Private AI Deployment & Sandboxed Infrastructure
Your pricing formulas, operating methods, strategic plans and client records are part of what makes your business valuable. You should be able to use AI on them without sending them to an outside model provider.
We design, deploy and govern AI systems in your facility or your own cloud account: offline on local hardware, physically air-gapped where connections are prohibited, or in an isolated private cloud. You keep the model and the setup that runs it.
Every outside service that processes your information adds a relationship to govern and another place it can be exposed. A private deployment keeps processing where you already control access.
These are three different arrangements with different protections. We agree which one applies, and test it, before anything goes into use.
Offline
Models run on dedicated workstations, servers or a local cluster at your site. Documents, company knowledge, model responses and logs stay on site. Outbound internet access and external telemetry are blocked and tested, and inference does not depend on an outside API.
On site and offline, but still on a network you run. Not the same as an air gap.
Physically disconnected
For environments that prohibit external connectivity, we scope a physically disconnected installation. Models, updates and files come in through controlled import procedures, including rules for removable media.
A firewall or isolated network segment alone is not an air gap. Maintenance must respect the same boundary.
Your cloud account
If your business permits cloud infrastructure, we deploy in your account with restricted network paths, sandboxed workloads and controlled administrator access.
Not air-gapped and not on your premises. Your cloud provider's access and data-location terms still apply.
The agreed scope can include any of the following. Included controls are documented at handover, with the isolation and performance test results.
What isolation does not do. It reduces exposure; it does not make leaks impossible. People, physical access, software vulnerabilities and imported files still need controls. A private deployment is not a compliance certification, and it does not make you compliant automatically. Sector-specific obligations are reviewed with your legal and security teams.
You receive a working system, not access to a hosted endpoint.
You own your deployed copies and custom work, not the original authors' intellectual property. Model and software licenses still apply; we select models whose licenses permit your intended use. Other products keep their own licenses and charges.
Each project receives a custom quote based on capacity, isolation requirements and support needs. Prices exclude travel costs.
Scoped under Custom AI & Automation, with its inclusions:
You do not have to run a pilot or buy a workflow assessment first. Past strategy or assessment payments are credited once, within three months of your first engagement.
We build in your environment. Isometric does not host your proprietary or customer datasets. The investment is in protecting business knowledge and honoring client obligations, not a promise of lower software bills.
Yes. With a complete model and inference software installed on dedicated local hardware, inference can run without an outside API. We block and test outbound connections and external telemetry for offline deployments. Model imports, updates and support follow agreed procedures; an air-gapped installation is physically disconnected from external networks.
A no-training commitment covers one use of your data. Processing, retention, support access and subprocessors need their own review. We read the actual terms with you rather than assume every hosted service handles data the same way. Sometimes a hosted service with the right terms is enough; a private deployment is for information that should not leave your boundary at all.
No. Offline on-premises means the system runs on dedicated hardware at your site with outbound connections blocked and tested. Air-gapped means the installation is physically disconnected, and models, updates and files enter only through controlled import procedures.
No. It runs in your cloud account with restricted network paths and sandboxed workloads, but it is neither air-gapped nor on your premises. Your cloud provider's access and data-location terms still apply.
Not by itself. Private deployment is not a compliance certification. It can support your client and compliance reviews with evidence of where information is processed and who can reach it. Sector-specific obligations are reviewed with your legal and security teams.
No. Isolation reduces exposure. People, physical access, software vulnerabilities and imported files still need controls, which is why access, media handling and incident response are part of the scope.
The complete open-source or open-weight model weights, your deployed model copies, any custom adapters, the inference configuration, your custom code and workflows, and the architecture, test results and maintenance runbooks. You can hand the system to your IT or security team, or arrange ongoing care under agreed access restrictions.
No. A complete open model can be deployed privately as it is. Fine-tuning is optional; when it is used, training data and model artifacts stay inside the same security boundary.
Yes. Ownership of deployed model copies does not transfer the original authors' intellectual property. Their model and software licenses still apply; we select models whose licenses permit your intended use.
No. There is no ongoing Isometric subscription or support contract required to keep using it. Your own hardware, hosting and usage costs continue as they would for any system you run.
New builds under Custom AI & Automation include three months of ongoing support after final project delivery. After that you can run the system yourselves, negotiate continued support, or choose a full-service partnership. For air-gapped systems, support follows the same boundary as everything else.
Hardware, cloud hosting, training compute and usage charges are identified separately in the proposal. Customization or support for a private AI system you already run is quoted separately.
No. You can commission the build directly; strategy and assessment for the agreed scope are included in Custom AI & Automation. If you have already paid for a strategy session or assessment, that payment is credited once, within three months of your first engagement.
No. We build in your facility or your cloud account. Isometric does not host your proprietary or customer datasets, and any support access is agreed and restricted in advance.
It depends on the task and the hardware. We size the hardware for your workload and test the chosen model on representative work before handover, and we will tell you where a private model is the wrong fit.
Book a free 30-minute call. Describe the kind of information involved, not the records themselves; please do not send customer data, credentials or financial documents.
