Use cases

Workload patterns for institutional AI.

These patterns explain how institutions can frame an on-premise AI requirement. They are not customer case studies, deployment claims or promised outcomes.

How to read this page

Everything below is a workload pattern, described by the kind of institution that runs it and the kind of question it gets asked. Hosn does not publish customer names, deployment counts or case studies, because the institutions this product is built for do not consent to being named and we will not invent a reference to fill a page. What follows is what the software does, framed by sector.

Audit Copilot, for oversight and regulatory bodies

An audit team's problem is rarely a shortage of documents. It is that the answer is spread across a budget file, three quarterly returns and a memorandum nobody indexed. An Audit Copilot workload puts the whole set behind retrieval and lets an auditor ask the question directly: compare a ministry's fourth-quarter spend against the approved budget and flag the line items that exceeded the ceiling.

Two properties make this workable rather than a demonstration. Spreadsheets are queried as structured data rather than summarised as prose, so figures are read rather than approximated. And every query and answer lands in the append-only audit log, which matters because an oversight body's own working papers are themselves subject to review. Related reading: anomaly detection and audit copilot patterns and how state audit institutions can use AI without compromising independence.

Legal Counsel, for law firms and legal directorates

Legal research is the workload where a cloud service is hardest to justify, because the corpus is privileged by definition. On-premise, a firm can index its own matter files alongside the Omani legal corpus and ask for a summary of a ruling with the precedents it draws from, in Arabic, with citations that point back at documents the firm can open.

The Arabic side is not incidental here. Retrieval folds diacritics, so a search typed without harakat still matches a judgment that carries them, and the models shipped are open-weight Arabic-strong models rather than an English model behind a translation step. Related reading: AI for legal research in Oman, on-premise indexing of the Omani legal corpus, and legal research inside the office.

Finance Analyst, for banks, regulators and sovereign funds

A analyst's version of the same problem is a folder of quarterly reports and a question that needs one number out of each: read these and extract every institution's capital adequacy ratio. That is a retrieval and extraction workload, not a chat workload, and it is exactly the kind of task where sending the source material to a third party is the part that cannot be approved.

For supervisory and AML work the audit trail is again the deciding feature: a supervisor has to be able to show what was asked, what was returned and against which version of the document. Related reading: credit memos, KYC and AML on private hardware, supervision without sharing regulated data, and the sovereign wealth fund thesis on on-premise AI.

Medical Scribe, for hospitals and clinical centres

Clinical dictation is a personal-data problem before it is an AI problem. Transcribing a consultation into a structured note means the recording, the transcript and the note are all patient data, and under the Omani Personal Data Protection Law the safest architecture is the one where none of it moves. On-premise transcription keeps the audio, the model and the output on hospital hardware.

The specific requirement clinicians raise is mixed-script handling: an Arabic note that must keep Latin drug names and clinical terms intact rather than transliterating them. Related reading: Arabic medical scribing in the Sultanate and medical scribing in health ministry workflows.

Archives, correspondence and records

The workload institutions underestimate is the archive itself. Decades of Arabic paper, scanned at varying quality, is the material most likely to be sitting outside any search system at all. Ingestion, OCR and classification are the unglamorous part of a deployment and usually the part that produces the first visible result. Related reading: Arabic NLP for government archive digitisation, OCR for ministry archive programmes, and bilingual correspondence workloads.

Which configuration a workload needs

Workload shape decides configuration more than headcount does. A single legal directorate with a few researchers fits Hosn Kernel. A floor running drafting, analysis and records at the same time needs both models resident, which is Hosn Tower. An institution that wants a model adapted to its own classified corpus needs training to run on its own GPUs, which only Hosn Rack does. The configurations page puts the three side by side.