Everyday office work
The AI tools you already license and govern, for general work with no client-confidential or proprietary material in it.
The private AI deployment pattern
FoxGrade
FoxGrade is the private AI deployment pattern Fuchs & Agents uses for clients. It runs where your data lives, keeps each client walled off, and a person signs off on everything it produces.
FoxGrade data line
Private AI, kept local.
Protected source material stays inside the boundary you approve: your hardware, your cloud tenancy, or a hybrid. We check that boundary for each deployment before we use the line about it.
Two lanes
Lane 1: your cloud AI
Lane 2: FoxGrade
A simple rule for which work goes where
Proprietary and client work
Anything that names a client, describes a client's process, or is your own proprietary material. Each client or matter gets its own workspace.
The only path between the lanes is a person publishing an approved document.
How it behaves
Five rules built into the system
AI drafts, people decide
Every output has a named approver. Nothing is published automatically.
Stop and flag
Unknowns are shown as unknown. A missing number or an unsourced claim blocks approval.
Isolation in code
Each client has its own storage, index, glossary and logs. Planted-leak tests must all be caught before a release.
Evidence first
Factual sentences carry a source reference. An append-only log records who did what.
Measured, not asserted
A set of your best past examples scores every model or prompt change. A change that scores worse doesn't ship.
These are design requirements. Before we call a specific deployment FoxGrade, its approval flow, isolation, sourcing, logs and regression checks are tested and accepted. See the records each rule leaves behind, shown on a fictional business.
Three of the rules can be tried in your browser on made-up data:
Deployment
Move the compute to the data
FoxGrade ships as one versioned package. You pick where it runs.
On your premises
On hardware you own, for your own material and for clients who agree to it.
In your cloud account
Inside your own Azure, AWS or Google Cloud tenancy, under your access controls. Only approved deliverables come out, and you pay your own compute.
Hybrid
Methodology in one copy, client data in another, and nothing raw crossing between them.
Cloud providers' own data terms apply inside each tenancy. We review them with you for each deployment.
The economics
Judge AI by cost per finished document
Per-seat prices and token prices don't tell you much on their own. What matters is the cost of each approved output: a drafted report, a tailored training module, a reviewed page.
At low volume, metered cloud AI is often cheaper on raw token cost, and we'll say so. The case for running locally at that volume is control: client data stays on your network, and a known hardware cost replaces an open-ended bill. At higher volume, local can also win on cost.
Read the worked example, with every assumption labelled.
What we measure first
Before anything is bought, the first phase measures:
- Reviewer minutes per draft
- Share of drafts that need only light edits
- Drafts per hour and speed on the candidate hardware
- Cost per approved output
Reviewer time usually costs far more than the AI. That is where the savings are, so that is what we measure.
"Audit-grade" describes logged, reviewable work. It is not a certification or an audit opinion.