Every week a business owner tells me their AI chatbot failed. The reason is almost always the same — they built the wrong thing.
They wanted an agent. They built a chatbot. A chatbot is a scripted conversation flow with a thin AI layer on top. It can answer questions it has been trained to answer and fall apart gracefully on everything else. That is fine for an FAQ widget on a low-traffic website. It is useless if you need the system to take action, make decisions, and complete tasks end to end without human oversight.
The distinction matters more than most people realise. A chatbot responds. An agent operates. One answers "what are your opening hours" — the other books the appointment, updates the CRM, sends the confirmation, and flags the job for your team. Same conversation, completely different outcome. The technology required to build each one is not dramatically different. The thinking required to design each one is worlds apart.
Why most chatbot projects fail within 90 days
The failure mode I see most often is this: a business spends £2,000–£8,000 on a chatbot, launches it with high expectations, and watches it give wrong answers, confuse customers, and get quietly switched off within three months. This is not a technology problem. It is a design problem.
Most chatbot builders start with the interface. They think about the widget, the colours, the welcome message. They think about what the bot should say. What they do not think about is what the bot should do — and more importantly, what should happen when the conversation reaches a decision point that requires action. A well-designed AI agent starts from the opposite end. You map the workflows first. What are the five things a human does in this role every day? Which of those things are repeatable and predictable? Which require genuine judgement? The repeatable ones become agent tasks. The judgement calls become escalation triggers. The interface is the last thing you design, not the first.
What a properly built agent looks like
I built an agent for a travel agency that handled 200+ WhatsApp conversations daily. The brief was simple: replace the support team. The execution was not. The agent needed to understand travel preferences from unstructured messages, cross-reference availability via API, generate personalised itinerary recommendations, handle pricing objections, and confirm bookings — all within the WhatsApp thread, all in the client's brand voice.
That required four things that a standard chatbot cannot do. First, genuine language understanding — not keyword matching, but intent classification with context retention across a multi-message conversation. Second, real API integrations — the agent had to read and write data in external systems, not just generate text. Third, a decision framework — rules for when to act autonomously and when to pass to a human. Fourth, failure handling — what happens when an API times out or an edge case appears that the agent has not seen before.
The result was an agent that the agency's clients interacted with daily without realising it was not a human. Bookings closed at the same rate as with a human consultant. Support costs dropped 60%. The agency redeployed three staff into sales roles.
The stack that actually works in 2025
After building more than a dozen production agents, my recommended stack for a serious deployment is: Claude Sonnet or GPT-4 as the language model (Claude is significantly better at following complex instructions without hallucinating), n8n or Make as the workflow orchestration layer, Supabase or MongoDB as the data layer, and Twilio or the WhatsApp Business API as the communication channel.
This stack is not the cheapest to build. It is the most reliable in production. The orchestration layer is what most cheap chatbot builds skip — and it is the entire reason good agents do not break when something unexpected happens.
The question to ask before you build anything
Before you spec out a single feature, ask yourself one question: what is the one workflow in your business that costs you the most time or the most money if it goes wrong? Start there. Build an agent that owns that workflow end to end. Measure the outcome. Then expand.
The businesses that have the most impressive AI deployments did not start by building everything. They started by building the one thing that mattered most, getting it right, and letting the results justify the next investment.
