Private & local AI for business.
Start with what your data is allowed to do. Then decide where the AI should run.
Soxoa helps teams evaluate local models, self-hosted systems and open-source options for a specific workflow. We define the data boundary, test the work and plan who operates it after handoff.
- Approved recordsOnly the inputs this job needs
- Model + document processingCheck every service and stored copy
- Named human reviewerApprove the result before it reaches a record or customer
Every connection across this boundary needs a decision.
Illustrative design. The boundary is agreed for your workflow.Four terms. Four different questions.
Local describes where it runs. Self-hosted describes who operates it. Private describes the data controls you need. Open source describes rights and access to the system. One implementation can combine these, but none substitutes for the others.
| Approach | What it means | What to verify |
|---|---|---|
| Local AI | The model runs on a workstation or another machine near the work. | Can the full workflow run there, including document reading, search and storage? Test memory, response time and access on the actual device. |
| Self-hosted / on-premise AI | Your team operates the model service. On-premise puts the infrastructure at your own site; self-hosting can also use rented cloud infrastructure. | Who patches it, controls access, restores backups and handles a failed job? A server needs an operator as well as a model. |
| Private AI | A design goal for who can access data and how it is processed, retained and used. The label alone does not establish those controls. | Trace every copy: prompts, files, embeddings, logs, backups and connected tools. Verify the provider terms and the running configuration. |
| Open-source AI | A question about the permissions and materials available for using, studying, changing and sharing an AI system. | Review the exact model and software licenses. Downloadable weights do not by themselves establish that the whole system is open source. |
For the licensing distinction, see the Open Source Initiative’s AI definition. For a concrete runtime example, Ollama documents local-only configuration; applications connected to it still need their own review.
Give the constraint a real job.
Private AI earns its operating burden when a defined workflow needs the control.
A requirement to keep records in a particular environment, an offline work site or a well-understood repeating task can justify testing a local or self-hosted approach. Start with approved examples of that job.
A document-intake pilot
An illustrative first workflow: read an approved document, propose a few structured fields, and put the source beside each proposed value. A staff member checks the result before anything enters the system of record. Unreadable documents and missing fields go to manual review.
When to keep the existing setup
If an approved tool already meets the data requirements, compare it fairly. If nobody can own a server, the work rarely repeats or the local model creates more checking than it removes, a new deployment may add work. A stop decision belongs in the evaluation.
A pilot with a boundary and a stopping point.
Draw the boundary
Name the records the workflow needs, who may see them, where processing may happen, and what must never leave the approved environment. Agree on test material before sharing it.
Compare the smallest useful options
Test a local model, a self-hosted service or an approved hosted option only where each fits the boundary. Include the existing manual process as the baseline. A narrow rules-based step may be enough.
Make the workflow pass
Agree on acceptable errors, review time, response time and failure behavior before the test. Keep ordinary cases, difficult cases and cases the system must refuse in the evaluation record.
Name the operator
Document accounts, access, versions, costs, updates, restore steps and the manual fallback. A passing pilot proceeds to implementation only when someone can own it after handoff.
Keep the decision record
You should be able to inspect the workflow map, data-flow inventory, candidate and version list, test cases and failures, operating-cost assumptions, review path, and handoff or stop recommendation. Client security, privacy and legal owners retain approval authority.
This work can fit an AI Build Sprint or a separately scoped evaluation. Scope and price are agreed before work begins. Start with what to ask before buying an AI agent if you are comparing proposals.
Read the decision before buying the hardware.
Questions before the first test.
Based in Sacramento. Working with teams across Northern California and the Bay Area.
Does private AI have to run on premises?
No. On-premise describes a location. Private AI describes intended data controls, which must be verified. Depending on the requirements, an approved hosted service or a self-hosted cloud environment may fit. If records must stay on site, that becomes a specific constraint for the whole workflow.
Can a local model work without an internet connection?
Some can, once the model and required software are installed. Test the complete workflow with outbound access blocked: document processing, embeddings, search, licensing checks and connected tools may have their own dependencies. Downloading a model is not proof of an offline workflow.
Is open-source AI automatically private or free to run?
No. The license, hosting and data handling are separate decisions. Hardware or rented compute, integration, access controls, evaluation, maintenance and human review still need owners and a budget. Check the exact model license rather than assuming a free download grants every intended use.
Can Soxoa help us choose and implement a private AI workflow?
Soxoa can scope evaluation and implementation around one bounded workflow, approved examples and your data requirements. A Build Sprint can cover the work when the scope fits. The result may be a tested implementation, a narrower recommendation or a written stop decision. Scope and price are agreed before work begins.
Do you work with Sacramento and Bay Area teams?
Yes. Soxoa is based in Sacramento and serves Northern California and the Bay Area. Remote work and on-site working days are agreed during scoping.
Bring the workflow and its data rules.
Describe the job, the information it needs and the boundary it must respect. We will discuss what is worth testing, who needs to approve it and whether a local, self-hosted or approved hosted approach fits.