Six things enterprises actually do with a private AI server — measured, end to end

Feature lists tell you what a server can do. What most teams actually want to know is: what would we use it for on Monday, and how long does the work take?

So we built a new section of the site around exactly that: six enterprise use cases, each one a complete end-to-end walkthrough — install, setup, the staged sample files, the exact prompts, and the verbatim output. Every command was run as shown on a CPU-only production server (32-vCPU arm, the 4B default model) before the page was published, and every timing on the pages is from those runs. Where a small model stumbled, the page says so and shows the guardrail.

• Document & contract extraction — 33 s per contract. Typed JSON out (ISO dates, numeric fields), self-verified by the agent, batch mode for folders. The workload with the strongest enterprise evidence for private inference: the documents legally can't leave.
• RAG with actions — 58 s. Retrieve from a knowledge base with the built-in reranker, answer with a citation, then act: our run caught a policy conflict unprompted and filed the exception ticket itself — valid JSON, exact schema, self-checked.
• Compliance & incident report drafting — 64 s. Policy + incident note in, review-ready draft out: clause-level citations, a remediation table with owners and timeframes, and open questions for the reviewer.
• Log & incident triage — 51 s. Raw application log to per-incident summaries with severity and recommended actions, counts verified with grep before writing.
• Write-and-run code — 66 s. The agent writes the script, runs it, fixes what breaks, and — in our run — flagged an underivable column instead of inventing values. With honest limits: long autonomous coding chains want the 27B tier.
• Assistants that remember — 2.6 s recall. One request field gives any app server-side conversation state; the memory API makes facts follow a user across sessions. In our run, a brand-new session recalled the account facts and silently applied a stored date-format preference. No agent needed — plain OpenAI-compatible calls.

All six walkthroughs share one design: one clear task → a few tool calls → the agent verifies its own output → a human checkpoint before anything irreversible. That shape is why a 4B model on the CPUs you already own handles these workloads — public benchmarks put small open models at near-frontier accuracy for single-turn tool calling, and the verify-and-checkpoint structure absorbs the occasional dropped instruction.

Where that stops being true, the pages say so directly — long autonomous multi-turn chains degrade steeply on all small generic models, which is what the 27B + GPU tier and human-in-the-loop framing are for.

Look at what these six workloads have in common: contracts, HR files, internal knowledge bases, incident records, application logs, customer conversations. This is precisely the data that is hardest to justify sending to a metered third-party API — and in regulated environments, often impossible. Running the whole loop — model, embeddings, reranker, memory, and the agent's tool calls — on your own hardware isn't a compromise for these use cases. It's the requirement.

Setup for all of them is the same ten minutes: one install line + a 20-line agent provider file. The memory use case needs no agent at all.

curl -fsSL https://inference-server.searchblox.com/install | sudo bash

Browse the section: inference-server.searchblox.com/use-cases.html. New use cases get added the same way these six were — validated first, published second. Timings measured 2026-09-01 – 2026-09-05 on CPU-only 32-vCPU arm servers with the 4B default model.