Message Vault is an important part of Zello Work. It's where an organization's communication history lives, used for everything from compliance to training. It's also the foundation for a lot of our AI features, like AI Digests and AI Insights.
At our most recent hackathon, the team I was working with took on two projects to make Message Vault even more useful: AI Message Search and Anonymized Message Vault.
The motivation for the two projects came from different directions, but they pointed at the same thing: making Message Vault work for more customers, in more situations.
AI Message Search came from wanting to give customers a simpler way to use Message Vault by letting them directly ask questions about their team's communications. Without something like AI Message Search, a customer who wanted to know "What incidents happened on the sales floor this week?" or "Did anything get flagged during the night shift?" had to filter and scroll through logs to find out.
Anonymized Message Vault came from a different problem. Standard Message Vault doesn't work for every customer. In jurisdictions like the EU, GDPR limits storing voice recordings tied to identifiable employees. Some organizations have union agreements that restrict the same kinds of data, and internal HR policies can do the same. For these customers, the choice has been "Use Message Vault and accept the legal exposure" or "Don't use it at all." We wanted to give them a third option.
The biggest call we made for AI Message Search was on the architecture. The usual way to build something like this is RAG, where you find messages that look similar to the user's question and ask the LLM to summarize them. That works for content questions, like "What incidents were reported on the sales floor this week?" But it falls apart for activity questions, like "What's the busiest time of day on the operations channel?" Even if RAG returned a hundred messages from the operations channel, summarizing them won't tell you when the channel is busiest. Answering involves a different set of tasks entirely: counting, grouping, and comparing across the whole archive.
So we gave the LLM five tools instead: semantic search, aggregation, counting, listing distinct values, and date histograms. The model gets the user's query along with the current date and timezone, and figures out which tools to call and with what filters. The result is agentic AI that figures out how to answer a question, not just something that summarizes whatever you put in front of it.
Where AI Message Search was an architecture problem, Anonymized Message Vault was a design problem: where do you draw the line on what to keep and what to drop? Too aggressive, and the archive becomes useless. Too permissive, and we haven't actually solved the legal problem.
Thankfully, our customers helped make that design problem easy by telling us what they couldn't store. So, we designed around that. Voice audio and image files don't get captured at all. Identifying details like sender names, recipient names, and participant lists are replaced in the stored record with a non-reversible "Anonymous user" label. What's kept is the operational record: timestamps, transcriptions, text messages, location messages, channel names, message types, and message durations.
The "non-reversible" part was a deliberate choice. We wanted the feature to be a real solution, not a soft one. If anonymization could be undone with the right access, it wouldn't satisfy the legal and policy obligations these customers were trying to meet.
We won the hackathon.
By the end of the week, we had both features working end to end. AI Message Search could answer real questions about real messages in seconds. Anonymized Message Vault was running on a test network, redacting what it needed to and keeping what it didn't.
We came into the hackathon with two ways we wanted to improve Message Vault. By the end of the week, both were real. I'm proud of the team that pulled it together.
If you want to see more of where voice and messaging are headed with AI, visit zello.com.