The question every SME manager or founder asks before starting an AI project: how long will it take before this is up and running? The short answer is that it depends on the type of system, the state of your data and how much control you want to have. The long answer is that a simple web chatbot can be live in 2 weeks, whereas an enterprise RAG system with complex integrations can take 6 months.
In this article I break down the real timelines for the most common AI projects in 2026: WhatsApp and web chatbots, phone agents, internal process automation and RAG systems. I’ll explain which phases each project goes through, which factors speed up or slow down deployment, and how to tell the difference between a supplier who promises two weeks of magic and one who gives you a realistic roadmap.
This page is for information only and does not constitute binding advice. Each case is adjusted after a diagnosis.
TL;DR
- An AI chatbot (web or WhatsApp) with a limited knowledge scope takes between 2 and 4 weeks from briefing to production.
- An AI phone agent needs between 3 and 6 weeks: conversational design, integration with the phone system, testing with real calls.
- Internal process automation with n8n and AI takes 4 to 12 weeks, depending on the number of systems to connect and the quality of the data.
- An enterprise RAG system (semantic search across internal documentation) takes 2 to 6 months: data cleaning, chunking, vectorisation, retrieval tuning, interface.
- Projects that overrun share three causes: unstructured data, changes of scope midway through, and integrations with legacy systems that have no API.
- You can speed things up by 30-40% by working in parallel and limiting the scope of the MVP, but the diagnosis phase and testing with real cases can’t be skipped without paying the price in rework.
Why a realistic timeline matters before you start
Most AI projects that drag on forever or end up in a drawer share the same original mistake: they were sold with a fantasy timeline. They tell you “chatbot in a week”, you start, and three weeks later you’re still waiting for the system to understand basic questions because nobody spent time structuring the knowledge base.
Having a realistic timeline isn’t just about managing expectations. It’s the difference between a project that goes into production, generates measurable value and gets iterated on, and one that eats up resources without getting anywhere. If you know that an AI phone agent takes 5 weeks to be properly tuned, you can plan the launch, train the team, design the escalation flow to humans and set aside budget for the first adjustments. If you’re sold 10 days, you end up launching something half-finished that frustrates customers and burns internal trust in the technology.
The timelines I’m giving you here are ranges observed in real SME and SaaS projects in 2026, working with commercial language model APIs (Claude, GPT-4, Gemini) and no-code or low-code tools such as n8n, Make or Voiceflow. If your project involves training models from scratch or building proprietary infrastructure, multiply by 3 or 5.
The 4 phases of any AI implementation
All AI projects, from the simplest chatbot to the most complex RAG system, go through these four phases. What changes is the length of each one.
Phase 1: Diagnosis and design (3 to 10 days)
This is where you define what the system is going to do, what data it will work with, which systems it will connect to and which use cases it has to solve. A well-done diagnosis includes:
- Mapping of current workflows (how enquiries, bookings and support are handled today).
- Inventory of data sources: where the information the system needs is held, in what format, and how up to date it is.
- Definition of integrations: CRM, ERP, product database, calendar, payment gateway, phone system.
- MVP scope: what the system has to do in version 1 to be useful, and what is left for later iterations.
- Success metrics: what percentage of enquiries it has to resolve without human intervention, what response time is acceptable, what error rate is tolerated.
In simple projects (an FAQ chatbot with 20 frequently asked questions), this phase can be wrapped up in 2-3 days. In complex projects (automation of internal processes that touch 5 departments), it can take 2 weeks of interviews, workshops and documentation.
Phase 2: Development and integration (1 to 8 weeks)
This is where the system is built. The duration depends on three variables:
- Complexity of the prompt or conversational logic: a chatbot with 10 clear intents is quicker than an agent that has to handle 40 variants of the same question, detect tone, personalise responses and remember context from previous conversations.
- Number of integrations: each external system the agent touches (CRM, database, booking API, payment webhook) adds between 2 and 5 days of development, testing and error handling.
- State of the data: if the knowledge base is in up-to-date, well-structured Google Docs, it’s vectorised in hours. If it’s in scanned PDFs, Word files from 2015 and scattered emails, preparation can take weeks.
A WhatsApp chatbot connected to a static knowledge base with no external integrations can be working in 1 week. A system that checks availability in your ERP, creates bookings, sends a confirmation by email and escalates to an operator if it detects frustration needs 3-4 weeks.
Phase 3: Testing and tuning (1 to 3 weeks)
This is where you validate that the system works in real cases, not just in the 5 happy-path examples from development. It includes:
- Regression testing: the system answers the 50 questions in the training set well, but what happens with the 200 variants that real users are going to throw at it?
- Stress testing: what happens if 20 enquiries arrive at the same time? Does it crash? Does it slow down? Does it lose context?
- Integration testing: does the data reach the CRM properly? Does the payment webhook fire correctly? Does escalation to humans work outside working hours?
- Prompt tuning: refining the system prompt, the tone, the escalation instructions and the few-shot examples.
In low-risk projects (a support chatbot that doesn’t touch sensitive data or transactions), this phase can be piloted in 1 week with a small group of users. In critical projects (a voice agent that handles payments or high-value bookings), you need 2-3 weeks of testing with real cases and iterative adjustments.
Phase 4: Deployment and training (3 to 7 days)
Launch into production, training of the internal team (if there is going to be supervision or escalation), setting up monitoring and metrics, and an iteration plan. If you’ve done the earlier phases properly, deployment is almost ceremonial. If you’ve skipped steps, this is where the problems appear.
Timelines by type of AI system
AI chatbot (web or WhatsApp): 2 to 4 weeks
This is the quickest project if the scope is limited. A chatbot that answers FAQs, qualifies leads or gives product information can be in production in 2 weeks if:
- The knowledge base exists and is structured (Google Docs, Notion, database).
- There are no complex integrations (at most, sending an email or logging to a spreadsheet).
- The tone and escalation rules are clear from the start.
If the chatbot has to check availability in real time, create records in a CRM, send confirmations by email and escalate to a human on WhatsApp outside office hours, the timeline rises to 3-4 weeks. You can find more information on the real capabilities of these systems in what a WhatsApp chatbot with AI can and can’t do.
AI phone agent: 3 to 6 weeks
Voice agents have an extra phase: conversational design and training of the synthetic voice. An AI phone agent that answers calls, qualifies interest, books appointments or gives product information needs:
- Weeks 1-2: design of the conversational flow (base script, handling of interruptions, escalation to a human, tone detection), voice selection, provider configuration (ElevenLabs, Deepgram, Vapi).
- Weeks 3-4: integration with the phone system (SIP trunk, phone number, routing), connection to the calendar or CRM if applicable, setting up recording and transcripts.
- Weeks 5-6: testing with real calls (internal first, then a pilot with real customers), latency tuning, refinement of interruption and goodbye prompts, team training.
If the agent has to transfer calls to human operators depending on calendar availability, take payments over the phone or query data in legacy systems, add another 1-2 weeks.
Internal process automation with AI: 4 to 12 weeks
The variability here is enormous because it depends on the number of processes to be automated and the quality of the data. A typical example: automating the qualification of leads that come in through a web form.
- Simple project (4-6 weeks): lead comes in through the form → webhook to n8n → call to the Claude API to analyse the message and assign a score → record in the CRM with tags → personalised welcome email.
- Complex project (8-12 weeks): lead comes in through the form → AI analysis → enrichment with external data (Apollo, Clearbit) → creation of a CRM record → assignment to a salesperson by sector and workload → sending of a personalised proposal as a PDF → automatic follow-up by email and WhatsApp → escalation to a human if there’s no response within 7 days.
The phase that stretches most here is integration with internal systems that have no API or whose API is poorly documented. If you have to build scrapers or intermediate APIs to extract data from an old ERP, you can add 3-4 weeks for that alone. More context in AI automation.
Enterprise RAG system: 2 to 6 months
RAG (Retrieval-Augmented Generation) is the technique that allows an AI agent to search your internal documentation before responding. It’s the system you need if you have hundreds of documents (manuals, contracts, policies, resolved cases) and want the agent to give accurate answers with references.
Timelines depend on the volume of documentation and its quality:
- Small project (2-3 months): 100-500 well-structured documents (Notion, Google Docs, Markdown), vectorisation with OpenAI or Voyage embeddings, vector database (Pinecone, Qdrant), search interface, retrieval tuning (top-k, reranking).
- Large project (4-6 months): 5,000-50,000 documents in mixed formats (PDF, Word, Excel, emails), custom cleaning and chunking, fine-tuned embeddings, on-premises vector database, interface integrated into an internal tool (Slack, Notion, custom dashboard), retrieval quality metrics, prompt iteration.
The critical phase here is data preparation. If your documents contain tables, scanned images or legacy formats, or mix languages, cleaning and chunking can take up 40-50% of the project’s total time. More on this in RAG.
Factors that speed up implementation
There are three levers you can use to shorten timelines without sacrificing quality.
Ready, structured data
If your knowledge base is already in a format that can be vectorised (Markdown, Google Docs, Notion, SQL database), you save between 1 and 4 weeks. If you have to clean scanned PDFs, extract tables from Excel or migrate information from an old system, add that time to the project.
Limited MVP scope
Launching with 70% of the functionality in half the time is better than waiting 3 months for a perfect 100%. A chatbot that resolves 60% of frequent enquiries already frees up your team’s time and gives you real data to iterate on. A voice agent that books appointments but doesn’t yet take payments can be ready in 4 weeks instead of 8.
The key is to define at the diagnosis stage what is critical for the MVP and what is left for version 2. A comparison of approaches in AI chatbot vs contact form.
Suppliers with stable APIs and clear documentation
Working with the Claude API, OpenAI API, ElevenLabs or Deepgram speeds up development because the integration is already solved, latency is predictable and support works. If you decide to build your own infrastructure or use open-source models without prior experience, multiply the timelines by 2 or 3.
Factors that slow down or block projects
Changes of scope midway through a project
Enemy number one. If in the third week you decide that the chatbot also has to take bookings, send reminders by SMS and handle cancellations, you’re adding 2-3 weeks to the timeline and blowing the budget. Changes of scope are inevitable, but they need to be managed as iterations, not as add-ons on the fly.
Integrations with legacy systems that have no API
If your CRM dates from 2010 and has no webhook or REST API, you’re going to need an intermediate solution: scraping, a mirror database, manual ETL or migration to a modern tool. Any of those options adds weeks. If you can, migrate before starting the AI project.
Lack of access to data or key stakeholders
If the product manager has no time to validate the prompts, if access to the database takes two weeks to be approved, if the documentation is in the head of someone who’s on sick leave, the project grinds to a halt. Implementing AI isn’t purely technical work; it needs business decisions and access to internal information.
How to tell whether a supplier is giving you a realistic timeline
When you ask for a quote, look out for these signs:
- They ask about your data before giving you a timeline: if they tell you “chatbot in 10 days” without knowing whether your knowledge base exists, run a mile. A serious supplier needs to see the state of your data before committing.
- They break down the phases: diagnosis X days, development Y weeks, testing Z days. If they only give you a round number with no breakdown, that’s a red flag.
- They propose a limited MVP: if everything you ask for fits into the first version with no cuts, they’re probably underestimating the complexity.
- They talk to you about post-launch iteration: the best AI systems are refined in production with real data. If the supplier doesn’t mention metrics, monitoring or an adjustment plan, they don’t understand how this works.
More selection criteria in how to choose an AI voice agent for your business.
When it’s worth extending the project
Quicker isn’t always better. There are three situations in which it’s worth investing more time:
When the system is going to handle sensitive data or transactions
If your AI agent is going to handle personal data, payments or medical information, you need an extra phase of security validation, compliance with the RGPD (the Spanish name for the GDPR) and adversarial prompt testing. It adds 2-3 weeks, but it saves you a fine or a data breach. Legal context in AI and data protection.
When the volume of interactions justifies fine-tuning
If you’re going to process 10,000 enquiries a month and every 5% improvement in resolution rate saves you thousands of euros in human support, it’s worth investing an extra 4-6 weeks in fine-tuning the model, optimising retrieval and tuning prompts with real data. The return justifies the investment.
When the system is the main interface of your product
If you’re a SaaS company and your AI agent is the way users get value (a data analysis assistant, a virtual tutor, a financial adviser), the quality of the system shapes how your product is perceived. There are no shortcuts here: invest the time needed in conversational design, latency, error handling and personalisation. Examples in solutions for SaaS.
Frequently asked questions
Can an AI chatbot be implemented in less than a month?
Yes. A web or WhatsApp chatbot with a limited knowledge scope and no complex integrations can be in production in 2 to 4 weeks: 3-5 days of diagnosis, 1 week of development, 3-5 days of testing and training. Projects that overrun tend to have integrations with CRMs, ERPs or old databases that require intermediate APIs.
How long does an AI phone agent take from briefing to launch?
Between 3 and 6 weeks. The conversational design and voice training phase usually takes 1-2 weeks, development and integration with the phone system another 2 weeks, and validation with real calls 1 more week. If there is transfer to human operators or integration with a CRM, add 1-2 weeks.
What delays an AI project the most?
Lack of access to structured data. If the information is in scanned PDFs, scattered Excel files or legacy systems with no API, the data preparation phase can triple the timeline. The second obstacle is a vague scope: changing requirements midway through a project adds weeks of redesign.
Is it quicker to implement AI with external APIs or with your own models?
With external APIs (OpenAI, Anthropic, ElevenLabs), development is between 3 and 5 times faster. A chatbot with the Claude API can be ready in 2 weeks; training your own model from scratch takes months. For most SMEs, APIs are the sensible option in 2026.
How long does the testing phase last before an AI system goes into production?
Between 1 and 3 weeks, depending on how critical it is. A support chatbot can be piloted with a small group of customers in 1 week. A voice agent that handles bookings or payments needs 2-3 weeks of testing with real cases, prompt tuning and validation of escalation to humans.
Can implementation be speeded up if the project is urgent?
Yes, but within limits. You can compress the timeline by between 30% and 40% by working in parallel (design and infrastructure development at the same time, live testing) and limiting the initial scope. What can’t be skipped is the diagnosis phase and validation with real cases; doing so generates rework that lengthens the project.
Next step
If you’re considering implementing an AI system and need a realistic timeline for your specific case, start with a free diagnosis. In 30-45 minutes of conversation you can find out whether your project is a 3-week one or a 3-month one, which factors will speed it up or slow it down, and what you need to have ready before you start.
If you already know you need to automate internal processes, take a look at the AI automation solutions and n8n. If your case is customer service by voice, see AI phone agent. Either way, the first step is to talk.