Customer Agent Recording Systems and Methods Stripping Sensitive Data Out of Call Recordings Before Anyone Analyses Them

This application describes a way of recording customer and agent interactions through a customer relationship management system, identifying sensitive information within the recording, removing it, and then analysing what remains to produce summaries and analytics. The applicant is Ujwal Inc., which trades as Level AI.

The Problem

Contact centres generate an enormous volume of conversation, and almost none of it gets reviewed. Traditional quality assurance samples a tiny fraction of calls, which means most of what customers actually say never reaches the people who could act on it. Automated analysis of every interaction would change that, producing feedback useful to managers, to agents and ultimately to customers.

The obstacle is what the conversations contain. Customers read out card numbers, quote account details, give addresses, confirm dates of birth and mention health or financial circumstances. All of that lands in the recording. Once it is there, the recording becomes a regulated asset, subject to privacy law such as the Australian Privacy Act and to payment card industry rules, and it cannot simply be fed into an analytics pipeline or a machine learning model. Many organisations respond by pausing recording during sensitive moments, which loses data, or by restricting who can access recordings at all, which defeats the purpose of collecting them.

What This Invention Does

The method inserts a redaction stage between capture and analysis. A recorder application records the customer and agent interaction. Sensitive information within that recording is then identified, and the portions containing it are redacted and removed from the recording itself rather than merely masked at the point of playback.

Only the redacted recording proceeds to analysis, which generates summary and analytics information from the conversation. The ordering is the point. Because the sensitive material is gone before the analytical stage runs, the downstream systems, the people reviewing output and any models trained on the material never hold the regulated data in the first place. That converts a compliance problem into an architectural one, and it is a considerably easier problem to defend to an auditor than a set of access controls sitting on top of a recording that still contains card numbers.

Key Features

• CRM-integrated recording. A recorder application captures the customer and agent interaction within a customer relationship management system.
• Sensitive information detection. Sensitive portions of the recording are identified automatically.
• Redaction and removal. Those portions are redacted and removed from the recording rather than hidden at playback.
• Post-redaction analysis. Analysis runs only on the redacted recording.
• Summary generation. The analysis produces summary information about the interaction.
• Analytics output. Actionable analytics are generated for managers, agents and customers.

Who Is Behind It

The applicant is Ujwal Inc., a Mountain View, California company founded in 2019 that trades as Level AI and builds analysis and quality assurance software for contact centres. The named inventors include the company’s founder and chief executive Ashish Nagar, along with Bhupesh Rohilla, Shlok Kapoor, Aadarsh Verma, Abhimanyu Talwar, Sumeet Khullar, Shubham Goswami, Neeraj Paliwal, Kishan Joshi and Manisha Barnwal. The application is a divisional of Australian application 2024304600.

Why It Matters

The economics of conversation analysis changed when natural language processing and speech recognition became cheap enough to run on every call rather than a sample. What has not changed is the regulatory position of the underlying recordings, and that gap is where most contact centre analytics projects stall.

Building redaction into the pipeline ahead of analysis is a recognisable pattern from other regulated data domains, notably de-identification in health research, where the same logic applies: remove the identifying material at the earliest possible point, then treat everything downstream as a lower-risk asset. Applied to voice, it is harder than it sounds, because sensitive content in speech is not neatly delimited the way a database field is, and a redaction system that misses a card number is worse than useless. But the direction is one regulators tend to favour, since it reduces the volume of sensitive data in existence rather than adding another layer of controls around it.

Related Concepts

• Customer relationship management – the system this recording capability sits inside.
• Data redaction – the practice of removing sensitive content from a record.
• Speech recognition – the technology that makes automated conversation analysis possible.
• Natural language processing – the field behind the summary and analytics stage.
• Information privacy – the regulatory pressure driving the redaction step.
• Office of the Australian Information Commissioner – the Australian regulator overseeing personal information handling.

---

AU 2026202054 was published in the Australian Official Journal of Patents on 9 April 2026 and is open for public inspection. Patent applications represent inventions that are sought to be protected and do not necessarily reflect commercially available products.