--- title: "Bespoke AI Chatbot: Build or Buy for Support Teams" description: "When an off-the-shelf assistant is enough, when a bespoke AI chatbot on your own data pays back, what bespoke should include, and how to run one." canonical: https://shipshapedata.com/resources/ai-feature-integration/bespoke-ai-chatbot/ language: en-GB author: "Emile Van Zyl" author_role: "MarTech Automation Consultant" author_url: https://www.linkedin.com/in/emile-vanzyl-enhancing-your-human-advantage/ date_published: 2026-09-12 date_modified: 2026-09-12 --- # Bespoke AI chatbot: build or buy for support teams > When an off-the-shelf assistant is enough, when a bespoke AI chatbot on your own data pays back, what bespoke should include, and how to run one. Section: Home > Resources > AI & integration Canonical page: https://shipshapedata.com/resources/ai-feature-integration/bespoke-ai-chatbot/ By Emile Van Zyl, MarTech Automation Consultant (https://www.linkedin.com/in/emile-vanzyl-enhancing-your-human-advantage/) | published 2026-09-12 Your website assistant only replies to specific scripted questions and cannot address other enquiries. Your support team is presently handling a large volume of repetitive tickets. You are looking at a chatbot and must decide between a ready-made product for immediate use or a custom system built on your own data. This guide shows how to make that decision, outlines the requirements for a custom build, and details the work needed to maintain the system. A bespoke AI chatbot is a conversational assistant that uses your specific content, rules, and systems rather than relying solely on a general model. It delivers answers based on the documents you manage, follows your business logic for routing or escalation, uses your brand voice, and works within your existing user channels. This type of chatbot is not created from scratch. Instead, the components that determine the accuracy of its answers are adapted to your organisation, including the data, retrieval methods, escalation rules, and evaluation processes. Shipshape Data develops conversational platforms using governed data. Two of our published case studies describe chatbots built with this approach. This guide is for individuals who must choose between buying a product or building a custom solution. It is also for those who must justify that decision when the system produces its first incorrect response. ## What a bespoke AI chatbot is Most chatbots belong to one of three categories. Scripted bots follow a decision tree and only answer questions that have been written in advance. General assistants use a large language model to answer any question. These assistants rely on the model's training data instead of your specific content, which makes them fluent but often inaccurate regarding your business. A bespoke AI chatbot uses a language model to handle the conversation and your own content to provide facts. Before the bot answers, it finds relevant passages in your documentation, product data, or knowledge base and uses that information to write a response. Our guide to conversational AI explains the differences between intent-based bots and model-based assistants in more detail. Bespoke products vary in complexity. Some options only involve adding your documents and logo to a standard configuration. Other systems include custom retrieval layers, integrations with your ticketing and account platforms, and specific rules for transferring users to human staff. You can use the following tests to determine which level of customisation your project requires. ## Build or buy: the five tests A product is the right answer more often than suppliers of bespoke builds admit, and a bespoke build is the right answer more often than product vendors admit. Most situations can be resolved by running through five tests. - Where the answers come from. If all potential user questions are answered in a single help centre, a product that indexes that site is sufficient. If the answers require data from a document store, a product database, and an account system, you must build a system that connects to all three sources. - Consider the consequences of incorrect bot responses. If an error only results in a follow-up email, the product's accuracy is likely sufficient. If an error leads to a refund, a compliance violation, or a lost contract, you require specific control over grounding, refusal, and escalation. Most products do not provide these controls. - Determine if the bot must take action. Providing answers to questions is a standard feature. Tasks such as checking orders, creating tickets, or updating records require integration and specific rules regarding the bot's permissions. - Languages and channels. A single language on one website is sufficient for some products. If you use multiple languages across a website, an app, and a support desk, you will likely need to create a specific build. This is necessary because each language requires a separate evaluation. - Whether a product has already failed on your content. If a pilot programme generated incorrect answers that appeared plausible, it has tested the interface and revealed issues with the data. A custom build becomes cost-effective at that stage. If a product passes all five tests, buy it and spend the saved budget on the content it will answer from. If a build is needed on two or more, treat the project as a bespoke build and define the scope around an initial channel and a specific, measurable goal. ## What bespoke should include If building a custom solution is the correct choice, it must include the following elements. Any proposal that excludes these components is merely a demonstration with a logo attached. - The system provides grounded answers. Every response is based on retrieved passages from your content. The sources are visible to the user or to the agent reviewing the conversation. - Refusal. When the answer is not in the content, the bot states this and provides a way to contact a person. A bot that always provides an answer is guessing. - Routing and escalation are handled by business rules. These rules determine which conversations the bot will manage directly, which conversations it will transfer to a human agent, and what information the agent will receive when a transfer occurs. This information includes the conversation transcript, the sources used by the bot, and the reason for the handover. - Permissions. The bot only displays information that the user is authorised to view. An account holder and an anonymous visitor will receive different answers when they ask the same question. - Tone and languages. The bot responds using your specific register. It maintains the same level of accuracy across every language that you support. - Evaluation and monitoring. You should use a test set of real questions with known answers. Run this set before you launch and after every change. You should also have a person review a sample of live conversations. - Ownership. Your team can update the content, modify the rules, and review the metrics without contacting the supplier. > The interface is the smallest part of a chatbot. The accuracy of its responses depends on the data it can access and the way that information is organised. ## The data behind a chatbot that answers correctly The information needs to be current, and there must be a single correct version. Help articles that conflict with a newer policy document produce contradictory information, and the system cannot decide which one is correct. Before any build, the content is audited to determine what is present, where it resides, which version is the official source, and who is tasked with its maintenance. The content needs to be accessible. A retrieval layer converts documents into a format that the model can query by meaning. Links to live systems provide facts not present in documents, for example order statuses. Our guide to what RAG consulting involves outlines the retrieval process, and our search and retrieval service builds that layer. You need to govern your content. You ought to exclude or limit personal data, contract terms, and internal notes prior to the first conversation. This is required because a chatbot will repeat any information it can retrieve. Two of our clients demonstrate how this appears in production. At 1NCE, static developer documentation was turned into a multilingual assistant, achieving 60% of tickets handled to a human standard. At Slimstock, SlimGPT provides 24/7 support in 11 languages on the company website. Both case studies outline what the build involved. ## Running it: evaluation, escalation, languages A chatbot is a system that needs continuous maintenance. The quality of its responses after six months relies on the routine used to manage its operations. ### Evaluation Before launch, we test the bot with a set of real questions that have known answers. We assess three specific criteria: whether the bot retrieved the correct content, whether the answer was backed by that content, and whether the bot declined to answer when appropriate. We execute this same test set after each content change and model update. Once in production, a staff member reviews a weekly sample of conversations, and users can flag incorrect answers. Our guide to preventing AI hallucination details the technical methods for this process. ### Escalation We design the handover to a human agent rather than leaving it to chance. The agent gets the complete conversation history, the sources the bot cited, and the reason for the transfer. This guarantees the user does not need to repeat information. We monitor the handover rate as a specific metric. If the rate is too high, the bot is not delivering sufficient value. If the rate is too low, the bot is probably trying to answer questions beyond its scope. ### Languages We assess each language individually because retrieval quality and tone differ among them. A bot that gives accurate information in English but fluent yet incorrect information in German is less useful than a bot that only operates in English. ### Content ownership A designated individual in your organisation is accountable for the content the bot uses. The bot's index must be refreshed whenever that content changes. Outdated information is the main cause of bots that performed well at launch giving incorrect answers a year later. ## What it costs, as drivers We do not disclose our prices. Any fixed price given before we review your content is an estimate. The cost of a custom AI chatbot depends on these factors. - The content: its quantity, the number of formats, its freshness, and the degree of cleaning required. - The systems it must reach, from the help centre to the account system. Each integration requires technical work and a permission decision. - The actions it is allowed to take. Providing read-only answers is simpler than creating tickets or modifying records. - Languages and channels. Each language needs its own evaluation, and each channel needs its own integration. - The quality bar. A bot assessed on its resolution rate needs a larger evaluation set and more tuning than a bot assessed on its deflection rate. - Operating costs comprise model calls, hosting, and retrieval infrastructure. These monthly expenses rise according to the volume of conversations and the length of the answers given. - Who runs it. A team that owns both the content and the performance metrics keeps costs low. A bot without an assigned owner becomes more expensive each quarter. Ask any supplier, us included, to give the operating cost separately from the build cost based on your expected volume. ## Where this leaves you You should purchase a product if your content resides in a single location, the cost of incorrect answers is low, and the bot only needs to supply information. You should develop a custom AI chatbot if the answers need data from multiple systems, incorrect answers pose high risks, the bot must perform actions, or existing products have failed to manage your content. Regardless of your choice, the result quality depends on the underlying content. You are responsible for managing that content in both cases. Our conversational platforms service creates chatbots of this type on governed data, using your tone, within the channels you already use. Talk to us and we will tell you whether your content is ready to hold a conversation, and what to fix first if it is not. ## Frequently asked questions A bespoke AI chatbot is a conversational assistant built from your particular content, rules, and systems. It delivers answers derived from the documents you control, follows your established routing and escalation procedures, employs your chosen communication style, and connects with your existing platforms. The customisation concerns the way the data is organised, the method of information retrieval, the permissions granted to users, and the metrics used for performance evaluation, rather than the AI model itself. The cost depends on the volume and condition of the content, the count of integrated systems, and the actions the bot is capable of performing. Additional considerations are the needed languages and channels, the quality standards, and the monthly fees for hosting and model usage calculated from conversation volume. You ought to obtain separate quotes for the initial build and for ongoing operations. A precise estimate can be given only after the content has been audited. Product configuration requires a few days. Our conversational platforms service typically needs four to six weeks to advance a custom chat system from concept to production. The initial content audit confirms that this timeline is accurate. Projects involving multiple systems, several languages, or particular actions such as ticket creation need additional time. That extra time is allocated to integration and evaluation tasks rather than to interface development. Yes. This process is called retrieval-augmented generation. The bot locates passages in your documents that correspond to a question and uses them to compose an answer that includes citations. The documents need to be current, formatted for parsing, and subject to access permissions. Preparing this content constitutes the main task in the process. Base each response on the retrieved information. Instruct the model to refuse to answer when the information is not present. Include the sources used for each response. Test the system with a set of real questions before the initial release and after any updates. Ensure a human reviews a sample of real-time interactions. These steps do not remove all risks. They reduce the risk to a level that you can monitor and control. ## Frequently asked questions ### What is a bespoke AI chatbot? A bespoke AI chatbot is a conversational assistant built from your particular content, rules, and systems. It delivers answers derived from the documents you control, follows your established routing and escalation procedures, employs your chosen communication style, and connects with your existing platforms. The customisation concerns the way the data is organised, the method of information retrieval, the permissions granted to users, and the metrics used for performance evaluation, rather than the AI model itself. ### How much does a custom AI chatbot cost? The cost depends on the volume and condition of the content, the count of integrated systems, and the actions the bot is capable of performing. Additional considerations are the needed languages and channels, the quality standards, and the monthly fees for hosting and model usage calculated from conversation volume. You ought to obtain separate quotes for the initial build and for ongoing operations. A precise estimate can be given only after the content has been audited. ### How long does it take to build an AI chatbot? Product configuration requires a few days. Our conversational platforms service typically needs four to six weeks to advance a custom chat system from concept to production. The initial content audit confirms that this timeline is accurate. Projects involving multiple systems, several languages, or particular actions such as ticket creation need additional time. That extra time is allocated to integration and evaluation tasks rather than to interface development. ### Can a chatbot answer from our own documents? Yes. This process is called retrieval-augmented generation. The bot locates passages in your documents that correspond to a question and uses them to compose an answer that includes citations. The documents need to be current, formatted for parsing, and subject to access permissions. Preparing this content constitutes the main task in the process. ### How do you stop a chatbot making things up? Base each response on the retrieved information. Instruct the model to refuse to answer when the information is not present. Include the sources used for each response. Test the system with a set of real questions before the initial release and after any updates. Ensure a human reviews a sample of real-time interactions. These steps do not remove all risks. They reduce the risk to a level that you can monitor and control. --- Shipshape Data is a London AI consultancy: we build the data foundation your AI depends on, then the AI on top. Site guide for agents: https://shipshapedata.com/llms.txt | Contact: hello@shipshapedata.com