Self-Sovereign AI
Graduate from the API.
If AI already runs in production, your people create useful material every day: the work they approve and the corrections they make. We use it to move tasks to a model you own, one at a time.
Should this task stay on the API?
A rented model API is how most production AI begins, and it is a sound beginning. Over time, a business learns which tasks repeat, what good output looks like, and what each task costs.
That knowledge is the starting point for an owned model. The question is task-specific: would an owned model clear the same acceptance criteria, under better terms for your business?
Signals worth measuring
- The same kind of request arrives many times a day.
- People already review and correct the output.
- API spending grows with volume on a narrow task.
- Data rules favor processing inside your boundary.
- The workflow needs a model that changes only on approval.
What you already have
Production use leaves a record. Much of what a training run needs is already in your systems.
Reviewed work
Outputs your people accepted, with the inputs that produced them and the reason they passed.
Corrections
Edits and rejections your team made. They show what the task requires more clearly than any prompt.
Rules and prompts
The instructions, formats, and business rules the current workflow already follows.
Volume and cost records
How often each task runs and what it costs today, so the comparison uses your numbers.
Acceptance decisions
What your reviewers accept and reject. This becomes the acceptance criteria for the owned model.
A working baseline
The current API is the benchmark. Every owned candidate is measured against it on held-out work.
How a task graduates
One task at a time, with the API still running until your approver accepts the switch.
- CaptureCollect reviewed work and corrections, with rights and provider terms checked.
- EvaluateMeasure owned candidates against the current API on held-out work.
- Run side by sideRoute real work to both and compare accepted outcomes.
- ApproveYour approver reviews the evidence and accepts or declines the switch.
- SwitchMove the task. Keep the API as the rollback route.
The free AI Workload Evaluation tells you whether one task is a candidate and what to do next. Deeper comparison and training are scoped separately, as fixed-price or hourly work.
Training itself runs through Model Training & Fine-Tuning. Sharper instructions, context, and retrieval come first; training follows when the evaluation shows it pays.
Each graduated task gets its own decision. The API can keep serving every task that has not yet cleared the bar.
What each stage leaves behind
Every stage produces a record your approver can read. The switch decision rests on those records.
Capture
A dataset of reviewed work and corrections. Each item records its source, its rights, and the version it belongs to.
Evaluate
A held-out test set drawn from your own work, and a scored comparison of each candidate against the current API.
Run side by side
Records of real work handled by both paths, with your reviewers' accept and reject decisions for each output.
Switch
An approved routing change for one task, a Model Passport for the owned model, and a tested rollback route.
Rollback stays in place after the switch. If accepted outcomes slip, your approver can route the task back to the API while the cause is found.
Check your provider's terms first
Some providers restrict using their outputs to train other models. Those terms come first, before any output is used for training.
We review the relevant terms with you at the start. Your counsel confirms the reading.
The API stays where it measures better
Graduation is a measurement, not a migration plan. Some tasks are best left where they are.
- Broad or unpredictable tasksWork that changes shape often may measure better on a large frontier model.
- Low-volume tasksWhen a task runs rarely, the effort of owning a model may not repay itself.
- Tasks without reviewed examplesStart by capturing reviewed work. Train once there is enough to measure against.
- No owner for the loopAn owned model needs someone to approve releases. Without one, keep the API for now.
A hybrid result is a good result. Many businesses finish with owned models on their core tasks and a rented model on the rest.
Governance for each switch
A person approves every switch. These records support that decision.
- Each dataset records its source, rights, version, and permitted uses.
- The owned candidate clears the agreed held-out tests before any switch.
- A Model Passport records the base model, license, data versions, results, limits, and approvals.
- The API remains the rollback route after the switch, retired only when your approver decides.
- Full engagement terms are finalized in a Master Services Agreement.
Typical applications by industry
Where graduation tends to fit
Typical applications across industries. They show where graduation applies, not past client work or results.
Insurance agency
Policy-change confirmations
The task repeats daily, and account managers already correct each draft. Those corrections form the test set. An account manager approves the switch, and the API stays as rollback.
Accounting firm
Client document classification
Staff already confirm how each incoming document is classified. An owned candidate runs beside the API on the same documents. A senior accountant approves or declines the switch.
Property management
Maintenance request triage
Managers correct urgency ratings against a written policy. The owned model takes this narrow task once it clears the held-out tests. Urgent cases still reach the on-call manager directly.
E-commerce
Product descriptions, split by line
A stable catalog line moves to an owned model, built from reviewers' edits. Seasonal campaigns with changing briefs stay on the API. A merchandising lead approves each routing change.
Medical billing
Denial reason coding
Sensitive records favor processing inside the company's own environment. Past corrections form the test set after a rights review. A billing lead approves the switch after the side-by-side run.
Law firm
Clause extraction from standard agreements
Attorneys already review extracted clauses for one agreement type. The firm's own reviewed work trains the candidate. The supervising attorney approves the switch, and broader research stays on the API.
Questions
Can we train on outputs from our current API?
Only where the provider’s terms permit it. We review the terms with you first, and build from your own work and corrections where they restrict it.
Do we have to move every task?
No. Each task gets its own measurement and decision. Many businesses keep a rented model for some work.
What if the owned model measures worse?
The task stays on the API. The evaluation still leaves you with a test set that measures every future candidate.
How is the side-by-side run handled?
Real work is routed to both, within the agreed data boundary. Reviewers compare accepted outcomes before anyone approves a switch.
Where does the owned model run?
In your cloud account, on hardware you buy, or in a hybrid arrangement. The choice follows your measured requirements.
Will an owned model cost less than the API?
It depends on the task, its volume, and where the model runs. The comparison uses your own volume and cost records, and the decision follows them.
Can we roll back after a switch?
Yes. The API stays available as the rollback route after the switch. Your approver can route the task back if accepted outcomes slip.
How long does a task take to graduate?
It depends on the reviewed work available and the length of side-by-side run your approver needs. The timeline is agreed when the work is scoped.
Who owns the trained model?
Model Training & Fine-Tuning delivers a trained model you own. Full engagement terms are finalized in a Master Services Agreement.
Keep exploring
Your data
Turn reviewed work into training and evaluation sets you keep.
Your models
How an open-source or open-weight model is adapted and released.
Your cloud or your servers
Where the graduated model runs, and its data boundary.
Model Training & Fine-Tuning
Adapt a model to your work, with an evaluation set you keep.
Agentic Engineering Diagnostic
$750: one hour of agent coding, the code, an Agentic AI Blueprint, and one hour of consultation.
Built to improve as models improve
The Canon rule for adopting new models with rollback preserved.