Is it possible to self-host AI?
Yes. A firm can run an AI model or agent process on hardware it controls. The label still needs a boundary map. LocalAI states that its agents run locally as part of the LocalAI process and can interact with external services. LocalAI agent documentation
Record the complete path for the selected workflow:
Write the workload as an input, a permitted transformation, and an expected output.
List the documents and systems the workload may read or change.
Name every model endpoint, tool, and outside service in the path. LocalAI documents local agents with tools and external-service access. LocalAI agent documentation
Name the operator and the person who approves the output.
Which consulting workflow should you test?
Use one bounded workflow: draft a client briefing from a folder of approved internal documents, then hold the draft for human approval. This makes the workload, data path, and approval point visible without treating the system as an autonomous consultant.
AI Jungle OS editorial test question: Can another reviewer replay the one-workflow test from its record? Write down the approved source folder, the requested briefing, the permitted transformation, the expected draft location, and the condition that ends the run. Keep the configuration examined for this test with the record and identify the evidence the reviewer should inspect. If the record quietly depends on another task or source, the test is not yet bounded.
| Decision field | What to record | Failure signal |
|---|---|---|
| Workload | Approved documents become a draft briefing | The task cannot be described without adding another job |
| Data path | Input folder, model endpoint, draft location | An endpoint or storage location is unknown |
| Operator | Named owner for the box and runtime | No one owns the operating decision |
| Approval | A person accepts or rejects the draft | The workflow can release the draft without approval |
| Maintenance | Named owner for updates and access | Ownership ends after installation |
| Failure recovery | Stop, inspect, and restore steps | The team cannot recover from a failed run |
| Exit test | Export the allowed records and remove access | Removal leaves an undocumented dependency |
Use the AI agent integration control map to document systems and approvals. The private box versus SaaS guide frames the operating choice without assuming that either route fits every workload.
What must the data path show?
The map must show where approved documents enter, where model processing happens, where the draft lands, and whether any call leaves the box. Ollama says that it does not see prompts or data when Ollama runs locally. The same FAQ distinguishes cloud-hosted models, which process prompts and responses to provide that service. Ollama FAQ
AI Jungle OS editorial test question: Does the inventory name every place the test data can touch? Ask for the source folder, temporary files, prompt assembly, model endpoint, retrieval store, logs, draft destination, backup location, and each external call that is enabled in the tested configuration. For every entry, record what enters, what leaves, and which person can inspect it. An empty field is a question to resolve, not evidence that no data passes there.
AI Jungle OS editorial test question: Where is the permission boundary for this single workflow? The review should identify the account used at each step, the exact folders or systems it may read, the destinations it may write to, and the actions it must not perform. Then ask whether the test still completes when unrelated access is removed. Treat this as a question about the configured path, not as a general security guarantee about the box or model.
That distinction is why “runs on our box” is not enough. Inspect the configured model mode and every enabled connection. Do not infer the route from the interface or product name.
Who operates the box and who approves the work?
Assign the operating decision and the consulting decision to named people. The operator owns the runtime, access, maintenance, and recovery. The approver accepts or rejects the client-facing draft. Keep those decisions visible in the workflow.
AI Jungle OS editorial test question: Which named human makes the release decision for this briefing? Put that person in the record, show them the source set and draft, and capture an explicit accept or reject decision before the output can move to a client-facing destination. Also ask who covers the decision when that person is unavailable. A role label without an accountable person leaves the approval boundary unclear.
AI Jungle OS editorial test question: Can the operator run the failure recovery drill from written steps? Choose a failed or interrupted test run, stop further processing, preserve the evidence needed for inspection, identify the last known valid input and output, and restore the workflow to its agreed starting state. Record what the operator checked before allowing another run. The question is whether this workflow can be recovered, not whether every possible failure has been eliminated.
AI Jungle OS editorial test question: Who owns maintenance after the initial test? Name the person who reviews access, configuration changes, model or tool changes, stored records, and the continued validity of the approval point. Define which change sends the workflow back through acceptance instead of being treated as routine upkeep. Ask where that decision is logged so the operating record does not end on installation day.
AI Jungle OS editorial test question: Can the firm complete the exit handback without relying on the departing setup? List the approved records that must be returned, their usable format, the person who receives them, the credentials and connections to remove, and the evidence that removal occurred. Then check whether the client briefing process can continue by its agreed fallback. This tests the handback for the named workflow only and does not claim that every dependency has disappeared.
Before the workflow is accepted, check these items:
The named input is the only approved input.
The configured data path matches the written map.
The draft cannot pass the approval point on its own.
The operator can stop the workflow and inspect the failure.
The exit test removes access and returns the agreed records.
What is the best self-hosted AI?
There is no useful universal answer for this consulting decision. “Best” must refer to the named workflow and its acceptance test. Select a candidate only after the firm has written the input, output, data path, approval point, maintenance owner, recovery method, and exit test.
A model result that looks good in an open chat does not verify the operating path. Test the complete configured workflow against the expected briefing and the stated failure conditions.
How expensive is it to host your own AI?
Do not estimate the decision from hardware alone. Build the cost scope from the chosen workload and the operating responsibilities attached to it. Include the box, the configured model path, the named operator, maintenance, failure recovery, and the exit work. This page does not claim that self-hosting costs less.
Is hosting your own AI worth it?
It is worth testing when the firm needs the chosen workflow on a controlled box and can own the complete operating record. Reject the fit when the data path remains unknown, no operator accepts maintenance, approval can be bypassed, recovery cannot be demonstrated, or the exit test fails.
The verdict belongs to this workflow. It does not prove that self-hosting fits every consulting task or that the box provides absolute sovereignty, security, performance, or savings.
What is the 30% rule for AI?
This decision does not need a percentage rule. Use observable acceptance conditions for the selected workflow. The draft must come from the approved input, follow the mapped data path, stop at the approval point, expose failures to the operator, and pass the exit test.
Can I build my own AI for free?
Do not use “free” as the acceptance test. A consulting firm still needs to assign the box, model path, operator, approvals, maintenance, recovery, and exit work. Record those responsibilities before choosing software.
If the workflow passes its checks, keep its scope fixed. If it fails, stop and document the failed condition before changing the box, model, or workflow.
Written by Tileo, who operates a portfolio of internet businesses on this same cockpit.