Agent Analytics helps you understand how your visitors interact with AI-powered agents in your application by capturing and reporting on visitor-submitted prompts and conversations.
This article outlines what data is collected, how it’s stored, and the privacy and security measures in place for Agent Analytics.
Prerequisites
The following prerequisites apply to Agent Analytics:
- You must be on agent version 2.300.0 or above. Data isn’t collected on earlier versions.
- You must be a subscription admin or have the AI Agent Admin role to view the actual prompt or conversation text that visitors submit to AI agents. Other users can still access analytics about prompts or conversations, such as counts and trends, without viewing the text contents.
What data is collected
Agent Analytics captures visitor-submitted prompts or full conversations from AI agent interactions in your application. Full conversation capture requires implementing the Conversations API.
Agent Analytics doesn’t capture or store the contents of files or attachments submitted during AI agent interactions. It captures metadata indicating whether a file was uploaded, including the file name and type.
Agent Analytics also respects your privacy and exclusion settings. Visitors and accounts marked as Do Not Process don't generate Agent Analytics data, and prompts from visitors and accounts on your exclude list aren’t included in your prompt counts and don’t count toward your monthly prompt limit.
Where data is stored
All data collected through Agent Analytics is stored using the same infrastructure and security practices as other Pendo product data.
Pendo uses Google Cloud Platform (GCP) to host customer data in a secure, multi-tenant environment. GCP services meet industry standards for security and compliance, including SOC 2 Type II and ISO 27001.
By default, Agent Analytics data is retained for 18 months from the date of collection. After this period, data is automatically deleted in accordance with Pendo’s data retention policies.
Customer data collected through Agent Analytics is never commingled with data from other subscriptions.
For details about data storage, encryption, and compliance certifications, see Data collection and compliance and Security and privacy at Pendo.
Data visibility by role
In addition to infrastructure-level protections, Pendo restricts access to raw conversation data based on user roles. Only subscription admins and users assigned the AI Agent Admin role can view visitor-submitted
prompt text, conversation transcripts, and visitor details. Other users in your subscription can access Agent Analytics reports, including aggregate metrics, use case summaries, and trend data, without seeing
individual prompt or conversation content.
For a full breakdown of what each role can access, see Who can access conversation data in the Agent Analytics overview.
Preventing sensitive data capture
Pendo applies multiple layers of protection to help prevent the capture and processing of sensitive or Personally Identifiable Information (PII). For each AI agent, you choose which categories of sensitive data Pendo removes from visitor prompts and agent responses before that content is stored.
How redaction works
For Pendo Web SDK version 2.330.0 and later, and for all supported Agent Analytics SDKs (including the Python and TypeScript SDKs), Agent Analytics redacts data during ingestion, before any captured content is stored. Pendo stores the redacted content rather than the original text. This applies to text in both visitor-submitted prompts (inputs) and agent responses (outputs).
Redaction uses a self-hosted machine learning model that identifies sensitive data based on context, together with regular expressions that match common formatting patterns. Combining the two lets Agent Analytics detect data types that don't follow a predictable format, such as names and physical addresses.
The redaction model runs on Pendo infrastructure. Redaction is applied before content is stored and before it's used by other Agent Analytics features, including use case grouping, issue detection, and summaries.
Pendo redacts the categories that are selected. If a category isn't selected, Pendo doesn't detect or remove that type of data, so it's stored as entered and sent to the third-party AI providers that power features such as use case grouping, issue detection, and summaries. For the list of providers, see Artificial intelligence (AI) at Pendo.
Applications running a Pendo Web SDK version earlier than 2.330.0 use the previous redaction method, which relies on regular expressions alone, covers a fixed set of data types, and can't be customized.
Where redacted content appears in Agent Analytics, including the Conversations tab, the prompts table, and use case detail panels, Pendo replaces the original text with a labeled placeholder. The label identifies the specific type of data that was detected, which can be more specific than the category name. For example, selecting Physical address can produce separate city and state placeholders within the same prompt.
What you can redact
You can select any combination of the following categories. All 14 categories are selected by default and available whether your agent captures prompts only or full conversations.
| Category | What it covers |
| Names | First and last names. |
| User names | Usernames and account handles. |
| Email addresses. | |
| Phone number | Phone and fax numbers. |
| Physical address | Street addresses, cities, states, postal codes, and countries. |
| Dates | All dates, including dates of birth and employment dates. |
| SSN | US Social Security numbers. |
| National ID | National ID numbers, such as passports and National Insurance numbers. |
| Tax ID | Tax IDs, such as EIN and TIN. |
| Credit/debit card | Card numbers and security codes. |
| Bank account info | Account numbers, routing numbers, and SWIFT or BIC codes. |
| Invoices | Invoice, payment, and reimbursement reference numbers. |
| Currency amounts | Salaries, wages, bonuses, commissions, and payment amounts. Detection is based on context, so amounts are detected with or without a currency symbol. |
| IP addresses | IPv4 and IPv6 addresses. |
Where you configure redaction
You select redaction categories when you add an AI agent, and you can change them at any time in Custom redaction in the agent's settings. Redaction settings apply to one agent at a time, so agents in the same subscription can use different categories.
To view or change redaction settings, you need subscription admin, app admin, or AI Agent Admin permissions for the agent's application. For more information about what each role can access, see Who can access conversation data.
For steps, see Set redaction categories for an agent in Add and configure AI agents in Pendo.
Before you save your selections, you can enter up to 4,000 characters of sample text in View redacted preview to see how your selected categories apply to it.
When redaction changes take effect
Changes to redaction categories apply to prompts and conversations after the change. Content that Pendo already collected keeps the redaction that was in place when it was captured. Clearing a category doesn't recover data that was already redacted, and selecting a new category doesn't redact data that was already collected.
Changes take effect within a few seconds. Because redaction is applied when data is ingested, content collected around the time of a change can include some data redacted under the previous settings and some redacted under the new settings.
Limitations of redaction
Redaction is a best-effort process. Consider the following limitations when you decide what information your AI agent should accept from visitors:
- Detection is probabilistic, so sensitive data can be missed, particularly when it appears in an unusual format.
- Pendo truncates prompts and responses longer than 100,000 tokens before redacting them, and appends
[truncated]to the stored content. Content beyond the limit isn't redacted or stored. - Some event metadata, such as the IP address associated with an event, might still be collected automatically when a prompt is submitted as part of standard Pendo event collection. Selecting the IP addresses category redacts IP addresses that appear in prompt and response text.
- Redaction detects sensitive data in other languages, but it's most accurate in English, French, German, Italian, Portuguese, and Spanish.
We recommend configuring your AI agents to discourage visitors from entering sensitive or confidential information, such as PII or customer data. Because redaction can't detect every possible sensitive data format and some metadata is still collected, it's your responsibility to configure your agent's messaging and behavior to prevent sensitive data collection whenever possible.
For more information about how Pendo uses AI and protects customer data, see Artificial intelligence (AI) at Pendo. For general information about Pendo's security and data privacy practices, see our Trust Center.