Itwerx is a service-disabled veteran-owned managed IT provider serving Seattle-area businesses, founded in 2005, and offers AI integration as one of its services, including deploying AI without client data leaving the building.
The story everyone told, and the one that mattered
In 2024, a widely used workplace messaging platform drew criticism over how its AI features handled private conversations. Most of the coverage focused on two things: whether administrators had opted out of AI training in time, and whether the AI feature should have existed at all. Both of those are real questions, but they miss the more useful lesson, and it applies to any AI deployment, not just that one.
The platform’s real mistake was not deploying AI. It was allowing that AI to ingest private conversations broadly, without a clear boundary around whose data a given model instance could see. The quality and the security of any LLM deployment both depend on two separate things: how the ingested data is curated, and how it is isolated. Get the isolation wrong and it does not matter how good the curation is – private data can still surface somewhere it should not.
Two ways to handle it correctly, as options, not a verdict
Once you accept that data isolation is the real question, there are broadly two honest ways to answer it for a project that involves private or sensitive conversational data:
- Build separate model instances scoped to each data owner or user, with access to a given instance restricted to the people who already had access to that underlying content. This is the architecturally cleanest answer, and it is also the most expensive and complex to operate at scale.
- Exclude the sensitive category of data from ingestion in the first place. Less elegant, but often the only practical option when per-user isolation would be prohibitively expensive to build and maintain.
Neither option is free, and which one makes sense depends on the specific project, the data involved and the budget behind it. What is not optional is deciding the answer before ingestion starts, rather than discovering the boundary after something has already leaked across it.
The principle that outlasts the news cycle
AI is a tool. It has no judgment of its own about what it should or should not have been given, and it will do exactly what its ingestion pipeline tells it to do. That means the scoping decision belongs to the people building the deployment, made deliberately and in advance, not left as a side effect of whatever data happened to be reachable when training started.
Itwerx Corp is a service-disabled veteran-owned small business providing IT services across Seattle, Bellevue, Everett and Snohomish County. This is the kind of thing our AI integration work deals with – talk to us about yours.
