What does no silent model substitution mean?

Every task class names its approved model, its budget, and its data class in advance. If the approved route is unavailable, the task stops and asks — it never quietly reroutes to a different model, and private company data is never uploaded to a destination the task class has not approved.

Which voice standard do GEN voiced assistants use?

Any future voiced assistant uses the Gemini 3.1 TTS standard, is started explicitly by the user, keeps public knowledge separate from private retrieval, and never captures a microphone automatically.

Knowledge & AI Efficiency

Spend AI well — in the open, on a budget you can see

GEN's differentiator is not a model. It is discipline: your knowledge kept properly, your AI spend visible in buckets and budgets, and routing rules stated as policy you can hold us to.

Request a Workflow Review

Knowledge first

AI is only as good as what it may read. The company Knowledge Vault discipline comes first: every record sourced, owned, versioned, and approved, with permission-scoped retrieval and citations. An assistant that cannot find an approved answer says so. That restraint is what makes the answers usable.

Credit efficiency, taught plainly

Most AI waste starts with a misunderstanding: "credits" are not one thing. App credits that work only inside a provider's apps, metered API usage billed per run, and CRM wallet credits that pay for messages and automations are three different buckets, and they are not interchangeable. GEN training teaches your team which bucket a task draws from before it runs — and your company sees its own usage in a cost ledger, working to an explicit per-company budget. Every AI run logs its model and its cost against that budget; going over budget stops the run class, it does not just produce a report about it.

We claim no savings percentages. None have been measured on your firm yet, and invented efficiency figures are exactly the habit this page exists to break. Your baseline comes first; the numbers follow from it.

Model routing, stated as policy

Approved models only, per task class. The routing table names the model, the budget, and the data class for each kind of task — before any run happens.
No silent substitution. If the approved route is unavailable, the task stops and asks. It never reroutes itself to a different model.
No silent upload. Private company data never leaves for a cloud model unless that task class explicitly approves that destination.
Local models for suitable tasks. For fed-content work — digests, classification, internal drafts on de-identified material — an approved local model keeps data on your own ground. The limits are stated honestly: local models are slower, they work in fed-content mode with a named human reviewer, and they never fetch their own inputs or make judgment calls.

Voiced assistants

Any future voiced assistant on the framework follows one standard: the Gemini 3.1 TTS voice standard, started explicitly by the user — never listening by default, never capturing a microphone automatically. Public knowledge and private retrieval stay in separate lanes: a voiced assistant answers public questions from public material, and private company knowledge is retrieved only inside the permissions of the person asking. Pre-recorded voice and live conversation are separate routing classes, and neither is ever swapped for the other silently.

Know what your AI is doing

Bring one workflow. We will map its knowledge, its buckets, and its budget — before anything runs.

Request a Workflow Review