Self-hosted vs cloud AI for regulated work

The distinction everybody collapses

There are two separate concerns and they get treated as one, which is where bad decisions come from:

• Will the vendor train on our data? Enterprise agreements say no, and those commitments are real and contractual.
• Is the data disclosed to a third party at all? Yes, necessarily — it has to reach their servers to be processed.

For most businesses the first question is the one that matters and the answer is satisfactory. For some obligations the second is the binding one: privilege, certain data processing restrictions, and contracts that name permitted sub-processors. No training commitment changes the fact of disclosure.

Work out which question your obligation actually asks before evaluating anything. It is usually a short conversation with whoever advises you on compliance, and it eliminates most of the options immediately.

What a cloud vendor agreement typically does cover

• A contractual commitment not to train on your business data.
• A data processing agreement, and in healthcare a BAA where the vendor offers one.
• Encryption in transit and at rest, and stated retention windows.
• Certifications — SOC 2, ISO 27001 and similar — which are real assurance about their controls.
• Admin logging within their platform.

Worth saying: This is a genuinely strong package and it satisfies many regulated use cases. Dismissing cloud AI as inherently non-compliant is wrong, and a vendor who tells you otherwise is selling.

What it does not cover

• The disclosure itself. Where your obligation restricts who may hold the information, a processing agreement does not remove the processor.
• Terms changing. Today's policy is today's. Your obligation is for the life of the matter.
• An audit trail you control. Platform logs are the platform's, with the platform's retention.
• Jurisdiction. Where processing happens is sometimes the controlling question and is not always yours to choose.
• Subpoena exposure. A third party holding your client's information can be compelled to produce it.

What self-hosting actually changes

One thing, completely: there is no third party. No processing agreement to negotiate, no retention policy to monitor, no sub-processor list to review, no terms to re-read next quarter.

It does not grant you compliance. You still need access control, audit logging, retention rules, a risk assessment and a policy — and now you are responsible for operating them rather than inheriting them from a vendor with a security team. That is a real trade and anyone presenting self-hosting as the automatically compliant choice is skipping it.

What we build includes those technical controls as standard, and we state plainly which obligations remain yours. See AI for law firms and AI for medical and dental practices for how this lands in each.

How to decide

1. Name the obligation. Which rule, statute, or contract clause. Not "we are in healthcare" but the specific requirement.
2. Ask whether it permits a processor. This usually eliminates half the options immediately.
3. Separate your use cases. Internal drafting with no client information is not the same as document review on a live matter.
4. Write the routing rule down. Which work goes where. "Use your judgement" is not a control and will not survive a review.
5. Document the decision with your reasoning. The question you will eventually be asked is not whether you used AI but whether you thought about it.