DBT: AI Chatbot
A support chatbot that provides AI powered responses to users based on their exporting or business support questions
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
The Deparment for Business and Trade
1.2 - Team
Business Growth Service Team
1.3 - Senior responsible owner
Chief Digital Officer
1.4 - Third party involvement
No
1.4.1 - Third party
N/A
1.4.2 - Companies House Number
N/A
1.4.3 - Third party role
N/A
1.4.4 - Procurement procedure type
N/A
1.4.5 - Third party data access terms
N/A
1. Summary
1 - Name
The Department for Business and Trade Webchat
2 - Description
An AI powered support chatbot that provides conversational support to businesses who are looking for support and guidance related to exporting or business support
3 - Website URL
www.business.gov.uk
4 - Contact email
ai.governance@businessandtrade.gov.uk
Tier 2 - Description and Rationale
2.1 - Detailed description
A conversational support agent that allows a user to input any question they have relating to exporting or business support in general.
2.2 - Benefits
More specific and personalised guidance than searching the website for answer
2.3 - Previous process
Users would have to manually navigate to the correct content or answers or utilise the site search functionality
2.4 - Alternatives considered
N/A
Tier 2 - Deployment Context
3.1 - Integration into broader operational process
The tool does not support or inform any specific decisions or processes and is limited to providing a conversational summary of already existing content available to businesses across business.gov.uk and gov.uk. Users can use the tool to access that content in a single place to be provided with the guidance that they require for their specific questions. The tool provides references and onwards links to those content sources to businesses using the tool.
3.2 - Human review
A sample of the tool’s outputs from the initial Private Beta have been manually reviewed by the Department’s monitoring and evaluation analysts.
3.3 - Frequency and scale of usage
The usage will be determined by users who engage with the tool upon creating an account on business.gov.uk and further engaging and the number of interactions and users will be tracked and monitored by the application
3.4 - Required training
There is no required training as the tool uses well established patterns of a conversational chatbot
3.5 - Appeals and review
N/A
Tier 2 - Tool Specification
4.1.1 - System architecture
Business.gov.uk webchat is built on the Microsoft Copilot Studio platform. This is a low code platform. There are two main aspects controlled by the development team.
-
Knowledge sources: The webchat tool has been provided with specific public urls as knowledges sources. Retrieval of knowledge relies on Microsoft Bing’s indexing.
-
Topics: The webchat tool allows the definition of topics that help maintain control over conversational flow. Topics can include conditions that may be triggered in different ways. The tools uses generative orchestration to decide which action to take based on the intent extracted from user input. This includes selection of the best knowledge source and topic as well as the generated answer. The platform provides the ability to configure instructions at the agent level and for individual topics. It also provides a selection across a number of LLM providers, including OpenAI and Anthropic. At the time of release, the chosen model is GPT-5. Access to business.gov.uk chat is restricted to those with an account on the website.
The AI Agent has an API endpoint and communication with the endpoint is with a token generated by a secret. The Agent is deployed onto the Website via JavaScript which facilitates communication with the vendor API. The low code platform is able to make POST API requests to specific API endpoints in business.gov.uk to facilitate site actions such as user lockout/ban.
4.1.2 - System-level input
Natural language input by logged in users of business.gov.uk
4.1.3 - System-level output
Natural language output generated by the selected LLM, grounded with content from specified public urls across gov.uk and business.gov.uk. In some conversational flows, natural language output written by humans is returned.
4.1.4 - Maintenance
The model used is not trained or further developed in any way. The team will monitor for any changes to the model by the foundation model provider and the Copilot studio platform.
4.1.5 - Models
The webchat tool uses GPT-5, provided through the Copilot studio platform
Tier 2 - Model Specification
4.2.1. - Model name
At the time of release this is GPT-5
4.2.2 - Model version
GPT-5
4.2.3 - Model task
Copilot studio enables development of agents that can perform various tasks (with or without the use of a model). For business.gov webchat, the agent is able to use the specified model in order to decide a plan of action in response to a user input. This includes which content from the allowed knowledges sources it should use, understanding the user queries, deciding if a particular topic is triggered, generating the response and ensuring the response meets certain standards.
4.2.4 - Model input
User input, system instructions (from the team), system instructions defined by Microsoft
4.2.5 - Model output
Response to user input, content extracted from knowledge sources, plan generated by generative orchestration
4.2.6 - Model architecture
https://openai.com/index/introducing-gpt-5/
4.2.7 - Model performance
Webchat was made available to a small group of users (approx 5000) for 8 week private beta in Autumn 2025. Prior to this prior beta, manual testing was conducted by the team, including employing subject matter experts to review a range of questions across the exporting domain and their generated answers. This involved reviewing answers across a range of metrics, including relevance, how often source content was cited and quality (on a qualitative basis). In addition, the agent was tested against a range of malicious questions, to ensure unintended behaviour was captured in as many cases as possible. Following the private beta, transcripts were reviewed using a mix of automated and manual evaluation. Automated evaluation was used to understand how source content was being used in responses, through metrics such as word overlap and semantic similarity. It was also used to understand if reference links provided in responses were valid. A comprehensive manual evaluation was also performed, reviewing accuracy of claims made against source material as well as relevance. A quality score was also assigned based on expert review. Overall, the approaches taken for testing and learning have provided the team with a broad understanding of risks and failure modes.
4.2.8 - Datasets and their purposes
No datasets have been used to develop the model or tune it to improve performance. Collections of example questions have been developed in order to aid testing and evaluation.
2.4.3. Development Data
4.3.1 - Development data description
4.3.2 - Data modality
4.3.3 - Data quantities
4.3.4 - Sensitive attributes
4.3.5 - Data completeness and representativeness
4.3.6 - Data cleaning
4.3.7 - Data collection
4.3.8 - Data access and storage
4.3.9 - Data sharing agreements
Tier 2 - Operational Data Specification
4.4.1 - Data sources
The webchat tool receives a user question via the webchat fronted which is then passed to the copilot agent. The copilot agent uses the Bing index to search for relevant content across the specified knowledge source urls
4.4.2 - Sensitive attributes
The webchat tool checks every input message for personal data by using regular expressions for common personal data types, such as email addresses and phone numbers. Where detected, this part of the user input is redacted before onward processing by the copilot agent.
4.4.3 - Data processing methods
Any structured personal data detected in the user input (using regular expression matching) is redacted before being sent to the copilot agent.
4.4.4 - Data access and storage
Conversation transcripts are stored in DBT’s Microsoft Dataverse tenant. From here, transcripts are ingested into DBT’s Data workspace (anaytical platform) via a daily pipeline. Transcripts are retained for 90 days. This data can be accessed by members of the AI webchat team who are responsible for monitoring and evaluating transcripts.
4.4.5 - Data sharing agreements
There are no data sharing agreements in place for this tool.
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessments
DPIA - completed September 2025 prior to private beta and repeated March 2026 as a routine reassessment prior to a rollout to a larger user base
5.2 - Risks and mitigations
A number of risks were identified and actions taken to mitigate them.
- Privacy and personal data: Users may intentionally or inadvertantly enter personal data into the webchat tool. To mitigate this, when a user accesses webchat they see a clear instruction that they should not input personal or commercially sensitive information. They are also given a link to the website’s privacy notice. In addition, cleansing of personal data is applied in the frontend prior to messages being sent to the Copilot agent.
- Risk of answering irrelevant questions: The underlying technology for current generative AI (LLMs) is general purpose and hence can be used for many different use cases. This presents a risk of the webchat tool responding to irrelevant questions, or questions unrelated to the intended domain. This has been mitigated by development of a range of topics to reflect different categories of questions which may be posed by users. In the case of irrelevant questions (malicious or not) the copilot agent can trigger such topics thereby enabling specific, pre defined actions or responses to such questions.
- Risk of innaccurate responses: this risk is present for any AI tool. For webchat, the agent has been configured to only use content from knowledge sources provided to it. Furthermore, the agent has been configured to provide reference links to the source content it uses and responses are restricted if they include references not part of the intended domain (business.gov.uk)