AI configuration
What this page is — the tab that sets AI behaviour for every document of a type: the drafting assistant, which analyses run automatically at submit, each counterparty round and finalize, and two advanced JSON settings.
What it is for — so the right analysis is waiting on a document when someone opens it, without anyone remembering to run it.
The problem it solves — useful analyses only happen when someone remembers to run them, which is usually after the document has gone out.
Route: /org/papers/designer/<type> → AI Config · Permission:
Create/edit/retire project types and template versions..
1. What it is
Four settings, and they are not equally real:
| Setting | Does | Actually used by |
|---|---|---|
| Drafting utility | The generator offered while authoring | The Drafting AI drawer |
| Auto-run analysis | Capabilities that run on submit, round and finalize | The lifecycle engine |
| Redline playbook (JSON) | Fallback positions and preferred language | Only the Compromise drafter |
| Extraction schema (JSON) | Intended to shape term extraction | Nothing at present — it is saved but not used. Use smart fields to shape extraction |
Redline grade grades against each clause's fallback positions in the Clause library. The playbook JSON here only seeds the Compromise drafter. If redline grading looks generic, add fallbacks to the clauses — not JSON to this tab.
AI configuration is saved on the type, with its own Save AI config button — the header's Save draft does not save it, and it does not need publishing.
2. Why you would use it
- Reviewers start from analysis, not a blank document. Risk analysis and missing clauses run on submit, so the approver opens a document with both already waiting.
- Every counterparty round is checked the same way. Version diff and shadow redline on each round catch the four-word indemnity change nobody reads for.
- Finished documents carry their own summary. Summarize and Extract key terms on finalize put a summary and seeded obligations on every executed agreement.
- Spend is predictable. Each event's selection is a known number of runs at a known cost per document.
3. How it spreads
| Event | Fires when | Results land in |
|---|---|---|
| On submit | A document is submitted | The document's AI Assistant, as cached results |
| On counterparty round | Each counterparty round is recorded | Same |
| On finalize | A document is finalized | Same — and Extract key terms seeds the obligation register |
Runs are best-effort: a failed capability is logged and never blocks the transition or the other capabilities.
4. Step by step
- Pick a Drafting utility, or leave No drafting assistant.
- In each event, pick the analyses to run.
- Leave the JSON boxes empty unless you use the Compromise drafter.
- Save AI config. Invalid JSON blocks the save with a message.
5. Field reference
Drafting utility
Draft Generator · Clause Suggester · Tone Rewrite · Translator · Annexure Generator — or none. The ? beside the picker opens the selected utility's guide.
Auto-run analysis
Nineteen capabilities can be selected on any event. Four cannot: Compromise draft, Fairness / balance, Compliance & policy check and Cross-document consistency.
| Group | Selectable | Sensible event |
|---|---|---|
| Analysis on any event | Summarize, Risk analysis, Extract key terms, Missing clauses, Blind spots, Entity & concept graph, Classify, Redaction suggestions, Retention advice, Approval routing, Cross-validate, Regulatory drift, Smart fields, Regional annexures | Submit or finalize |
| Needs a round or trail | Version diff, Redline grade, Shadow redline, Negotiation summary, Negotiation analytics | On counterparty round — they produce nothing useful on a first submit |
On a round event, Redline grade is fed the newest open counterparty redline automatically.
Rules that apply to every automatic run
| Rule | Effect |
|---|---|
| Credits | Each run costs the capability's normal credits — usually 450 |
| Monthly cap | papers.ai_monthly_credit_cap stops automatic and manual runs alike |
| Confidential documents | Automatic runs are blocked unless System Config papers.ai_confidential_cloud is allow. Silent — only the server log says so |
| Manual runs | Always available, whatever this tab says |
Read-only
Without Create/edit/retire project types and template versions. the tab says it is view-only and names the permission.
6. Worked example
A legal operations lead configures a Master Services Agreement type.
| Event | Selected | Credits |
|---|---|---|
| On submit | Risk analysis, Missing clauses | 900 |
| On counterparty round | Version diff, Shadow redline, Redline grade | 1,350 per round |
| On finalize | Summarize, Extract key terms, Entity & concept graph | 1,350 |
A typical MSA here goes through three counterparty rounds, so each agreement costs about 900 + 3 × 1,350 + 1,350 = 6,300 credits automatically — worth knowing before 200 of them a year run against the monthly cap.
They leave the Extraction schema empty: it would do nothing. They paste their liability and payment fallbacks into the Redline playbook, knowing it will only seed the Compromise drafter, and add the same fallbacks to the clauses in the library so Redline grade uses them.
One document is marked confidential. On submit, no analysis appears — the cloud AI gate blocked it, and the reviewer runs Risk analysis by hand, which is allowed.
7. The admin contract
| Must be true | Where | What breaks without it |
|---|---|---|
| The selected utilities are active and enabled | Orbit AI Flow, /org/ai-utilities | Those runs fail, silently |
papers.ai_monthly_credit_cap has room | System Config | Automatic runs stop mid-month |
papers.ai_confidential_cloud = allow, if confidential documents should auto-run | System Config | Nothing runs on them |
| The type's lifecycle records counterparty rounds | Lifecycle | Round events never fire |
| Save AI config was pressed | This tab | Nothing is saved — Save draft does not cover this tab |
8. Downstream
| When you… | Then |
|---|---|
| Save a new selection | It applies from the next event on every document of the type — no publish |
| Select Extract key terms on finalize | Obligations are written to the register on every finalize |
| Select Entity & concept graph | The knowledge graph's AI edges for each document are replaced on each run |
9. Don't confuse this with…
| How to run an AI action | Manual runs, always available |
| Orbit AI Flow | Where utilities and models are administered |
| Clause library | Where Redline grade's fallbacks actually live |
| Smart fields | The real way to define what gets extracted |
10. Troubleshooting
| Symptom | Cause |
|---|---|
| An expected analysis never appears | The utility is disabled, the cap is reached, or the document is confidential |
| My extraction schema changed nothing | The schema is not currently used — define smart fields instead |
| Redline grading ignores my playbook | It uses clause-library fallbacks, not this JSON |
| Changes did not save | Save AI config is separate from Save draft |
| Round analyses never run | The lifecycle does not record counterparty rounds |
| A capability is missing from the picker | It is one of the four that cannot auto-run |