In Sinfonia, an AI agent prepares its proposals on a working copy of the real data. A person reviews the impact, then applies or discards. Agents follow the same permissions as screens and the API, and the server checks every rule, whatever the path.
Agents that notice and build
A field added as a comment, a task repeated every Monday, a discount granted off the price list: the Sinfonia assistant notices these gaps between how teams work and what the software does. It proposes the change, then builds it: reference data, rules, menus, dashboards, workflows.
A change takes hours, not months. The software keeps pace with the business instead of holding it back.
Every proposal on a working copy
Example: the 2027 price list. The agent prepares it on a copy of the real data. The sales director sees the effect on open quotes, field by field, before and after. They apply the price list in one go, or discard it without a trace.
All or nothing: the server checks every rule before writing. A record changed in the meantime stops the update until it is reconciled. Critical changes require a typed confirmation.
Your AI assistants, plugged into your rules
Your teams query Sinfonia from the AI assistant they already use, through the MCP protocol, as an option. Three scopes to choose from: pseudonymised reading, plain reading, preparing working copies.
An external agent follows the same path as the built-in assistant: it prepares, a person applies.
The same guards on every path
Roles and permissions apply per person, in the same way to screens, the API, the AI assistant and the MCP server. The server enforces every guard on every path: screen, AI, scripts, API.
A refused action appears greyed out, with its reason. For an agent as for a person.