Platform · Observability & Audit
Every agent action, on the record.
Full tracing of every AI call and every tool an agent touches — who, what, which data, at what cost. Tamper-evident retention, SIEM-ready exports, discovery of the AI you don't yet govern, and evidence packs when the auditor asks.
The question "what did the AI do?" has an answer.
Illustrative log — field names and detail depth follow your retention and privacy configuration. Note the third row: the trail records what was refused, not only what ran.
Capabilities in depth
Four layers of record, one system.
Request tracing
Every model call carries model, tokens, latency, and cost, attributed to user, team, tool, and agent. Filter by any dimension; follow a single conversation or a whole department's month. This is the layer finance uses for chargeback and engineering uses for debugging.
Action audit
Above the calls sit the actions: which records an agent read, what it changed, what it was refused. Entries are tamper-evident and retained on your schedule — the layer security and compliance live in.
Policy events
Masks applied, routes diverted to local models, budget caps hit, requests blocked — the control system's own activity is logged, so you can show not just rules on paper but rules firing in production.
Shadow AI discovery
Discovery surfaces the AI running outside the governed path — unmanaged tools, personal keys, stray OAuth grants, agents nobody owns. The output is a migration list, not a blame list: the goal is to bring usage onto the path employees already prefer.
SIEM & API export
Stream events to the security stack you already run, or pull them via API into your own warehouse. Leapforce is a source of truth, not a silo — your existing dashboards and alerting keep working.
Evidence on demand
Logs, inventories, and access states export as evidence packs mapped to EU AI Act requests, ISO/IEC 42001, NIST AI RMF, and SOC 2. When the auditor asks, the answer is an export — not a quarter of screenshots.
Who reads the record
Five audiences, five very different queries.
Show every agent that touched customer records this week, and anything that was blocked.
Incident scoping becomes a filter, not a forensics project.
Produce the AI system inventory and usage records for the audit.
One export, mapped to the framework's language, generated from live state.
Attribute this month's AI spend to cost centers, and flag anything pacing over budget.
Chargeback comes from the same records as audit.
Which tools are actually used, by whom, and what did discovery find outside the path?
License decisions and migrations run on evidence.
Is the triage coworker actually helping — volume handled, escalation rate, cost per ticket?
Adoption is managed with numbers, not anecdotes.
Questions we hear
Observability & Audit — frequently asked.
That's your call, per data class. Metadata (who, what model, cost, actions) is always recorded; prompt and response bodies can be retained, truncated, redacted, or dropped by policy. Regulated teams often keep full bodies for sensitive workflows and metadata-only elsewhere.
Audit entries are written append-only and integrity-chained, so alteration or deletion is detectable. Retention windows are configurable to your policy, and legal holds can freeze specific ranges.
Observability is scoped to corporate AI activity, and visibility is role-based — a team lead sees team aggregates, not colleagues' prompt text, unless your policy says otherwise. Works councils and privacy teams can review the configuration; the goal is accountability, not surveillance.
Multiple signals: identity-provider app grants, expense and SaaS inventories, network indicators where you enable them, and what teams self-report once there's an approved path to migrate to. Manual surveys alone undercount — the combination is what works.
Yes — policy events (blocks, masks, budget thresholds, unusual volumes) can notify owners directly or flow to your SIEM where your existing alerting takes over.
Events export in standard structured formats over streaming and batch APIs, which the mainstream SIEM and logging platforms ingest. If you ingest JSON events, you can ingest Leapforce.
Embedded AI features that never touch the gateway are visible only through discovery signals, not full tracing — an industry-wide limit worth being honest about. The mitigation is routing what matters through governed connectors, where actions are fully recorded.
Customer data access is restricted, logged, and contractually bounded; enterprise cloud deployments keep the record entirely inside your infrastructure. Details are covered in the trust review on a strategy call.
Leapforce is in active development — per-capability build status (live, in development, roadmap) is disclosed honestly on request. Industry figures above are drawn from public 2025–2026 research and cited ranges vary by study.