approvals, plus testing of real order lookups. Watch out for: the order system’s API access. If it takes two weeks to get, the launch moves two weeks.
What delays chatbot projects (and how to spot it early)
When a project runs late, it’s rarely because “the AI is hard”. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, weak risk controls, rising costs, and unclear business value (Gartner, July 2024). The same issues that kill projects slow down the ones that survive.
Messy or conflicting documents
If your shared drive holds three versions of the refund policy, the bot may quote the wrong one. Someone on your side has to choose the current version. That decision often takes longer than the technical cleanup.
Waiting for access
Connecting to a helpdesk, CRM, or database needs API keys and a permissions setup from your admin. A week of waiting for one login is common, and it pushes every step after it.
Security review started too late
If the bot touches customer or employee data, your IT or legal team may ask for a security questionnaire and a data processing agreement. They should also ask how the bot handles prompt injection. OWASP lists it as the top risk for LLM apps and warns that instructions can arrive hidden inside external content like websites or files (OWASP). These checks are healthy, but if they start in week 4, they freeze the launch. Start them in week 1.
No agreed definition of “done”
Without a fixed test set, every review adds new questions and the finish line moves. Agree the questions and the pass mark in week 1, and add new ones to the next round, not the current one.
Run the 20-question test before you hire anyone
Write down the 20 questions your customers or staff ask most. For each one, find the single current document that answers it. Count the questions with no document, or with two documents that disagree. Each one is cleanup work that must happen before the build. More than four or five means you should plan extra time for content.
How to shorten your chatbot timeline
Most of the time you can save is on your side of the project.
- Launch one channel for one group first. Website or internal staff first. Add WhatsApp, the mobile app, and more teams in phase two.
- Name one owner who can decide within a day. Group email threads add days every week.
- Send documents and access before kickoff. Shared folder, API keys, admin contacts, and brand files.
- Write the test questions now. They become the finish line, and writing them often reveals content gaps.
- Start security and legal paperwork in week 1. Questionnaire, data processing agreement, and data storage questions.
- Work with a team that has shipped RAG before. We reuse a pipeline we’ve already run in production (upload, chunking, embeddings, vector search), so your budget goes into the parts unique to your business.
Red flags in a chatbot quote
A timeline in a proposal tells you a lot about how the team will build. Be careful if you see any of these.
- “Custom RAG chatbot in 2 days.” That’s usually a template with your files uploaded. Fine for a test, risky for customers.
- No test or evaluation phase. If the plan goes from “build” straight to “launch”, nobody has checked the answers.
- A fixed price before anyone has seen your data. Content quality is the biggest variable. A good team asks for samples first.
- Six months for an FAQ bot. Either the scope is unclear, or the team is learning on your budget.
- No plan for hand-off to a person. Every bot needs a way out when it doesn’t know.
For the full list of questions to ask, read how to choose an AI development company.
What to have ready before day one
- One clear goal, such as “answer the top 50 support questions without a human”.
- Current documents in one folder, with old versions and drafts removed.
- 20 to 50 real questions with correct answers.
- One named owner who can approve content and answers.
- Logins or API keys for every system the bot must read from or write to.
- Access rules: who is allowed to see which documents.
- Topics the bot must refuse, and when it should hand off to a person.
- Security and legal contacts, plus any forms they need.
- Where the bot will live: website, web app, iOS, Android or WhatsApp.
- A rough monthly usage guess, so running costs can be estimated before the build.
Running costs are separate from the build: you pay the AI model provider for what you use, plus hosting. See how much an AI chatbot costs for a full breakdown.
What building Abee and Sakkar AI taught us about timelines
We’ve shipped retrieval-based assistants in production apps, not just demos.
Abee Assistant is an enterprise AI assistant with a full RAG pipeline: document upload, chunking, embeddings, vector search, and grounded answers. The lesson from Abee: access rules belong in week 1. Deciding who can see which documents changes how data is stored and searched, so adding it late means rebuilding part of the pipeline.
Sakkar AI answers questions from published medical research instead of the model’s memory. The lesson from Sakkar: in high-stakes subjects, the test set is the project. Every answer has to trace back to a source, and “I don’t know” is a correct answer when the sources don’t cover a question.
Across both, the retrieval pipeline was the repeatable part. The time went into content, access rules, and testing answers. That’s why our first week is always about your data, not code.
About this guide
This guide was written by Huzaifa Imran, CTO at Emergen, who leads our AI chatbot builds. The working-day estimates come from how we scope RAG projects. Every outside figure links to its source. We update this page when tools, review times, or our process change.
Frequently asked questions
A website FAQ chatbot takes about 1 to 2 weeks. A RAG chatbot that answers from your own documents takes 3 to 5 weeks. An assistant with sign-in, user roles, and CRM, HR, or WhatsApp connections takes 6 to 10 weeks. Data readiness and access approvals affect the calendar more than the code does.
About 15 to 25 working days. That covers discovery and test questions, ingesting and indexing documents, retrieval tuning, prompts and guardrails, the chat interface, testing, and launch. Scanned PDFs, role-based access, or extra data sources push it toward the top of the range.
A no-code template can be set up in a few days. A custom chatbot that answers correctly from your own data, cites sources, respects permissions, and hands off to a person usually takes at least 1 to 2 weeks, mostly because its answers have to be tested.
WhatsApp needs a business account, phone number setup, and message templates that Meta reviews. Meta says a template decision can take up to 24 hours, and a rejection means changing and resubmitting it. The build itself usually adds 3 to 5 working days.
Messy or conflicting documents, waiting for system access, security or legal reviews that start late, and no agreed test set. Starting access requests and paperwork in week 1 and fixing the test questions early avoids most delays.
A chatbot is faster. It answers questions. An agent plans and completes tasks across your tools, which needs more permissions, approval steps, and testing. Many teams launch a chatbot first and add one action at a time.
One clear goal, current documents in one place, 20 to 50 real questions with correct answers, one person who can approve content, access rules, and logins or API keys for every system the bot must use.
Get a realistic timeline for your chatbot
Tell us what your chatbot should answer, where your data lives, and which channels you need. On a free 15-minute call, we’ll tell you which type fits, how many weeks it will take, and a cost range. Prefer email? Send us a message.
