Editor's note: This case study has been redacted using Tonic Textual at the customer's request.
[ORGANIZATION] was unable to pursue AI innovations in the tax solutions they offer their clients due to the sensitive nature of the data those workflows would need to access. By deploying Tonic Textual as a privacy layer within LLM workflows to prevent sensitive data from leaking into data models or cloud environments, they unblocked the development of solutions to optimize the efficiency and precision of their services. Tonic Textual enables [ORGANIZATION] to securely handle all of their data through industry-leading accuracy in sensitive data detection and realistic data synthesis, opening the doors to innovation and a stronger competitive edge.
The challenge: AI initiatives blocked by inaccessible data
Based in [LOCATION] in Europe, [ORGANIZATION] is a leading tax agency offering full-service tax solutions with a tech-forward approach to innovation. “We have a very startup-like culture in the way we do things, prioritizing technical knowledge among all our team members and the rapid prototyping of internal solutions,” explained [OCCUPATION], [NAME_GIVEN] [NAME_FAMILY].
With this mindset, they were keen to leverage large language models (LLMs) to prototype and implement AI-driven processes that would enhance the services they offer their clients. But using real-world, customer data to build and train these solutions was a non-starter. The sensitive nature of the financial data they handle and their client’s trust in how they handle that data are paramount to [ORGANIZATION].
Their AI initiatives were blocked and could not get off the ground, despite all the ideas they had. “We had a lot of discussions around what we wanted to do, but they always ended in data privacy concerns. It was a lot of, ‘That would be nice, but we can't do that. That would be nice, but we can't do that,’” shared [NAME_FAMILY]. “Ultimately, we found ourselves asking, ‘What can we even do?’”
To unblock innovation, they needed a secure and compliant way to strip their unstructured data of sensitive or personally identifiable information (PII) to make it safe for use in AI initiatives, including prototypes of their envisioned workflows. A very high degree of accuracy in sensitive data detection and anonymization was critical, in addition to the ability to process the [LANGUAGE] language and the ability to self-host the solution, rather than using it on the cloud.
[NAME_FAMILY] led the search, prioritizing solutions with an active Github repository. It was here that he found several promising options that he put to the test, including open-source tools and Tonic Textual.
“We quickly found that the performance of even well-known open-source solutions was quite underwhelming and not accurate enough to be something we could use in good conscience for our use cases,” stated [NAME_FAMILY]. “Those open-source tools were definitely worse in [LANGUAGE], as well.”
Turning to Tonic Textual, [NAME_FAMILY] dove into validating its effectiveness directly on the product’s website. From there, he got up and running with a free trial to further verify that the solution fulfilled his use case.
“Tonic.ai was a perspective shift,” [NAME_FAMILY] affirmed. “It’s clear that your company and solutions are very developer focused. You have a team that's very passionate about what they’ve built, and your solution met all our criteria. This solidified our decision to go with Tonic Textual.”
The solution: a self-hosted privacy proxy embedded within LLM workflows
[ORGANIZATION] deployed Textual self-hosted as a privacy proxy for LLMs and immediately unblocked work on several AI initiatives. For each initiative, they first built and tested a prototype before rolling out the AI-driven workflow with Textual embedded as the privacy layer.
The first initiative involved building an in-house workflow to better transcribe and summarize meetings securely. They rely on Krisp AI, hosted locally, for transcription services. But to enhance the quality of the transcription and generate summaries with clear action items, they needed to send the data to a cloud app to be processed by an LLM. This is where Textual steps in to ensure the utmost in data security.
Before being sent to the cloud, the initial transcripts are fully de-identified in Textual, which replaces sensitive information in the transcripts with synthetic replacements. When asked about the choice between redacting vs synthesizing their data, [NAME_FAMILY] replied, “We are using data synthesis, as a matter of fact. Textual makes a seed parameter accessible for us so we can have a deterministic output for the synthesis.” This deterministic, or consistent, data synthesis preserves relationships found within the data, as well as contextual information, to maintain the data’s realism and enable effective LLM processing.
Once the de-identified transcripts are polished and summarized by the cloud-based LLM, they return to [ORGANIZATION]’s local environments, where they are de-encrypted to replace the synthetic values with the original data so employees can leverage the transcripts and summaries in serving their clients.
A second initiative for which they’ve built a prototype and will use Textual as the privacy layer in their live workflow involves automating the process of filing documents from various sources. Here, they’re leveraging AI to learn from existing files on their server to streamline file placement and optimize management within the server, whether those documents come from an internal team member or directly from a customer.
“Basically, Tonic Textual unblocked all of our AI initiatives,” stated [NAME_FAMILY]. “We’ve been able to build the infrastructure that will power them all, abstracting the Textual SDK and creating a wrapper that integrates well with our systems. So now we have a button that we can press to activate data privacy, allowing us to move forward with AI. It is quite transformative.”
The results: unlimited AI initiatives unblocked with realistic data synthesis
With Tonic Textual, [NAME_FAMILY] explained, “We have gone from: ‘We cannot do that. We cannot do that. We cannot do that.’ To: ‘We can do almost anything now; we will just have Tonic as a privacy layer in between. Let’s build a prototype and see what works.’”
Accelerating these innovations is Textual’s ease of use, which has made it accessible to less technical members of [ORGANIZATION]’s team who are able to redact documents without any additional effort or technical support.
The first two initiatives that [ORGANIZATION] have rolled out are just the beginning. They’ve already identified multiple future initiatives for which Textual will play a critical role in safeguarding data privacy, including improving internal document search capabilities for their employees. “We want to make that much easier with an AI vector-based search.” Whether leveraging an LLM or a RAG system to power that search, Textual will serve as the process’s privacy proxy.
Thinking further into the future, [NAME_FAMILY] said, “As a mid- to long-term vision, we want to have a specialized AI client that has a deep understanding of our customers but is safe to use from a data privacy perspective because Textual will be there as a protective layer, preventing sensitive information from entering the model.”
The results of these initiatives are heightened efficiency and an elevated quality of service that [ORGANIZATION] is able to offer its clients. “We want to improve efficiency but also accuracy in the work that we do,” stated [NAME_FAMILY]. “AI allows us to work at a pace that was not possible before, and with a very high quality output.”
A high quality output made possible by high quality data synthesis—this is the promise that [ORGANIZATION] needed and that Tonic Textual delivered.
.svg){kind=icon}
