Glacis joins PACT AI as a founding member.

AI is moving from answering questions to doing things: handling patient information, making recommendations, calling tools, approving transactions and acting inside businesses.

That changes the trust problem.

It is no longer enough for an organization to have an AI policy, say it has guardrails, or produce a report describing how the system is supposed to behave. When AI does something consequential, somebody outside the software stack needs to be able to establish what supervision was actually in place.

That is why Glacis has joined the Partnership for Assurance, Credibility, and Trust on AI — PACT AI — as a founding member.

An assurance market needs more than one kind of participant

PACT AI’s launch announcement brings together organizations deploying AI, technical assurance providers, AI insurers and civil-society voices. That mix matters. The people building assurance methods cannot define credibility alone; the organizations using AI, the institutions absorbing risk, and the people affected by these systems all have a stake in what counts as evidence.

PACT says its founding members spent 90 days shaping its governance structure, core principles and operating model. Its initial work is organized around three connected priorities:

1. 01Developing a policy platform for federal and state conversations.
2. 02Professionalizing AI assurance while preserving quality, rigor, independence and accountability.
3. 03Growing a market in which credible assurance is recognized and rewarded.

Those are PACT’s priorities, not claims that the work is finished. The assurance ecosystem is still young. Shared methods will need to prove useful across different systems, risks and sectors.

What Glacis brings to the work

We built Glacis around a simple problem: the controls surrounding AI are usually hidden inside the same systems claiming that those controls worked.

Our contribution is the operational evidence layer. For actions routed through configured paths, Glacis can apply rules and create a signed record of the reported decision — what was allowed, blocked or escalated, which control was involved, and enough information for another party to check supported properties of the record. A data-minimizing configuration can omit protected payloads from that record; wider deployment data flows remain separate.

That is one part of assurance, not all of it. Evaluation, governance, standards, domain expertise, legal accountability and insurance each answer different questions. The useful work is making those parts connect without collapsing independence between them.

From promises to evidence

PACT’s launch is a chance to make AI assurance legible outside the small group of people who already work in it.

For Glacis, the test is practical: can a customer, auditor, insurer or regulator inspect a scoped record of what configured supervision reported when an AI system acted? Can they check supported integrity and key-attribution properties without relying on the system’s own pass/fail assertion? Can that record travel between organizations with its trust assumptions intact?

Those questions are narrower than “can we trust AI?” They can be answered for the supported properties of a covered record; source truth, execution, effectiveness, timing, and completeness still require separate evidence. We are glad to work on them alongside PACT AI’s other participating organizations.