"40% reduction in support tickets." "24/7 customer service without additional headcount." "Answers customer questions in under 3 seconds."
These outcomes are real. They're also only achievable with the right use case, the right knowledge base, and a properly scoped implementation. The gap between what chatbot vendors promise and what most deployments actually deliver comes down to one thing: how honestly you scoped the project before you built it.
IBM's research on AI in customer service found that chatbots can handle up to 80% of routine customer queries without human intervention. That same research confirms the ceiling is only reachable with quality knowledge base data and tight use case definition. Without those two things, you get a chatbot that frustrates customers and gets turned off within six months.
Before you commit to a build, you need a real cost-versus-savings model. Here's one.
What a Business-Grade AI Chatbot Actually Costs
There are three cost categories: build, run, and maintain.
Build: A RAG (Retrieval Augmented Generation) chatbot trained on your knowledge base, deployed on your website or internal tool, with an admin interface to update content—starts from £8,000 at the lower end of scope. That covers scoping, data preparation guidance, build, testing, deployment, and 60 days of post-launch support.
Run: Ongoing LLM API costs depend heavily on usage. For a business chatbot handling 500 queries per day at moderate token counts, expect £80–£250/month in API costs.
Maintain: Prompt performance degrades as your content evolves. A monthly maintenance retainer from £600/month covers monitoring, content updates, model upgrades, and performance optimization.
Year 1 total: approximately £8,000 build + £3,600 maintenance + £1,800 API = £13,400 fully loaded for a well-scoped SMB chatbot.
That's the honest number. Not the headline figure vendors advertise. The full cost.
What It Actually Saves
The savings calculation depends entirely on what the chatbot handles. The most common—and most reliable—use case is first-line customer support.
Here's how the math works with realistic numbers. Your support team currently handles 200 inbound enquiries per week. 60% of these are answerable from your knowledge base: product questions, pricing, process queries, FAQ-type requests. Each currently takes 8 minutes to respond to.
200 enquiries × 60% automatable = 120 queries per week × 8 minutes = 960 minutes = 16 hours of support time per week.
At a fully-loaded support cost of £20/hour: £320 per week, £16,640 per year saved from support team time alone.
Against a £13,400 Year 1 cost, that's a positive ROI in Year 1—with the maintenance cost dropping to recurring API and support fees from Year 2 onward. Year 2 economics are substantially better than Year 1.
And that's before accounting for 24/7 coverage, sub-3-second response times, and the improved customer experience that comes with instant answers.
Where Chatbot Projects Fail
Three failure modes kill the majority of AI chatbot implementations. None of them are about the technology.
Knowledge base quality. A chatbot can only answer questions as well as the information it's been trained on. If your internal knowledge is unstructured, outdated, or incomplete—scattered across Notion pages, email threads, and someone's head—the chatbot will hallucinate or deflect. Data preparation is the most underestimated phase of any AI chatbot project. We've seen builds delayed by four to six weeks because the knowledge base wasn't ready.
Use case misalignment. Chatbots handle structured queries from known knowledge exceptionally well. They struggle with complex negotiations, nuanced complaints, and situations requiring human empathy and judgment. Deploying a chatbot to handle the wrong type of query is worse than no chatbot—it damages customer trust and creates more escalations than it prevents. Be precise about what the chatbot will and won't handle.
No feedback loop. A chatbot without a mechanism to identify unanswered questions and continuously improve the knowledge base will degrade over time as your business evolves. Product changes, pricing updates, new processes—your chatbot needs to stay current or it becomes a liability.
Is It Right for Your Business Right Now?
The clearest signal that a chatbot is a good investment: your support team is spending 30%+ of their time answering the same 50 questions repeatedly, you have a knowledge base or FAQ that could serve as training data, and you have the operational infrastructure to maintain the system after launch.
If those conditions are already met, the ROI calculation above almost certainly works for your business.
If they're not met yet—if your knowledge base is fragmented, your support categories are unclear, or you don't have someone to own the system post-launch—the right first step is an AI Readiness Audit rather than a build. The audit tells you exactly what's needed to reach launch-ready conditions, and which AI investments will pay back fastest in your specific context.
Building a chatbot on a weak foundation produces a chatbot that gets turned off. Getting the foundation right first produces one that generates positive ROI in Year 1 and improves for years after.
If you're handling 100+ repetitive support queries a week and want to see a scoped cost model for your specific situation, explore custom AI tools for customer service →