About the AI Feedback Assistant
Introduction
The nps.today AI Feedback Assistant provides the capability to move beyond simple data aggregation, enabling the rapid extraction of actionable, thematic insights and sentiment trends from large volumes of unstructured text data. This allows users to accelerate decision-making, prioritize high-impact improvements, and simplify the understanding of Customer Experience drivers.
Early Release
The AI Feedback Assistant is currently available as an early release until July 1st.
This means users can start using a preliminary version of the feature before the full release. The purpose of this early release period is to make the assistant available sooner, while we continue to improve the product, monitor performance, collect user feedback, and add more context to how the AI understands and summarizes customer feedback.
Until July 1st, the AI Feedback Assistant should be considered a preliminary product. Functionality, wording, performance, and the way results are presented may still change before general availability.
Responsibility
As the AI output is based on a constantly evolving advanced language model, it is designed to help analyze data faster and more effectively. We encourage users to treat the generated insights as critical supporting evidence in their decision-making process. While we strive for accuracy, the responsibility for verifying and acting upon the results remains with the user.
We recommend that users review the generated output before using it for business decisions, customer communication, reporting, or follow-up actions.
AI Model and Data Handling
Core AI Technology
The AI Feedback Assistant is powered by OpenAI, a large language model used to understand, summarize, and categorize written survey feedback.
The nps.today platform is hosted on Microsoft Azure infrastructure. According to the nps.today security documentation, data is stored in Microsoft Azure datacenters in the EU, in Amsterdam, Netherlands, with failover to Dublin, Ireland. Read more in nps.today Security & Requirements.
Account Data Separation
Every survey response in nps.today belongs to a specific customer account. The AI Feedback Assistant can only retrieve and summarize responses from your own account.
This means your feedback data is kept separate from other nps.today customers' data. This separation is built into how nps.today stores and retrieves data, not something the AI model has to decide or remember.
OpenAI Data Governance
When you use the AI Feedback Assistant, the relevant survey feedback is securely sent from nps.today to the OpenAI API so the assistant can summarize the feedback and answer your question.
The integration is configured to use OpenAI's European data residency options for supported API projects, models, and endpoints. OpenAI describes European data residency as supporting regional storage and regional processing for eligible API use cases.
OpenAI does not use the submitted API data to train its models.
You can read more here:
Personal Information and PII Handling
When generating a summary, the assistant is instructed to redact personally identifiable information or sensitive personal data that may appear in survey comments. This includes, but is not limited to:
- Names of individuals, including customers, employees, agents, and family members
- Personal addresses, including home addresses, work addresses, or any street-level address
- Phone numbers, including mobile numbers, landline numbers, and fax numbers
- Personal email addresses
- National identification numbers, such as Social Security numbers, CPR numbers, NHS numbers, NRIC numbers, or NIN numbers
- Passport numbers, driving licence numbers, or other government-issued ID numbers
- Banking, payment card, or financial account details, including IBANs, account numbers, card numbers, or CVV codes
- Date of birth
- Health, medical, or biometric information
- Login credentials, passwords, API keys, or tokens
- IP addresses and precise geolocation data
- Special-category data under GDPR Article 9, including race, religion, political views, sexual orientation, trade union membership, and similar sensitive information
If a survey response contains this type of information, the assistant is instructed to leave it out of the generated summary.
Important distinction
Survey responses are stored in your account as they were submitted.
This means that if a respondent writes a phone number, email address, or other personal detail in a free-text comment, the original response is saved in your account as typed.
The PII protection described here applies when the AI Feedback Assistant generates its summary. At that point, the assistant is instructed to filter personal details out of the output shown to the user.
AI-generated safeguards
The PII filter relies on the AI following strict instructions. This is a strong safeguard, but no AI system is perfect.
If you notice personal information appearing in an AI-generated summary, please contact our Support Team so we can review and tighten the rules.
PII Redaction Method
When the AI Feedback Assistant generates a summary, it is instructed to redact personal information instead of repeating it in the output.
If personal information appears in the source feedback, the assistant replaces each piece of PII with a category-specific placeholder. This allows the summary to preserve the meaning of the feedback without exposing the personal detail itself.
Examples of redaction placeholders include:
[name]for names of individuals[email]for personal email addresses[phone]for phone numbers[address]for personal or street-level addresses[national-id]for national identification numbers[government-id]for passport numbers, driving licence numbers, or other government-issued ID numbers[payment-info]for banking, payment card, or financial account details[date-of-birth]for dates of birth[health-info]for health, medical, or biometric information[credentials]for login credentials, passwords, API keys, or tokens[location]for precise geolocation data[ip-address]for IP addresses[sensitive-data]for special-category data under GDPR Article 9[redacted]when the information is sensitive but does not clearly fit one of the specific categories above
For example, a comment such as:
John Smith asked us to call him at +45 12 34 56 78 about his invoice.
Would be redacted in the AI-generated output as:
[name] asked us to call him at [phone] about his invoice.
This redaction method helps keep the feedback summary useful while reducing the risk of exposing personal or sensitive information in AI-generated output.
Company and Account Context
Each response record includes company or account context. The assistant can see which company or account the response belongs to in order to retrieve and summarize the correct feedback.
This company/account context is not personal information about an individual, but it is part of the information available to the assistant during analysis.
Assistant Scope
The AI Feedback Assistant only summarizes survey feedback available in your account. It is instructed to stay on topic and base its answers on the feedback data provided.
It should not answer unrelated questions, provide unsupported advice, or speculate beyond what respondents actually said in their survey comments.
AI Operational Limitations
Data Sampling Constraint
The feature has a hard limit on the amount of raw text data it can process in a single request, primarily driven by the LLM's token window size and the need for prompt processing efficiency.
The system strategically fetches a maximum of 2,000 responses for each NPS category: Promoter, Passive, and Detractor. This response selection is optimized based on user-defined topics and employees semantic search to ensure the most representative and valuable feedback samples are chosen for analysis.
Rate Limit
A system-wide API rate limit, provided by OpenAI, is enforced to ensure fair usage and service stability. This limit has a maximum capacity of token usage per minute. This total capacity is dynamically shared and allocated equally among all active users utilizing the AI Feedback Assistant.
This design ensures equitable resource distribution. It means that as the volume of concurrent users increases, the effective token allowance per minute available to each individual user will commensurately decrease.
If a user's request is temporarily denied due to the system reaching its peak rate, the user must wait for approximately one minute and try again. After the waiting period, the prompt can be safely retried.
This pause allows the system's token counter to reset and rebalance the available capacity across the user base.
Data input and output
Data Inputs
The AI uses the following structured data points from the feedback results to generate comprehensive insights:
- Rating
- Text comment
- Category
- Rating date
- Company/account context
No additional data points beyond those explicitly listed are utilized by the AI for analysis.
Output Format
AI-generated results are delivered exclusively in a text-based format. The feature does not generate or process images, files, or other media.
In Short
The AI Feedback Assistant is designed to help you identify patterns in customer feedback, such as sentiment, recurring themes, what is working well, and what needs improvement.
Your account data is kept separate from every other customer's data, conversations are not used to train AI models, and the assistant is instructed to filter personal information out of every generated summary.