Security and trust
Governed by construction, not by policy.
aprèsvente is an agentic platform, so trust cannot be a promise in a PDF. It has to be structural. These six guarantees are enforced in code and re-proven by 122 tests across 26 suites on every commit.
01 · The boundaryAgents can never contact your customers
Agent tools can only stage drafts into an approval queue. The single outward dispatch path is a human-triggered, authenticated action. The rule is enforced three times: a permission gate on every tool call, a runtime hook in front of every tool, and a CI test that proves every registered tool passes through the gate.
02 · Citation honestyAn uncited number cannot ship
Every metric, chart, and table in every artifact must reference a gathered source. Saving a section with an unbacked number throws a CitationError in code. Model predictions enter the system as cited context like any other source, so they are covered by the same rule.
03 · Tenant isolationProven per route, on every commit
Every table carries an organization id. A route-manifest harness calls every endpoint as the wrong tenant and fails the build on any cross-organization read or write, including endpoints added after the harness was written.
04 · CredentialsEncrypted at rest, masked in flight
Integration credentials are encrypted with AES-256-GCM and masked at the API layer. Secrets never appear in artifacts, logs, or agent context.
05 · Autonomy under capsAutopilot is opt-in, capped, and killable
Autonomous runs are off by default and enabled per organization. When on, they run under daily caps, per-account cooldowns, a customer denylist, per-playbook allowlists, and a kill switch that stops everything at once.
06 · Honest metricsNo accuracy theater
Churn and expansion models expose every factor and weight behind a score, and the platform refuses to display accuracy claims until real outcomes exist to measure against. The same standard applies to this website: every number on it is counted in the codebase.