AI chatbot development cost for a SaaS product in 2026 runs from about £4,000 for a scripted FAQ bot up to £80,000+ for an agentic assistant that calls multiple internal tools. Most SaaS teams land in the £18,000-£40,000 band for a retrieval-augmented generation (RAG) bot that actually answers questions from their own documentation, not a generic script.
| Chatbot Type | Typical Cost (GBP) | Timeline | Best For |
|---|---|---|---|
| Scripted FAQ bot | £4,000 - £8,000 | 2-3 weeks | Static help center questions |
| Workflow bot with API calls | £8,000 - £18,000 | 4-6 weeks | Order status, account lookups, ticket creation |
| RAG-powered assistant | £18,000 - £40,000 | 6-10 weeks | Product docs, onboarding, in-app support deflection |
| Agentic multi-tool assistant | £40,000 - £80,000+ | 10-16 weeks | Copilots that act across CRM, billing, and internal APIs |
| Ongoing LLM API usage | £200 - £2,000+/month | Recurring | Scales with conversation volume |
| Support & maintenance retainer | £500 - £2,500/month | Recurring | Monitoring, re-embedding docs, guardrail tuning |
Before anyone quotes you a number, figure out which tier your bot actually needs to be. Most founders ask for "an AI chatbot" when what they mean is one of four very different products with very different price tags.
A scripted FAQ bot answers a fixed set of questions using decision-tree logic or simple keyword matching. No large language model call is required. It's cheap because the scope is fixed: you write the questions, someone maps the answers, and it ships. It's also brittle. Ask it something outside the script and it either fails or loops back to "contact support."
A workflow bot adds API calls into the mix. It can look up an order status, pull a subscription tier from your billing system, or open a support ticket. This is where most SaaS teams should start if their real goal is reducing support ticket volume rather than sounding impressive in a demo.
A RAG-powered assistant retrieves relevant chunks from your own documentation, help center, or codebase and feeds them to an LLM to generate a grounded answer. This is what most people mean when they say "AI chatbot" in 2026 — it can answer novel questions in natural language instead of matching a script.
An agentic assistant goes further still. It plans multi-step tasks, calls several tools in sequence, and can take actions like updating a record or triggering a workflow, not just answering questions. Think of it as a copilot embedded in your product rather than a chat widget bolted onto it. This is the tier covered in more depth in our guide to the technologies shaping SaaS products in 2026.
Five variables move the price more than anything else, and none of them is "which AI model you pick."
LLM choice and API costs. A hosted model API (OpenAI, Anthropic, or similar) is cheap to start with and expensive at scale. A self-hosted open-weight model has higher upfront infrastructure cost but flatter per-conversation economics once volume grows. Most SaaS teams start hosted and revisit later.
Data sources for retrieval. If your documentation is scattered across a wiki, a help center, old PDFs, and support ticket history, someone has to clean, chunk, and embed all of it before the bot can retrieve anything useful. This data prep step is routinely underestimated and often costs more developer hours than the retrieval pipeline itself.
Number of integrations. Each system the bot needs to talk to, your CRM, your billing platform, your helpdesk, an internal database, is its own authentication flow, error handling path, and test surface. Three integrations cost meaningfully more than one, not linearly, because failure modes multiply.
Conversation design and testing effort also add up fast. Someone has to write the prompts, define what the bot should refuse to answer, and run enough test conversations to catch edge cases before real users hit them. For teams handling EU customer data, add guardrail and hallucination testing plus a review of where conversation logs are stored under GDPR, that review alone can add a week or two to a RAG build.
A RAG chatbot retrieves relevant passages from your live documentation using a vector database, then has an LLM generate an answer grounded in that text. An FAQ bot matches a fixed question to a pre-written answer with no live retrieval or generation involved.
That difference shows up in cost and in behavior. The FAQ bot never improves on its own; you edit the script when policy changes. The RAG bot updates as soon as you re-embed a changed document, but it needs embeddings, a retrieval pipeline, and ongoing monitoring for hallucinated answers that the FAQ bot simply doesn't require. Our tech stack selection guide covers how vector database choice (Pinecone, pgvector, Supabase) fits into a broader architecture decision.
Use a no-code platform like Chatbase or Intercom Fin when you need a generic support widget live within days and don't need it to touch your own database. Build custom when the bot has to act inside your product, own its data, or follow logic no template supports.
No-code tools are genuinely good at what they're built for: point them at a URL or a PDF, and you get a working widget in an afternoon. The trade-off shows up once you need something specific to your SaaS product. Custom fields in your billing system, a multi-step onboarding flow, or a requirement to keep all conversation data inside a chosen EU region tend to hit the ceiling of what these platforms expose through their settings panel.
This is the same build-vs-buy question we cover for marketing sites in our comparison of website redesign cost by feature and scope, the pattern repeats: templated tools win on speed, custom code wins on control once requirements get specific.
Timeline tracks cost tier fairly closely, but not perfectly, a workflow bot with five integrations can take longer than a simple RAG bot with one clean data source.
A scripted FAQ bot typically ships in 2-3 weeks: a short discovery call, script writing, and a QA pass. A workflow bot with two or three API integrations takes 4-6 weeks, most of it spent on integration testing rather than the conversational layer itself.
A RAG assistant runs 6-10 weeks in our experience, with data preparation and embedding pipeline work often taking longer than the chat interface. An agentic assistant that plans multi-step actions across several systems runs 10-16 weeks, closer to a full product build than a feature add.
At Axire Infotech, every chatbot project runs through the same four-step process we use for any SaaS build: discovery and planning to scope the actual use case, design and prototyping to map the conversation flow, development and testing under agile sprints, then launch and support once it's live. This structure is what keeps a RAG build from sliding past its estimated timeline, you can read more about how we run discovery in our piece on MVP development questions startups ask.
The build cost is the easy part to budget for. The ongoing cost is where teams get surprised three months after launch.
LLM API usage bills by tokens processed, not by seat or by month. A bot handling a few hundred conversations a day can cost a few hundred pounds monthly; one embedded in a high-traffic app can run into the thousands once every user opens it by default. Model choice, prompt length, and how often it retrieves large document chunks all move that number.
Documents change, so embeddings go stale. Someone needs to own re-embedding your knowledge base whenever product docs update, or your RAG bot starts confidently citing last year's pricing page. Budget a support and maintenance retainer, typically £500-£2,500 a month depending on conversation volume and how many integrations need monitoring, similar in structure to what we outline in our website maintenance cost breakdown.
Cut cost by scoping down to one use case first, reusing documentation and integrations you already have, and choosing a managed vector database over self-hosted infrastructure until volume justifies the switch.
The single biggest lever is scope. A bot that answers billing questions well beats a bot that half-answers everything. Prove the narrow version deflects tickets, then expand it into a second use case with a second budget cycle, rather than trying to fund the full agentic vision before you've shipped anything.
Reuse matters too. If your support team already maintains a clean help center, that's most of your RAG data prep done. If you already have a billing API your team calls elsewhere, reuse the existing auth flow instead of building a new one for the bot.
A £15,000 RAG bot that deflects 30% of tier-one tickets is a better first project than a £60,000 agentic assistant that sits half-used because nobody trusts it with real customer data yet.
People researching this question are often really asking whether their existing site platform matters for a chatbot build. It mostly doesn't, a chatbot's backend logic is separate from whatever renders your marketing pages.
Whether your marketing site runs on Webflow or custom React code, the chatbot itself needs its own backend: a database or vector store, an API layer, and hosting for the LLM calls. Custom development gives you full control over that backend and the data it stores. A no-code chatbot widget can usually be embedded on a Webflow site with a script tag, but it inherits all the same custom-vs-template trade-offs discussed above once you need it to do more than answer generic questions. If you're weighing this for your whole site, not just the chatbot, our agency comparison guide covers how teams evaluate custom code against platform-based builds more broadly.
Whichever tier fits your product today, get the scope defined before you ask for quotes. According to the UK government's data on generative AI adoption, adoption is accelerating fastest among SMBs that scope narrow pilots first rather than committing to enterprise-wide rollouts. The NCSC's guidelines on secure AI system development are also worth reading before you finalize a data handling plan with any vendor.
Axire Infotech builds RAG and workflow chatbots for SaaS teams across the UK, Ireland, and the Nordics, using the same React, Node.js, and Supabase stack we use for full product builds. If you want a scoped estimate for your specific use case rather than a generic range, get in touch to discuss your chatbot project and we'll walk through which tier actually fits your support volume and budget. You can also browse our recent projects or see our full range of development services before you decide.
Let's discuss your project and create something amazing together.
