What does an AI chatbot development company in the UK build for you?
A custom AI chatbot is a chat window, on your website or WhatsApp, connected to a language model that answers visitors using your own information, plus the plumbing around it: lead capture, hand-off to staff, logging, privacy controls and a way to keep answers current.
The model is the part everyone talks about, and the easiest part to buy. What an AI chatbot development company in the UK should really be selling is everything else: preparing your content so the bot can find the right passage, instructions that stop it wandering off-topic, tests that prove it answers correctly, and a route to a person when it cannot. That is where most of the build time goes.
With us, another of us designs the retrieval and model side, one of us builds the widget, WhatsApp connection and integrations, and the third of us runs the content gathering and testing with your team. You end up with a bot that runs in your accounts, a document explaining how it works, and a test set you can rerun whenever content changes.
A generic widget sends the visitor’s question to a model that answers from its general training, perhaps with some pasted text about your business. A retrieval-based bot first searches your own content for the relevant passages, then asks the model to answer only from those. The second is far more accurate on questions about you.
Retrieval-augmented generation, usually shortened to RAG, works in three steps. Your content is split into sections and indexed so that meaning, not just keywords, can be matched. When a question arrives, the most relevant sections are retrieved. The model then writes an answer using those sections, ideally citing which page it came from. If nothing relevant is found, the bot says so.
The practical differences are large. A generic widget will happily describe services you do not offer, quote prices from nowhere or give general advice that contradicts your policy. A retrieval bot can still make mistakes, but they are traceable: you can see which passage it used and fix the passage or the instructions. For regulated or price-sensitive businesses that traceability is the whole point. The comparison table further down sets the options side by side.
What content should a custom AI chatbot be trained on?
Use the content you would be happy for a new member of staff to quote to a customer: your website pages, FAQs, service descriptions, price lists, delivery and returns terms, opening hours and policies. Leave out drafts, internal notes and anything you would not publish.
“Trained on” is a slightly misleading phrase. In most business chatbots the model itself is not retrained; your content is indexed and looked up at question time. That is good news, because updating the bot means updating a document, not paying for a new training run.
Quality matters more than quantity. Contradictions (two different delivery times on two pages) produce contradictory answers. Out-of-date PDFs produce out-of-date answers. So the first week of a chatbot project is often a content tidy-up: we list every source, flag conflicts and gaps, and ask you to decide which version is right. Questions customers ask that no page answers are added as short approved answers. It is also a useful audit of your website.
- Service and product pages, with current prices where public
- FAQs, delivery, returns and cancellation terms
- Opening hours, locations served and contact routes
- Policies customers ask about: warranties, bookings, complaints
- Short approved answers for common questions no page covers
How do you stop an AI chatbot making up answers?
You reduce made-up answers by limiting what the bot can draw on, telling it plainly what to do when information is missing, and testing it against questions designed to tempt it into guessing. No method removes the risk entirely, so the bot should also be allowed to say “I don’t know”.
Our guardrails work in layers. The instructions say to answer only from retrieved content and to offer a person otherwise. A relevance threshold means weakly matched passages are not used. Topics you rule out (medical diagnosis, legal advice, individual pricing, competitor comparisons) get fixed polite responses. Answers involving numbers such as prices or dates are checked against the source text. And every answer can show a link to the page it relied on.
Then we attack it. The test set includes normal questions, awkward phrasing, questions just outside your content, attempts to make the bot ignore its instructions, and questions where the honest answer is “we don’t offer that”. We share the results with you, fix the failures and rerun the set. After launch, flagged conversations feed new test cases.
Lead capture: turning chatbot conversations into enquiries
A chatbot earns its keep when a useful conversation ends in a captured enquiry: name, contact details and what the person needs, sent to wherever your team works. Asking for those details at the right moment matters more than asking early.
We usually let the bot answer first, then offer the next step when intent is clear: “Would you like a quote? I just need a few details.” Asking for an email address before answering anything tends to put people off. The fields captured are the ones your team actually uses, such as postcode for a trade business, preferred dates for a clinic, or order number for an online shop.
Captured leads go to email, a shared inbox, a spreadsheet, or directly into HubSpot or another CRM, with a short AI-written summary of the conversation so staff do not have to read the transcript. If you want those leads processed further, the AI automation guide for UK SMEs shows how they can feed quote drafting and follow-ups.
Handing off to a human: when the chatbot should step aside
A good chatbot knows its limits: it should pass the conversation to a person when it cannot answer, when the customer asks for a human, when the topic is sensitive, or when the conversation signals frustration or urgency.
How hand-off works depends on the channel and your staffing. On a website, the bot can collect contact details and promise a reply within your stated hours, or pass the live chat to an agent if you run a helpdesk with live chat. On WhatsApp, the same thread moves from bot to person, which customers find natural. Outside working hours, the bot says clearly when someone will respond.
We write the hand-off rules with you. Complaints, safeguarding concerns, anything about health, and requests involving personal account details usually go straight to a person. The hand-off message includes the conversation summary, so the customer does not repeat themselves. That single detail does more for customer satisfaction than most AI features.
Website chatbot or WhatsApp chatbot: which should you build first?
Build the website bot first if most enquiries start on your site from search; build the WhatsApp bot first if customers already message you there. Many UK businesses end up with both, sharing one knowledge base.
A website widget is quicker to launch and costs less to run, because there are no per-message platform fees. It catches visitors at the moment they are reading about you. WhatsApp suits repeat customers and conversations that continue over days: bookings, deliveries, service updates. It uses the official WhatsApp Business Platform, which requires a verified business account and a phone number dedicated to it.
Meta moved the platform to per-message pricing from 1 July 2025: businesses are charged when template messages are delivered, while replies inside the 24-hour window that opens when a customer messages you are free. See Meta’s WhatsApp pricing documentation for current rates. That structure favours bots that answer inbound questions rather than send lots of outbound messages.
UK GDPR transparency: telling visitors they are talking to AI
Visitors should know they are chatting with an automated system, what happens to what they type, and how to reach a person. Under UK GDPR you must give people clear information about how their personal data is used, and chat messages often contain personal data.
In practice that means a short line when the chat opens (“You are chatting with our AI assistant. A person can take over at any time.”), a link to your privacy notice from the widget, and a privacy notice updated to cover the chatbot: what data is collected, why, which AI provider processes it, where, and how long logs are kept.
If the widget sets cookies or similar storage that is not strictly necessary, the cookie banner should cover it, and the widget can load only after consent where needed. We build the disclosure, widget behaviour and the data-flow description; you and your adviser confirm the privacy notice wording. If your cookie set-up needs attention generally, the UK GDPR cookie banner page covers it. We do not give legal advice.
How long should chatbot conversation logs be kept?
Keep chat logs only as long as you have a reason to, and write that period down. The ICO’s storage-limitation guidance says you must not keep personal data longer than you need it, and suggests reviewing data at the end of a standard retention period and erasing or anonymising it unless there is a clear reason to keep it.
Chat logs are genuinely useful for a while: they show which questions the bot handles badly and which content is missing. That usefulness fades. Many businesses settle on a short period for full transcripts, then keep anonymised question text longer for improving the bot. Leads captured by the bot are a separate record that follows your normal customer-data rules.
We make retention automatic rather than a promise: transcripts are deleted or anonymised after the period you choose, the AI provider account is configured with the most restrictive data-retention option available to you, and access to logs is limited to named staff. The retention table below gives a starting structure to discuss with your adviser.
Do you need a DPIA before launching an AI chatbot?
Possibly. The ICO says a Data Protection Impact Assessment is required where processing is likely to result in high risk to individuals, and it lists innovative technology, including AI, as a factor when combined with others such as vulnerable users, special category data or large-scale processing.
A chatbot answering questions about opening hours on a café website is low risk. A chatbot on a clinic, care home, legal or financial services site, where visitors may type health details or personal circumstances, deserves a proper assessment, and many organisations choose to do a DPIA for any customer-facing AI anyway as good practice.
We provide what the assessment needs from the technical side: a data-flow diagram, the AI provider and its data-handling settings, where logs are stored and for how long, who can access them, and how users are informed. The ICO’s guidance on when a DPIA is needed is the place to start. The decision and sign-off are yours and your adviser’s.
How much does an AI chatbot cost to run each month?
Monthly running cost is made up of AI model usage, which scales with the number and length of conversations, plus hosting, the search index and, on WhatsApp, Meta’s per-message charges for template messages. For a typical SME website bot, the model usage is often modest; high-volume or WhatsApp-heavy bots cost more.
Model providers charge by the amount of text processed, both the question plus retrieved passages going in and the answer coming out. Retrieval keeps this efficient, because only relevant passages are sent. Choosing a smaller, cheaper model for straightforward questions and a stronger one only when needed can cut costs further without hurting quality.
Before you commit, we estimate running costs from your expected monthly chats and set spending limits on the AI account so a spike cannot surprise you. All of these costs are billed to your own accounts, not resold through us. Care after the free two months starts at US$120/mo if you want us to keep tuning. For a fuller breakdown see AI chatbot costs in the UK.
How much does AI chatbot development cost in the UK?
Quotes from an AI chatbot development company in the UK vary widely, depending on content volume, channels, integrations and testing depth. With us, a custom chatbot build starts at US$600; bots that need logged-in customer portals or bespoke admin screens move toward custom software pricing from US$900.
What raises the build cost: a large or messy knowledge base, several languages, WhatsApp as well as web, integrations with a helpdesk, CRM or booking system, access to customer-specific data (such as order status), and stricter review processes in regulated sectors. What lowers it: tidy, current content, a clear list of what the bot should and should not do, and one decision-maker for approving answers.
Compare quotes on what is included, not just the total: content preparation, the test set, hand-off, privacy work and post-launch tuning. A low price that leaves those out usually produces a bot that goes quiet after a few embarrassing answers.
AI chatbot development testing: checking answers before customers see them
Every chatbot should pass a written test set before launch: a list of real questions with the answer you expect, rerun after every content or instruction change. Without one, nobody can say whether the bot is getting better or worse.
We build the test set with you from three sources: questions your team actually receives (from email, calls and existing chat), questions the website ought to answer, and adversarial questions designed to trip the bot up. Each is marked pass, fail or partial, with notes. We aim for correct answers on the questions it should handle and a polite hand-off on the ones it should not.
A soft launch follows: the bot goes live on a few pages or to staff only for a few days, and every conversation is reviewed. Then it opens fully. The same test set is rerun monthly during the free aftercare period, which catches problems caused by content edits.
Keeping an AI chatbot accurate after launch
Accuracy decays when your business changes and the content the bot reads does not. The cure is a routine: update the source content, re-index it, rerun the tests and read a sample of conversations each month.
We connect the knowledge base to the source where possible, so when you edit a page or FAQ the bot’s index updates automatically. For PDFs and documents, an admin page or shared folder lets your team replace files without calling us. Price and date changes should be made once, at the source, not patched into the bot’s instructions.
Each month, someone should read a sample of transcripts, especially those where the bot handed over or the visitor left abruptly. Those conversations show missing content and unclear answers. During the two free months after launch we do that review with you and adjust; afterwards, your team can do it, or we can continue on a care plan.
Working with an AI chatbot development team in India from the UK
Chatbot projects run well remotely because the work happens in shared documents, a staging chat widget and test spreadsheets. There are no site visits and no UK office; you get video calls, a WhatsApp group and a weekly written update.
India is four and a half hours ahead of the UK in summer and five and a half in winter. Calls usually sit in your late morning or around lunchtime, and test results you request in the afternoon are often waiting when you start the next day. The team works in English and Hindi; if your bot needs Welsh or another language, you supply or approve the translated content.
Accounts are opened in your name: the AI provider, hosting and, for WhatsApp, the Meta business account. We are added as users. Quotes are in USD, invoices come from India, and you can pay from a GBP account by Wise, wire or PayPal, with nothing billed before written approval. The offshore team page explains the general working model.
Week one
Content inventory, list of allowed and banned topics, hand-off rules, and the first draft of the test set.
Week two
Retrieval index built, chatbot on a staging page, first test run shared with pass and fail notes.
Worked example: a hypothetical veterinary practice in Kent
This is a made-up scenario for illustration. Picture a three-surgery veterinary practice in Kent whose reception team answers the same questions all day: opening hours, vaccination prices, what to do if a pet eats chocolate, whether they see exotic pets, how to register.
The bot is built on the practice’s website content, price list and registration guide. Anything that sounds like an emergency triggers a fixed response with the practice phone number and out-of-hours arrangements; the bot never offers clinical advice. Registration enquiries are captured with owner name, pet type and preferred surgery, and emailed to reception with a summary. Visitors are told at the start that they are talking to an AI assistant.
Chat transcripts are kept for a short period to improve answers, then anonymised. A DPIA is completed by the practice with technical inputs from us, because owners may mention personal circumstances. A build like this would start from US$600, with WhatsApp added later if clients want it.
Checklist for choosing an AI chatbot development company in the UK
Before you sign with any AI chatbot development company in the UK, ask how answers are grounded, how accuracy is tested, what happens when the bot does not know, and whose accounts everything runs in. Vague answers to any of those are a warning.
Also ask to see the test set approach, the hand-off design and the retention settings. A supplier who promises the bot will “never make mistakes” is overselling; one who explains how mistakes are caught and fixed is being straight with you.
- Does the bot answer only from my approved content?
- What does it say when it does not know?
- How and when does a person take over?
- Is there a written test set I can rerun?
- Which AI provider processes the chats, and with what retention?
- Are the AI, hosting and WhatsApp accounts in my name?
- How do I update content without paying for a rebuild?