Social Care Wales: Qualification Chatbot
The Qualifications Chatbot is used to support the public with an interest in social care qualifications find the appropriate qualification and information to work in the social care, early years, and childcare sector in Wales.
Tier 1 Information
###Name Qualifications chatbot
Description
How the tool is used:
The chatbot tool is used to support the public with an interest in social care qualifications find the appropriate qualification and information to work in the social care, early years, and childcare sector in Wales.
The chatbot uses an OpenAI integration in a Logicdialog software to search data made available on the qualifications finder to provide answers to user queries.
The chatbot is made available to users via a web application on the Qualifications Pages of the Social Care Wales website.
The chatbot uses data that is publicly available from Social Care Wales to provide responses.
Why the tool is used:
With the chatbot, customers can receive support and assistance anytime and anywhere without having to wait on hold for a phone operator or send an email and wait for a reply. This helps to provide an improved customer service experience that extends beyond the business hours and is in keeping with the 24-hour nature of the social care industry.
Website URL
https://socialcare.wales/qualifications-funding/qualification-framework
https://gofalcymdeithasol.cymru/cymwysterau-ac-ariannu/fframwaith-cymwysterau?
Contact email
Digitalteam@socialcare.wales
Tier 2 - Owner and Responsibility
###1.1 - Organisation or department Social Care Wales, Corporate Services
1.2 - Team
Digital Team
1.3 - Senior responsible owner
Assistant Director (Digital and Technology Delivery)
Corporate Services
1.4 - External supplier involvement
Yes
1.4.1 - External supplier
Logicdialog
1.4.2 - Companies House Number
10899737
1.4.3 - External supplier role
The supplier has developed chatbot software that is widely used by public sector organisations, like local authorities. The supplier has supported the restructure of the data and developed the prompts used by the chatbot.
This involved working with the Senior Digital Officer to provide:
* evidence and training on structuring the data to be used.
* developing and testing a minimum viable product through iterative feedback sessions
* presenting the minimum viable product (MVP)
1.4.4 - Procurement procedure type
Competitive quotes were sought to develop the chatbot from three companies. The chosen provider operates on a monthly subscription basis.
1.4.5 - Data access terms
The supplier does not have access to Social Care Wales data. The data supplied is publicly available.
Tier 2 - Description and Rationale
###2.1 - Detailed description The chatbot tool is populated with textual knowledge units, pre-categorized by topic. When a user submits a query, the system attempts to match the input with relevant knowledge units. These, along with the original query, are then processed by OpenAI to generate a response. The OpenAI privacy policy governing this process is accessible within the platform settings. In instances where the chatbot is unable to identify suitable knowledge units, a default response indicating a lack of understanding is provided. These occurrences are logged for administrative review and analysis. By examining these instances, the team can identify knowledge gaps and introduce new textual data to enhance the chatbot’s capabilities. Upon integration of this new knowledge, the chatbot’s response repertoire is expanded.
2.2 - Scope
The chatbot has been designed to support those with an interest in social care, early years and childcare. It is designed to:
- Find relevant qualifications to work in social care, childcare and early years in Wales.
- Provide high-level guidance on process of registering to work as a social care worker
The chatbot tool is not designed to: * Decide the suitability of an individual to work in the social care, early years or childcare sector * Decide the suitability of an individual to be include on the Register of social care workers * Provide a decision on a registration application * Provide progress on a registration or renewal application, or a Fitness to Practise case.
2.3 - Benefit
The primary benefits the tool is expected to deliver is: * Easier access to information for those seeking information about qualifications, currently the process is reliant on internal e-mail exchange * Access to a 24 - hours qualifications enquiry service – currently the service is complicated and if telephone assistance is needed, it’s limited access Monday to Friday. * Increased efficiency for the enquiries team to deal with complex queries
2.4 - Previous process
Prior to piloting the tool, the process was reliant on checking several pages of content on the website, contacting the enquiries line that operates 10am-3pm Monday to Friday and often a follow-up check via e-mail to the qualifications team.
The piloting of the tool is taking place at the same time as streamlining the information available on the website to improve customer journey.
2.5 - Alternatives considered
Use of the OpenAI Generative AI approach was chosen because an intent-based approach, which tries to decipher the users meaning when submitting questions, limits the responses to one answer. The data provided has some nuances and variables so supplying one answer that may or may not fit the customer’s question precisely. None of the data publicly available on the website is in a format that could be queried by an API, so the Generative AI approach fit the use case better in terms of getting better quality replies.
Tier 2 - Decision making Process
3.1 - Process integration
The tool presents a number of options related to a user query. It does not provide a final decision, but instead provides a body of information that an end user can use to decide on a course of follow up action.
3.2 - Provided information
The tool will in most cases provide a list of qualifications relevant to the user, who is the decision maker, or a list of suitable jobs that are relevant to a qualification.
The information used to present information is taken from a knowledge bank that lists all qualifications and job roles relevant to the Qualification Framework for Wales.
3.3 - Human decisions and review
Decision and review process for the tool is as follows: The team at Social Care Wales decides on the information to give to the tool as ‘knowledge/text data’. Once the knowledge is in use, review sessions are scheduled to analyse unanswered questions from users. The team then decide if the knowledge that the bot is using needs to be reviewed, deleted or new information needs to be supplied to the tool.
The team also decides on the tool’s role, personality and approach which designed to reflect the organisation’s brand guidelines for dealing with the public, A summary of the brand guidelines is provided to the tool as a prompt instruction for how it should respond to all queries.
3.4 - Required training
The project lead has been trained in: * Generative AI foundations * Generative AI prompts * Awareness in managing bias in AI, * Developing knowledge articles * Analytics
A user guide is also provided for the review team.
3.5 - Appeals and review
The appeals and review process for decisions made by the tool is the same as the organisation complaints procedure:
https://socialcare.wales/cms-assets/documents/Complaints-Policy-202295.docx
Tier 2 - Data Specification
###4.1 - Method The tool uses an unsupervised generative AI model that uses a data set provided by Social Care Wales about social care, early years and child care qualifications.
4.2 - Frequency and scale of usage
The tool is currently in Beta and has handled 75 queries in the first 5 days of use. This exceeds the expected number of users. Based on existing email data for qualifications it was predicted that the tool will have around 50 users per week
4.3 - Phase
The platform is in Beta (since 5 June 2024).
4.4 - Maintenance
The reviewing of any unanswered queries can be done as frequently as we wish. It is all accessible within the platform. In terms of the supplier Logicdialog, they have releases every 2 weeks, so if anything is flagged within the system that needs attention, we respond as quickly as possible. If the issue is critical, it’s likely we will provide a hotfix once a fix is developed if we’re mid sprint, otherwise it will be incorporated into the next available sprint.
4.5 - Model performance
The model has been tested against the following queries for performance and better analysis: * Does the chatbot have potential to achieve 80% or above for answer accuracy * Can the chatbot reduce the number of queries directed to qualifications team * The quality of Welsh language responses * If the responses are easy to understand
Number of queries that are directed to teams that are out of scope of the tool’s knowledge.
No biases have been identified in the data model.
4.6 - System architecture
4.7 - Source data name
Qualifications job roles
4.8 - Source data description
The main source of data is the job descriptions and relevant qualifications supplied by Social Care Wales and also publicly available as part of the Find a Qualification tool.
The variables used as part of the data set are: * Job role * Job purpose * Registration required * Qualification scenario(s) * Service requirements * Continuous professional development * Alternative qualifications * Other information (which relates to other pertinent information related to a specific job role, this might include regulatory obligations, service requirements)
4.9 - Source data URL
https://socialcare.wales/qualifications-funding/qualification-framework
4.10 - Data collection
The data used as part of the tool if from the Qualification Framework for social care early years and childcare. The Qualification Framework content is also the data source for the Find a qualification page on the Social Care Wales website. The data source is intended to inform users of the qualifications that care valid to work in the social care and early years sector in Wales.
The qualification framework and finder can be used to find: * the required or recommended qualifications in a service area * the required or recommended qualifications for a job role * what induction is required * what will be useful or required for updating, maintaining and progressing skills and knowledge.
The qualification framework and finder can also help: * raise staff knowledge, competence and confidence * make sure people hold appropriate qualifications during recruitment * set service standards during commissioning * standardise training and assessments that are being commissioned and provided * the development and agreement of an organisation’s policies * succession planning for crucial people in services * personal development and progression planning for staff * workforce planning and career advice * the checking of older qualifications and those from other UK nations * inspection and regulation of services by Care Inspectorate Wales (CIW) * quality assuring and monitoring by commissioners or responsible individuals.
For some job roles and service areas there are specific requirements for practice: The qualifications required by regulations, legislation and national minimum standards (NMS) in social care or early years childcare the qualifications workers and managers are required to have to register with us. Other services and roles where these requirements do not apply (non-regulated roles) have recommended qualifications. These will ensure people have appropriate and transferable qualifications, knowledge and skills.
4.11 - Data cleaning
The data used on the website under the Find a qualification page is unstructured and without a legible hierarchy that the Tool can understand. To remedy this the data was restructured to introduce a heading structure that offered a numbered list of qualification options for each job role. All other content remains the same.
4.12 - Data completeness and representative-ness
The data in use by the tool is periodically reviewed as qualifications are updated by the service designer following consultation from the qualifications team for social care and early years. Queries are reviewed on a fortnightly basis.
4.13 - Data sharing agreements
See the open AI privacy policy: https://openai.com/policies/privacy-policy
4.14 - Data access and storage
We have a cookie which is stored for a week. This allows the conversation to continue if the user comes back to use the bot within 7 days of using it the last time. You can turn the cookie off if you wish but it will mean every time the user opens a new tab/window etc, the bot will start from the beginning.
The conversation data is stored in our platform for 12 months. Personal data in the form of email addresses are stored for 12 months. A retention automation has been put in place to ensure the data is removed within the set timeline.
Tier 2 - Risks, Mitigations and Impact Assessments
###5.1 - Impact assessment impact assessments that have been completed as part of the pilot for the tool include: * Equality Impact Assessment * Welsh language impact assessment * Ethical data framework assessment * Digital accessibility audit report * Data Protection assessment
5.2 - Risks
The risks associated with the tool are: * end user understanding of the chat bot’s limits – as the pilot is focussed on a discreet area of our work there is a chance users will attempt to ask the chat bot questions beyond its scope of use. However, there are opportunities to mitigate this by including introductory guidance that appears when the tool is opened. We will also be able to use the queries and data gathered to improve on the responses provided and to make a case extend use of the chat bot to meet new user needs, should they arise.
-
accuracy of answers – the efficiency and accuracy of the answers is dependent on an effective regular review of the content by internal teams. Establishing an efficient and regular review and ongoing team working model to review questions from users to improve the content will be vital to the success of the pilot.
-
prompt injecting is mitigated by using Typesense to scrutinise a user’s query by performing a semantic search ensure it matches content in the database of qualifications, similar to a key word search, before using OpenAI. This means that if user asks a question that doesn’t match our database it would no response would be given. Additionally, the prompts used in the chatbot are structured to not to take instructions from users. These things combined have proven very strong so far.
-
personal data – If a user were to enter information that was not needed for the purposes of resolving their query we can mask the content. This means if the personal data follows a pattern, e.g. telephone number, post code, date of birth etc we can mask it, so it isn’t stored or sent to the team.