# AI Chatbot Development UK

Source: https://dijituldevelopments.co.uk/ai-chatbot-development/
Updated: 2026-10-10

> dijitul developments builds custom AI chatbots for UK businesses: assistants for your website, customer portal or staff intranet that answer from your own content, look up orders or bookings through your systems, and hand over to a person when they should. Every chatbot is a fixed-price quote after a free chat.

## Problems this solves

- Your team answers the same twenty questions every day by email and phone
- An off-the-shelf chatbot gives confident wrong answers about your products
- Customers want order or booking status out of hours
- Staff cannot find the current version of a policy or procedure
- The old decision-tree bot frustrates people and nobody maintains it
- You worry a chatbot will say something that embarrasses the business

## Key facts

- Answers are grounded in your own content using retrieval, with sources shown to the user
- Connected to live systems through tool calling, with read-only access by default
- Clear handover to a human by live chat, ticket or callback request
- Tested against a set of real customer questions before launch
- Protected against prompt injection and topic drift with input and output checks
- Conversation logs, retention settings and privacy notice wording designed for UK GDPR
- Works with OpenAI, Anthropic Claude, Google Gemini or open-weight models

## A chatbot that knows your business

Most chatbot disappointment comes from one of two things: a scripted bot that only understands exact phrases, or a general AI model that makes up answers because it has never seen your price list or returns policy. We build the third kind. The assistant uses a modern language model for understanding and writing, but it answers from **your** content, retrieved at the moment of the question.

That retrieval step is called retrieval-augmented generation. Your web pages, help articles, product data and PDFs are split into passages, converted into embeddings and stored in a vector index, often PostgreSQL with pgvector. When a question arrives, the closest passages are found and given to the model with an instruction to answer only from them and say so when it does not know. We show the sources under each answer so customers and staff can check. The RAG knowledge base page goes into more depth.

## Doing things, not just talking

The most useful chatbots look things up. Using tool calling, the assistant can ask our code to run a specific function, for example:

- `get_order_status(order_ref)` after the customer has verified their email or logged in
- `check_availability(date, party_size)` against your booking system
- `create_ticket(summary, category)` in your helpdesk when it needs a human

Our code, not the model, enforces who can see what. Tools are read-only unless there is a clear reason otherwise, and anything that changes data, such as cancelling a booking, asks the user to confirm and is written to an audit log.

## Knowing when to hand over

A good assistant knows its limits. We set explicit handover rules: complaints, anything involving health or legal matters, vulnerable customers, a repeated "that's not what I asked", or simply the user asking for a person. The conversation summary goes with the handover so your team does not ask the customer to start again. Out of hours, it becomes a ticket or callback request instead.

Internally, the same pattern works as a staff assistant for HR policies, product specifications or technical procedures, usually behind single sign-on with Microsoft Entra ID so answers respect who is asking.

## Keeping it safe and on topic

Public chatbots get tested by curious and sometimes hostile users. We plan for that:

- **Prompt injection**: users or uploaded content may try to override instructions. We separate instructions from user content, restrict tools and validate every output.
- **Topic limits**: the assistant politely declines questions outside your business rather than offering medical, legal or financial advice.
- **Personal data**: we minimise what is stored, set retention periods, mask card numbers and other identifiers in logs, and help you update your privacy notice.
- **Cost and abuse**: rate limits per session and IP, message length limits and monthly spend caps stop a bot attack becoming a large bill.

## Choosing the model and where it runs

Chatbots do not need the largest model available. For most customer questions, a fast, mid-sized model from OpenAI, Anthropic Claude or Google Gemini answers well when it has the right passages in front of it, and it responds quickly enough to feel like a conversation. We compare two or three candidates against your real questions and pick on accuracy, response time and cost per conversation.

Where data must stay in-house, for example an internal assistant over HR cases or client files, we can run an open-weight model on your own servers or private cloud instead. The chat widget, retrieval pipeline and admin area stay the same either way, because the model sits behind a provider-neutral layer. If you later move provider, we rerun the evaluation set and switch over without rebuilding the bot.

The widget itself is lightweight, accessible by keyboard and screen reader, and loads after the rest of your page so it does not slow your site down.

## Testing before launch, measuring after

Before launch we collect a set of real questions from your inbox, call notes and search logs, and write the correct answer for each. The chatbot is scored against this set on every change of prompt, content or model. After launch, the admin area shows unanswered questions, thumbs-down answers and handover reasons. Those gaps usually point to missing content, which your team can add without a developer.

Every build is quoted at a fixed price after a free chat, and you own the code and content index at the end.

## What we deliver

- A chat widget for your website or portal, or an assistant inside Microsoft Teams or your intranet
- A content pipeline that indexes your pages, PDFs and help articles and keeps them in sync
- Tool connections for order, booking or account lookups with proper authentication
- Human handover to live chat, your helpdesk or a callback form
- An admin area to review conversations, flag bad answers and add missing content
- An evaluation suite of real questions with expected answers, run on every change
- Guardrails, rate limiting, spend caps and abuse protection

Technologies: OpenAI API, Anthropic Claude API, Google Gemini API, PostgreSQL, pgvector, Laravel, TypeScript, Microsoft Entra ID

## FAQs

### Will the chatbot make things up?

Language models can produce plausible wrong answers, so dijitul developments designs against it: answers come from your retrieved content, the assistant is instructed to say when it does not know, sources are shown, and we score it against real questions before launch. No chatbot is perfect, which is why we also build handover to a person.

### Can the chatbot see customer orders or bookings?

Yes, through tool calling to your order, booking or CRM system. The customer verifies who they are first, and our code checks they can only see their own records. Tools are read-only by default. Anything that changes data asks for confirmation and is logged.

### Can it go in Microsoft Teams as well as our website?

Yes. The same assistant back end can serve a website widget, a logged-in portal and a Microsoft Teams app. Internal versions usually sit behind Microsoft Entra ID single sign-on so answers can respect the user's role and department.

### How do we keep the chatbot's knowledge up to date?

We build a content pipeline that re-indexes your website, help centre and document folders on a schedule or when content changes. Your team edits the source content as normal; the assistant picks it up automatically. The admin area shows which questions went unanswered so you know what to add.

### Is an AI chatbot compliant with UK GDPR?

It can be, with the right design. We minimise the personal data stored, set retention periods, mask sensitive identifiers in logs, choose a provider and region with suitable processor terms, and help you update your privacy notice. dijitul developments also documents the data flows for your DPIA.

### How much does a custom chatbot cost?

It depends on the content, the systems it connects to and where it runs. dijitul developments gives a fixed-price quote after a free chat and a short scoping stage. Ongoing model usage is billed per token by the provider, and we put spend caps in place.

## Pricing and contact

Every project gets a fixed-price quote after a free initial chat and a short scoping stage. You own the code and the data. Book a free chat: https://dijituldevelopments.co.uk/contact/ · 01623 650333 · info@dijitul.uk