Transparent to clients: consent based on clear information

The criterion has three parts, and MSPs commonly satisfy the first while ignoring the other two.

• Which actions are AI-driven. Not the fact that AI is used somewhere, but which categories of work an agent performs, and which of those it performs unattended.
• Which are human-reviewed. The client should be able to tell the difference between an agent that drafts for approval and an agent that executes against their environment.
• What data goes to which models. Which providers process client data, what categories of data reach them, whether that data is retained or used for training, and in which jurisdiction the processing happens.

Timing and placement are part of the criterion rather than incidental to it. Disclosure happens at onboarding, in a form the client reads, and it is updated as tooling changes. A clause added to a master services agreement two years ago, in a document signed once and never re-read, is not what this criterion describes. Nor is a change of model provider that nobody was told about.

For MSPs with clients or operations in the European Union, the AI Act's obligations for general-purpose AI have been phasing in since 2024, with further application dates through 2026, and they establish a baseline for disclosure that did not previously exist in contract law. MSPs serving regulated clients in other markets face the same pressure through their clients' own obligations rather than directly: a bank or a hospital that must document its processors will ask its MSP for exactly this information whether or not the MSP's own jurisdiction requires it.

The index does not assess legal compliance and is not qualified to. What we take from the regulatory direction is simpler: disclosure is becoming a procurement question, and an MSP that cannot answer it in writing will lose deals to one that can. Any MSP relying on this for its own obligations should take its own advice on what applies to it.

It would be easy to read transparency as a governance nicety attached to a list of operational capabilities. It is in the list because it is the criterion that decides whether the rest of the model survives contact with a client's risk function.

There is a commercial reading as well. The argument in Automation Arbitrage is that the durable question for an AI-native MSP is what remains inside its value wedge once the automation is purchasable by anyone. Being the party that can account for how AI touches a client's estate, in writing, is one of the few positions in that wedge that a platform vendor selling to the client directly cannot easily occupy.

The failure mode worth naming is the opposite of secrecy. Over-disclosure, in the form of a fourteen-page appendix nobody reads, satisfies a lawyer and fails the criterion, because the criterion is about the client understanding what happens, not about the MSP having said it somewhere.

We ask to see the artefact the client actually receives, redacted as needed. We look for the three parts above in language a non-technical client can act on, and for a stated mechanism that triggers an update when the tooling changes.

We also ask what a client can decline. Meaningful consent implies at least one thing the client can say no to, whether that is autonomous execution against production, a particular model provider, or the processing of a sensitive data category. Where nothing can be declined, the disclosure is a notification, and we record it as such.