The AI proposes and explains. Code and people decide.
Vestoris uses large language models where they are strong (reading requests, explaining numbers, drafting text, connecting signals) and keeps every decision that matters in deterministic code, behind permissions and human approvals.
| Where | What the AI does | What decides |
|---|---|---|
| Rebalancing | Writes the reason for a proposal from the optimizer's numbers. | The optimizer (drift, tax lots, costs, limits); a person accepts. |
| What-ifs | Turns a sentence into structured optimizer parameters. | The parameters are bounded and shown for confirmation; the optimizer runs. |
| Intent pipeline | Interprets "Rebalance growth accounts" into a structured intent. | Entitlements, optimizer and compliance engine, stage by stage; orders on your click. |
| Pre-trade checks | Explains a failure and proposes a passing alternative. | The compliance engine, which re-checks the alternative before it is shown. |
| Rules | Drafts a structured rule from plain English. | A Compliance user who did not draft it approves it; until then it never applies. |
| Execution | Suggests TWAP or VWAP and a slice count. | The trader confirms; the simulator executes; compliance is re-checked first. |
| Surveillance | Writes the audit narrative of an alert. | Deterministic detectors raise the alerts; a person reviews them. |
| Access reviews | Drafts the review and its recommendations. | An Admin applies what they agree with and signs off. |
| Copilot | Answers questions and may propose a request for the pipeline. | Its context holds only data the user is entitled to see; it cannot act. |
Four hard requirements, each covered by automated tests
No bypassing a hard limit
AI output is only ever a proposal or text. Submitting, approving and executing an order each re-check entitlements and compliance in code. An AI-drafted rule is stored as a draft whatever the model says, and a copilot action can only start the pipeline.
Humans approve above thresholds
Orders above the firm's order-value or percent-of-account thresholds, and orders that breach a "warn" rule, wait for a person with "approve" on every account, who is not the person who entered the order. Every AI-drafted rule needs a Compliance user.
Every AI call on the record
Each call is stored with its purpose, full input and output (encrypted at rest, with a SHA-256 of each), model, fallback, latency, tokens and user, for the firm's chosen retention period. That supports SEC/FINRA-style recordkeeping if the platform serves clients.
Works without the AI
If the AI is unavailable, switched off for the firm, or the demo's daily allowance is used, every feature falls back to deterministic code: templates for explanations and drafts, a built-in parser for common requests, the engine's own passing quantity.
What the model receives
- Only the data a feature needs, from accounts the user may see.
- No passwords, keys or session data, ever.
- No other firm's data: every query is scoped to one firm.
- Calls go to the model provider's API only; the provider's commercial terms apply to the data sent.
Models and availability
- Claude models from Anthropic, in a fallback chain: when one model is unavailable the next is tried, and the one that failed is skipped for ten minutes.
- The model is told it can only propose or explain, and its answer is parsed and validated like any untrusted input.
- A firm can switch AI assistance off entirely.
See it with your own eyes, in ten minutes.
Open a private sandbox firm with sample accounts, rules and orders. No sign-up, no email. Switch roles and follow the guided tour.