NS&I: PolyAI
PolyAI delivers AI agents that respond to customer phone calls with lifelike voice interactions.
Tier 1 Information
1 - Name
PolyAI
2 - Description
What is PolyAI? PolyAI delivers lifelike voice assistants to handle complex customer service inquiries. It focuses on voice-first technology to deliver natural, human-like conversations that improve customer experience and reduce operational costs.
Why is it being used? NS&I uses PolyAI to automate and personalise customer interactions, enabling efficient, consistent, and scalable support through advanced natural language understanding and contextual memory.
3 - Website URL
4 - Contact email
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
National Savings and Investments (NS&I)
1.2 - Team
Enterprise Service Management
1.3 - Senior responsible owner
Head of Service Operations
1.4 - Third party involvement
Yes
1.4.1 - Third party
PolyAI & Sopra Steria
1.4.2 - Companies House Number
PolyAI: 11048129 Sopra Steria Limited: 04077975
1.4.3 - Third party role
PolyAI are the vendor of the tool and provided further support and consultancy. Sopra Steria are the service delivery partner who provide ongoing development, staffing, support and maintenance.
1.4.4 - Procurement procedure type
Specified and procured by our service delivery partner, Sopra Steria.
1.4.5 - Third party data access terms
N/A
Tier 2 - Description and Rationale
2.1 - Detailed description
How it Works:
The system uses natural language processing (NLP) to understand customer questions, identify intent, and provide relevant responses based on NS&I’s public product information. It combines rule-based flows for straightforward queries with machine learning to improve over time, and can escalate complex or sensitive issues to human agents.
Purpose:
To automate responses to common, non-account-specific customer queries, improving efficiency and customer satisfaction while freeing up human agents for more complex tasks.
Intended Users:
NS&I customers seeking information about products and services via web chat or phone, especially outside normal business hours.
Key Aspects and Functions:
- 24/7 availability, handling over 100,000 interactions per month
- Provides text responses and links to further information or self-service options
- Maintains conversational context for multi-turn interactions
- Escalates to human agents when needed
- Trained on NS&I’s public-facing product and service documentation
Scope:
Covers general product and service information, guidance on using NS&I’s services, and routing to further help if needed.
Limitations:
- Does not handle account-specific or highly specialized inquiries
- Cannot provide financial advice
- Customers must contact a human agent for detailed, personal, or error-related issues
2.2 - Benefits
Provides 24/7 instant answers to general product and service queries Improves customer convenience and satisfaction Automates routine inquiries Reduces contact centre workload and allows staff to focus on complex issues Delivers consistent, accurate information based on NS&I’s official resources Can handle over 100,000 interactions monthly without extra staffing costs Escalates complex queries to human agents
2.3 - Previous process
NS&I primarily relied on traditional staffed contact centres, and written correspondence for handling customer inquiries. Customers could call NS&I’s helpline to speak with agents during set hours, or write to them for support, including requests for extra assistance
2.4 - Alternatives considered
As an outsourced business running in an assurance role, NS&I specified business requirements and desired outcomes. PolyAI and other solutions were evaluated by our incumbent technology and operational outsourcer, Atos. NS&I validated that their chosen solution met the business requirements
3.1 - Integration into broader operational process
- Integrates with NS&I’s operations via the Genesys DX platform, connecting directly to the contact centre and digital channels for seamless customer interactions
- Informs routing decisions, escalating complex or urgent queries to human agents based on intent, sentiment, or topic analysis
- Supplies users (customers and agents) with accurate, up-to-date product and service information drawn from NS&I’s public documentation
- This information is used by customers for self-service and by agents to provide consistent, informed responses, improving efficiency and customer satisfaction.
- Analytics help NS&I monitor customer behaviour, optimise workflows, and enhance service quality in real time.
- Cannot handle account-specific queries, ensuring sensitive cases are directed to human support
3.2 - Human review
NS&I’s customer service and compliance teams regularly review PolyAI’s chatbot output. They monitor conversations, audit responses for accuracy and appropriateness, and use analytics dashboards. Quality assurance processes include sample audits, feedback loops, and regular updates to training data to ensure the chatbot meets regulatory and service standards.
3.3 - Frequency and scale of usage
Available 24/7 to all NS&I online customers seeking product and service information. Handles over 100,000 customer interactions / month.
3.4 - Required training
Contact centre agents receive training to intervene when needed and ensure smooth handover from the bot to human support.
3.5 - Appeals and review
Customers cannot formally appeal because PolyAI does not make binding decisions; it provides information and guidance on general queries. If a customer is unsatisfied with its response or needs further help, they can request to speak to a human agent or use other NS&I contact channels. For complaints or unresolved issues, customers can follow NS&I’s standard complaints process, which includes escalation to the Customer Care Team and, if necessary, referral to the Financial Ombudsman Service
Tier 2 - Tool Specification
4.1.1 - System architecture
PolyAI uses a cloud-based, modular architecture. Key components include: - Data Ingestion: Public-facing NS&I website data is ingested into the cloud for processing. - Natural Language Understanding (NLU): PolyAI’s conversational AI processes customer queries, performing intent recognition, entity extraction, and maintaining context using advanced NLP and machine learning models. - Dialogue Management: The system manages multi-turn conversations and generates user-friendly responses by summarising relevant information. - Routing: Straightforward queries are handled automatically, while complex or sensitive issues are escalated to human agents based on topic or sentiment analysis. PolyAI official architecture overview: https://poly.ai/conversational-ai/architecture/ PolyAI Help Center with technical guides and integration details: https://docs.poly.ai/home Infosys deep dive on PolyAI architecture (PDF): https://www.infosys.com/services/data-ai-topaz/insights/ai-imperatives-2024/poly-ai-architecture.pdf PolyAI’s approach to multilingual conversational systems: https://poly.ai/blog/towards-composing-multilingual-conversations/
4.1.2 - System-level input
- User queries in voice form.
- Contextual information from ongoing or previous conversations to maintain continuity.
- Structured or unstructured data from back-office systems (e.g., product details, FAQs).
- Dialogue data collected for ongoing system training and improvement.
4.1.3 - System-level output
- Natural language responses in synthesised speech, answering user queries or guiding next steps.
- Actions or transactions triggered within integrated systems (e.g., updating records, routing calls).
- Escalations or handoffs to human agents when queries exceed system capability or require sensitive intervention.
- Analytics and conversation logs for business review, system tuning, and compliance monitoring.
4.1.4 - Maintenance
NS&I requires Atos to maintain PolyAI at no more than n-1 behind the most current version. Retraining schedules are dynamic: the AI continuously learns from new customer interactions, with updates made as needed to address changes in product information. We test, review, and deploy updates across historical, draft, sandbox, pre-release, and live environments.
4.1.5 - Models
The architecture incorporates: Natural Language Understanding for intent recognition and entity extraction using neural networks LLMs to maintain context, manage multi-turn interactions, and manage and evaluate prompts
Tier 2 - Model Specification
4.2.1. - Model name
PolyAI
4.2.2 - Model version
Not publicly disclosed for security reasons
4.2.3 - Model task
Answering general product and service queries from NS&I customers by phone.
4.2.4 - Model input
Customer queries in voice form.
4.2.5 - Model output
Natural language responses with relevant product or service information; escalates to human agents for complex cases.
4.2.6 - Model architecture
PolyAI uses natural language processing and natural language understanding to interpret user input, extract intent, and manage dialogue. Dialogue management and machine learning models generate relevant, context-aware responses. It integrates with backend APIs for real-time data and escalates complex queries to human agents. Inputs are user queries via speech recognition. Outputs are natural language responses using synthesised speech. See https://poly.ai/conversational-ai/architecture/ https://poly.ai/technology/ https://poly.ai/conversational-ai/deployment/ https://poly.ai/the-ultimate-guide-to-generative-ai-platforms/
4.2.7 - Model performance
Following a phased rollout, 100% of inbound customer calls are now answered by PolyAI, with PolyAI also available 24/7 outside contact centre business hours. We report daily on performance against performance indicators for: call response; handling time; and call containment (whether calls are handled solely by PolyAI or require subsequent agent interaction). Currently, approximately 22.23% of calls are contained, being simple queries, e.g. enquiries about premium bond prize wins. Operational SMEs use additional journey analytics to drive continuous improvement and optimise containment, e.g. by adding further FAQs, with these improvements validated by user acceptance tests with observer validation.
4.2.8 - Datasets and their purposes
NS&I product information and terms and conditions Used for training, validation, testing, and production
Tier 2 - Data Specification
4.3.1 - Development data description
NS&I product information and terms and conditions https://nsandi.com
4.3.2 - Data modality
Text
4.3.3 - Data quantities
A knowledgebase containing circa 200 web pages and document links.
4.3.4 - Sensitive attributes
N/A - all information is public domain.
4.3.5 - Data completeness and representativeness
The data includes all NS&I’s public domain product and service information which is factual and correct. Questions not answerable by the chatbot from these sources may prompt a human agent intervention.
4.3.6 - Data cleaning
N/A
4.3.7 - Data collection
NS&I gathered together all of its NS&I’s public domain product and service information. As this data is another form of publishing the same information to the same audience, the purpose is identical.
4.3.8 - Data access and storage
NS&I and Atos staff have access to this dataset alongside NS&I customers and the general public due to the information being publicly available. As mandated by NS&I’s knowledge management and data retention policies, information assets are subject to regular review. Additionally, there are ad-hoc reviews of documentation as required when product and service terms change.
4.3.9 - Data sharing agreements
N/A
Tier 2 - Operational Data Specification
4.4.1 - Data sources
User inputs, API calls
4.4.2 - Sensitive attributes
N/A
4.4.3 - Data processing methods
N/A
4.4.4 - Data access and storage
Logs and analytics are stored. Logs can be accessed by designated technical and operational staff. Analytics are available internally for service improvement. Storage is subject to NS&I’s standard retention policies. No sensitive data is stored
4.4.5 - Data sharing agreements
N/A
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessments
A DPIA assessment was undertaken and confirmed that this tool contains and collects no Personally Identifiable Information.
5.2 - Risks and mitigations
The main risks are:
- data becoming out of date, and
- inaccurate outputs.
How are we mitigating against these risks: - User and customer testing to assure chats are delivered in accordance with the scripts and sources - As mandated by NS&I’s knowledge management and data retention policies, information assets are subject to regular review. Additionally, there are ad-hoc reviews as required when product and service terms change.