AI chatbot development
AI chatbots that answer from your content — and know when to hand over
We build LLM-powered assistants for customer support, sales and internal knowledge. They answer from your documentation, policies and product data rather than the open internet, cite where an answer came from, and pass the conversation to a person when they are unsure.
Why most chatbots frustrate the people using them
The old generation of chatbots followed scripted decision trees and broke the moment a customer phrased something unexpectedly. The new generation, built on large language models, has the opposite problem: it always has an answer, including when it should not. A support bot that confidently invents a refund policy is worse than no bot at all.
A useful chatbot is mostly retrieval and guardrails. It needs the right content, a way to find the relevant passage for each question, rules about what it may and may not say, a clean hand-off to a person, and a test set that tells you how often it is right before customers find out.
Where AI chatbots work well
AI customer service agent
Answering the questions your team answers every day — orders, accounts, how-to, policies — from your help centre and past tickets, with a link to the source article and an escalation path.
Pre-sales and lead qualification
Answering product and pricing questions on your site, asking the qualifying questions your sales team would ask, and booking a call or creating a CRM record when the lead is a fit.
Internal knowledge assistant
Staff ask about policies, SOPs, contracts or product specs in Slack, Teams or an internal tool and get an answer that cites the document it came from.
In-product help
An assistant inside your SaaS that understands your product's screens and the user's own account, and can walk them through a task or do it for them.
What separates a reliable chatbot from a liability
These are the design decisions we make on every chatbot project.
Grounded in your content (RAG)
Answers are generated from passages retrieved from your knowledge base — help articles, PDFs, tickets, product data — using hybrid search. Our RAG development service covers this layer in depth.
Permission to say "I don't know"
When retrieval finds nothing relevant, the bot says so and offers a person, instead of improvising. This one rule removes most embarrassing answers.
Hand-off with context
Escalation passes the transcript, detected intent and retrieved sources to your helpdesk or inbox, so the customer never has to repeat themselves.
Actions behind confirmation
If the bot can do things — look up an order, change a booking, create a ticket — it calls your APIs with the customer's own permissions and confirms before anything irreversible.
Measured accuracy
A test set of real questions with expected answers and sources, scored automatically on every change to prompts, content or model, plus LLM-as-judge rubrics checked against human ratings.
Review loop
Unanswered questions and thumbs-down responses are logged and reviewed, so gaps in your content get fixed instead of repeated.
What we build
Website chat widgets
An assistant on your marketing site or help centre, styled to your brand, with lead capture and hand-off.
In-app assistants
A chat panel inside your web or mobile product that understands the signed-in user's data and context.
Workplace assistants
Bots in Slack or Microsoft Teams that answer from your internal documentation with existing access permissions respected.
Helpdesk integrations
Draft replies for agents inside your support tool, or automatic first responses with agent review.
Chatbot rescue
An existing bot that gives wrong or vague answers: we build a test set, find where it fails, and fix retrieval, prompts or content.
How a chatbot project runs
- 1
Collect the real questions
We pull 100–200 genuine questions from tickets, chats or sales calls, with the correct answers. This becomes the test set and tells us what content the bot needs.
- 2
Prepare the content
We audit and ingest the help articles, documents and data the bot will answer from, and flag gaps where no good answer exists yet.
- 3
Build and measure
A working bot in a staging environment, scored against the test set. We iterate on retrieval and prompts until accuracy meets the bar we agreed.
- 4
Launch carefully
Internal testing first, then a share of real traffic with hand-off enabled, then full launch.
- 5
Improve from real conversations
We review failed and escalated conversations, update content and prompts, and re-run the test set to catch regressions.
Experience behind this service
Our founder has built LLM pipelines gated by LLM-as-judge evaluation and multi-tenant RAG in production on an AI EdTech platform used across 20+ US schools. The client work below shows the same discipline applied to structured AI output and to letting an assistant act inside a business system.
Connecting an operations platform to Claude with a remote MCP server
A remote Model Context Protocol server that lets the consultancy's staff run their admin platform by asking Claude — invoicing, scheduling, follow-ups, time entries and contracts — with the same permissions they have in the app.
Read the case studyAI application · Structured LLM outputAI-generated assessments and executive reports for a leadership consultancy
A team-assessment app where admins generate assessment templates from a topic with an LLM, participants answer through share links, and the system writes executive-summary reports.
Read the case studyIntegration and deployment considerations
Where conversations go
Transcripts can stay in your own database and analytics; escalations go to the helpdesk, inbox or CRM you already use.
Privacy
Personal data the model does not need is redacted before it is sent, and we use providers' zero-retention or enterprise endpoints where your obligations require it.
Keeping answers current
Content changes are re-indexed automatically, so an updated policy is reflected in answers without redeploying.
Cost per conversation
We measure tokens per conversation during testing and choose the smallest model that passes the test set, so running costs are predictable.
Technology we use
Models
- Anthropic Claude
- OpenAI GPT
- Google Gemini
- Groq-hosted open models
Retrieval & orchestration
- pgvector (PostgreSQL)
- Hybrid search
- LangChain
- LangGraph
- Vercel AI SDK
Evaluation
- promptfoo
- Langfuse
- LLM-as-judge with human calibration
Application
- React
- Next.js
- Node.js
- Python / FastAPI
What drives the cost
A chatbot answering from one well-organised help centre is a modest project. Cost grows with the number and messiness of content sources, the actions the bot can take, the channels it runs in and the accuracy bar. We quote after seeing a sample of your content and questions.
- Number and quality of content sources
- Whether the bot only answers or also takes actions
- Channels: website, in-app, Slack/Teams, helpdesk
- Hand-off and CRM/helpdesk integrations
- Accuracy bar and size of the test set
- Conversation volume and model cost targets
Who this is for
- Support teams answering the same questions every day.
- SaaS companies wanting in-product help that understands the user's account.
- Organisations whose staff struggle to find answers in internal documentation.
- Teams with an existing chatbot that customers have stopped trusting.
Further reading from our engineers
- How much does an AI chatbot cost? Build price, running cost and when not to build one
Off-the-shelf tool, configured platform or custom build — what each costs, what moves a custom quote, and worked token arithmetic for running a chatbot at current API prices.
- AI agent vs chatbot vs automation: which one does your business actually need?
Plain definitions, a side-by-side comparison and a decision table — plus why most good "AI agents" are mostly ordinary automation with AI in two or three steps.
Need an engineer inside your own team instead of a project? Hire AI developers from DablewDevs on an hourly or dedicated basis.
Related services
AI Chatbot Development: common questions
Tell us who the chatbot is for and what it should know
Tell us what you are building or what needs fixing. You will hear back from an engineer — not a sales rep — within one business day, usually with questions or a rough approach.
Your enquiry goes straight to our engineering team — the people who will scope and build your project, not a sales desk.