Skip to content
Multigrid

Privacy, compliance & data residency

Where a prompt physically goes, who processes it on the way, what is kept and for how long — and which of those are decisions you control.

Almost everything written about AI and privacy is a summary of the law with no engineering in it. That is the wrong half. A regulator’s question and an engineer’s question turn out to be the same question asked in different vocabulary: what happens to this record. Answer it as a data flow — the hop, the processor, the retention window, the sub-processor behind the processor — and most of the compliance work becomes filling in a table you should have had anyway.

So these pages start from the request. They inventory it, follow it, redact it, log it and delete it, and they treat the legal text as the specification that says which of those you must be able to prove. Where a fact about a specific vendor is needed, they hand you the question to ask rather than an answer that would be stale by the time you read it.

None of this is legal advice. It is the engineering half, written so that the person who does give you legal advice has something concrete to read.

Does Your AI Provider Train on Your Data?

The question is really four questions against a stack of documents. Here is which document decides, and the clause types that hold the answer.

5 min read

Zero Data Retention: What It Actually Means

Retention is not one setting. It is six points in a request path where a copy can exist, and ZDR usually covers three of them.

5 min read

GDPR and AI APIs: A Practical Compliance Guide

Roles, lawful basis, transfers and records — worked through as properties of a request path rather than as a summary of the regulation.

6 min read

Data Residency: Keeping Inference in the EU

Residency is four separate guarantees that get sold as one. Here is how they come apart, and what to verify for each.

5 min read

DPAs and Sub-Processors for AI Vendors

A review checklist for an AI vendor's processing agreement, plus the sub-processor questions that are specific to brokered inference.

5 min read

The Right to Erasure vs a Trained Model

Deleting a record is solved. Deleting an influence on model weights is not. The engineering answer is to keep the two problems apart by design.

5 min read

PII Detection Before It Reaches a Provider

Why a bare pattern match over-fires, how check digits turn a guess into a verified match, and where detection stops working entirely.

5 min read

Redaction and Pseudonymisation in an AI Pipeline

Reversible pseudonymisation that survives the round trip: swap identifiers for stable tokens, let the model reason over them, restore on the way out.

5 min read

The EU AI Act: What Actually Applies to Your App

A decision path through role, risk tier and general-purpose model obligations, for ordinary software that calls a model API.

5 min read

AI Transparency Obligations and User Disclosure

Four triggers create a duty to tell someone AI was involved. Map them onto your product surfaces and most of the question answers itself.

5 min read

SOC 2 for an AI Product

What changes about scope, system boundary and evidence when the product calls a model — and which artefacts you will need that ordinary SaaS does not.

5 min read

HIPAA and LLMs: What Is Actually Possible

There is no such thing as a compliant model. There are business associate agreements, de-identification, and a set of questions to put to a vendor.

5 min read

Financial Services and AI: Regulatory Constraints

Model risk management was built for statistical models and mostly transfers to language models — except for the three places it does not.

5 min read

Confidential Computing for Inference

What a hardware attestation actually proves about a running inference workload, what it leaves open, and the questions that tell the two apart.

5 min read

On-Premise AI for Regulated Industries

A cost model you fill in with your own quotes, and an honest account of which compliance questions self-hosting closes and which it merely relocates.

5 min read

Designing an AI Audit Trail That Holds Up

An append-only, hash-chained record of what the system decided and why — with the fields that matter and the ones that must never be in it.

5 min read

Copyright and AI Output: Who Owns It

Ownership, protectability and infringement risk are three separate questions. Most confusion comes from answering one and thinking you answered all three.

5 min read

Training Data Copyright: The Shape of the Dispute

What is actually being argued about training data, the legal hooks each argument hangs on, and how to track it yourself instead of trusting a summary.

5 min read

Terms of Service Traps in AI APIs

Five clause families that are harmless at prototype scale and expensive in production, and what to negotiate or engineer around for each.

5 min read

Building an AI Usage Policy for Your Company

A short internal policy people will actually follow, with the reasoning behind each rule so you can adapt rather than copy.

6 min read

Other topics

Privacy, compliance & data residency · Multigrid