Private AI — Grok
Grok, on a leash.
Grok — xAI's frontier model and chatbot — can't be self-hosted in any practical sense. That doesn't mean banning it, and it definitely doesn't mean pretending your team isn't already using it. We put it behind a policy gateway: sensitive data redacted before it leaves, every request logged, access by allowlist.
The problem with frontier bots
Your team already uses it. The question is whether it's governed.
People paste contracts, customer lists, and half-finished code into public AI chat because it's fast and it's right there. A written policy against it changes nothing; a gateway changes the math. With one in place, the safe path and the convenient path become the same path — and "what has Grok seen?" stops being an unanswerable audit question.
Policy gateway, not a ban
Grok access flows through a gateway we operate — in your office or our data center. Direct browser use gets redirected to the governed path, not just discouraged.
Redaction before egress
Names, account numbers, card data, health fields — sensitive patterns are detected and masked before a request ever leaves your network. The model does the work without seeing the PII.
Team allowlists
Who can use Grok, for what task categories, through which integrations — decided per team, reviewed quarterly, revoked in one ticket.
Full request logging
Every call — who, when, what tool, what was redacted — lands in the same centralized logging as everything else, feeding your SOC 2 evidence automatically.
Contract and terms review
We configure provider settings toward no-training-on-your-data options where offered, and we document what the provider actually does with the data — so your risk register says something real.
One gateway, any model
The same gateway governs other online models too. When the hot new assistant arrives next quarter, it's a policy change — not a new shadow-IT incident.
How we set it up
Four steps to governed access.
Classify the use
Which teams need frontier capability, for which tasks — and what data categories must never reach an online model. This becomes the gateway policy.
Typical: 1 week
Deploy the gateway
A small hardened gateway — your office, or our data center next to your local model node — with redaction rules tuned to your data.
Typical: 1–2 weeks
Connect the team
Allowlists applied, integrations wired into the chat and tools people already use, direct access redirected. Users notice convenience, not restrictions.
Typical: 1 week
Log and review
Every request logged, monthly usage report, quarterly allowlist review — evidence your auditor can read without a translator.
Ongoing
The line
Where Grok belongs — and where it doesn't.
| Good through the gateway | Never leaves your network |
|---|---|
| General research, drafting, and summarization | Customer records and PII |
| Code assistance on non-proprietary work | Contracts and legal documents |
| Marketing copy and public-facing content | HR, payroll, and health data |
| Brainstorming and non-sensitive analysis | Credentials, keys, and internal network details |
The sensitive side of that table runs on local models — Hermes and peers — on your own hardware. One architecture, a clear line between the two.
Make the safe path the easy one.
We'll map which AI tools your team is already using, then deploy the gateway that keeps the capability and removes the leak.