For assurance and audit firms

Work that used to be performed by people, in sequence, with reviewable artifacts, is now performed or assisted by systems that produce a conclusion and very little else, and you are still engaged to opine on it.

So you ask whether the control operated, and what comes back is a policy document, a management assertion, and a system log written by the process under examination.

All three are evidence of intention, while none of them is evidence of operation.

That leaves two bad options: accept the assertion and carry the exposure quietly, or refuse it and expand scope, extend fieldwork, and explain to the client why this year costs more without a finding to justify it.

Neither is a professional failure on your part, because this is a supply problem: the evidence you need is not being produced by anyone.

The instinct is to ask for deeper access to the client's environment, which helps at the margin and then stops helping.

The party that hosts the workload frequently also controls identity, holds the log, operates the orchestration layer, and carries commercial exposure in the outcome.

Each of those roles is legitimate on its own, but held together they mean the record you are testing is produced, interpreted, and retained by the party with an interest in how it reads.

You already apply this principle everywhere else in your practice, because you do not accept a financial statement audited by the company that wrote it.

The same logic applies one layer down, to the evidence itself.

We issue signed, tamper-evident records of individual governed decisions, where tamper-evident means later alteration is detectable by anyone holding the record.

Each one is issued at the moment of the decision rather than assembled during fieldwork, and its origin can be verified without relying on the word of the party being examined.

ISO/IEC 42001, the international management system standard for artificial intelligence, and the NIST AI Risk Management Framework, the United States voluntary risk framework, both turn on demonstrated operation rather than documented intention.

The EU AI Act tightens that further for high-risk uses, through record-keeping and human-oversight obligations.

Different vocabularies, one recurring demand: show that the control fired on this decision.

We are the evidence layer beneath that demand, not another framework competing for the same shelf.

We do not map your client to a standard and we do not tell you whether they meet one, because that is your work and your judgment.

We do not audit, certify, attest, opine, or issue findings, and we decline engagements that would put us in that position.

That is not modesty but the entire architecture, because an evidence layer that also rendered opinions would carry the same conflict we exist to remove.

Your reliance on it would then be worth exactly what your reliance on the client's own log is worth.

Your independence is unaffected because we never enter the assurance chain.

We produce artifacts, and what they mean is yours to determine.

Most start on a single engagement, usually the one where the client's AI adoption has outrun the evidence and the partner already knows the file is thin.

We work alongside the client rather than through your engagement, so the evidence exists independently of the audit and stays available to other parties the client has to satisfy.

Firms that see it work once tend to raise it early on the next planning conversation, because the alternative is discovering the gap in the field.