# Governed: Three jobs. The same platform underneath.

Tileward's use cases start with what the assistant has to do: answer from approved knowledge, carry a long-running thread, or run privately on hardware you control. This page names the three job patterns, the part of Tileward that answers each one, and three worked examples by industry.

Page: https://tileward.com/use-cases/

## Three job patterns

The same three jobs recur in finance, healthcare, manufacturing, public-sector work and support. Each one has a cost or a control problem, and one part of Tileward that answers it.

- **Private knowledge assistants: answer from approved files.** A line in a model's prompt is not a control anyone can check. With Tileward Documents and Tileward Governance, only approved files can answer, every answer comes back with the passages it was given, and requests on locked tiles are refused before the model writes a word. 2,201 of 2,251 requests written to get past a locked tile were stopped before the model saw them.
- **Long-running agents: keep the history without resending it.** Every step of an agent resends the whole history, so the bill grows with the thread rather than the work. Tileward Context keeps the history and sends each step only the part that answers: 68% less context (3.2×) sent on agent work, with the same work done.
- **Private model deployment: run the model where the data lives.** When the model the job needs does not fit the GPUs you have, or the data cannot leave the building, Tileward Models' compressed weights run a bigger model on the cards you already own: hosted, in your VPC, or air-gapped. Tileward 35B-A3B scores level with the publisher's own FP8 build on MMLU; the [model library](https://tileward.com/model-library/) has the current sizes.

## Worked examples

Each assistant answers only from the files it was given, and the passages it was given come back with the answer. Its locked tiles are switched on per API key, refused before the model answers, and every decision is recorded.

| Assistant | Answers from | Refuses, for example |
|---|---|---|
| [Financial services](https://tileward.com/use-cases/financial-services/) | Product and suitability policies, approved client-facing wording, escalation runbooks | Personalized investment advice, options pricing, personal tax advice |
| [Healthcare](https://tileward.com/use-cases/healthcare/) | Formularies and dosing policy, care pathways, coding and claims guidance | Personal medical advice, medical devices, health insurance and claims |
| [Customer support](https://tileward.com/use-cases/customer-support/) | Support handbooks, refund and warranty policy, product manuals | Personal legal advice, personal tax advice, data privacy and PII |

## Try it

- Ask TWFP4 and NVIDIA's NVFP4 the same question, live, on the [comparison page](https://tileward.com/use-cases/compare/).
- See it on your own policy: bring the documents, and what it must not discuss. An engineer will ground the answers, lock what it must not discuss, try to get past it, and show you the record. Thirty minutes, no deck. [Talk to an engineer](https://calendar.app.google/hg2DzFTxKJgwFWm19).

## Related

- [Tileward Documents](https://tileward.com/documents/): answers from approved files, with citations
- [Tileward Governance](https://tileward.com/governance/): locked tiles refused before inference
- [Tileward Context](https://tileward.com/context/): send only the conversation that matters
- [Tileward Models](https://tileward.com/models/): run larger models on existing GPUs
- [Model library](https://tileward.com/model-library/): every model, its size and its price
- [Start free](https://app.tileward.com/?join=explore): the Tileward console, no card