HAIEC demonstrates action-path assurance against a working multi-tenant AI SaaS

Multi-tenant AI services can connect users, agents, integrations, credentials and business workflows. These connections make them relevant validation environments for consequential action assurance. HAIEC’s September 2026 MunichTech evaluation used KestrelVoice, a working AI reception and website chat SaaS application, as its evaluated system.

The documented snapshot reports 44 action paths meeting their stated criteria, with 1,322 evidence traces and 555 verified checkpoints. It also preserves 99 unresolved facts instead of converting absent evidence into affirmative conclusions. This is a bounded engineering evaluation of a particular frozen system and control scope, not a claim that every possible action or runtime outcome was established.

The engineering value is the method: examine request and policy, verify access and source capabilities, separate possible actions from observations and preserve the exact evidence used for the determination. Enterprise buyers can use this pattern to scope an assurance proof of concept around a consequential workflow they actually need to govern.

Assurance diagrams are useful for explaining a concept, but a production-oriented software service contains the difficult interfaces that determine whether the concept has engineering value. A multi-tenant application has real handlers, role checks, tool paths, stored records and isolation boundaries. Examining those details exposes potential differences between stated permissions and implementation behavior.

HAIEC used a versioned KestrelVoice application snapshot as an assessment subject to evaluate selected action paths and evidence qualifications. This approach gives reviewers concrete source paths and a bounded set of questions rather than an abstract assertion that an AI agent is governed.

The method separates requested operations from policy authorization, effective grants, code reachability and observed evidence. A path that can be invoked is not automatically an authorized business choice, and a recorded invocation is not proof of a durable change at the destination. Treating these as separate dimensions helps engineers identify the specific missing link.

The documented award case reported 44 defined action paths, 1,322 evidence traces and 555 checkpoints. These numbers describe one versioned evaluation and its evidence model. They should not be read as universal coverage of every KestrelVoice account or future software version.

For an organization deploying consequential AI, the practical first step is to select one meaningful system and one important business action. Connect the relevant code and policy evidence, evaluate the qualified paths, and produce a reviewable record of established facts and unresolved questions.

This bounded workflow offers a disciplined way to expand assurance. Security, platform and governance teams can prioritize which additional authority records, telemetry and target-side verification would materially strengthen the next decision rather than treating incomplete evidence as a passing result.