The EU AI Act entered into force in August 2024 and its general application date passed in August 2026. For most people delivering custom software the Act does not regulate you directly, but it has changed what your buyers ask: how much of this build was AI, who checked the result, and can you evidence your answer. This guide explains what an AI provenance record is, what it should contain, and how to generate one for free in about two minutes.
Precision matters here, because most of what is written on this topic overstates it. The AI Act regulates AI systems by risk class: prohibited practices, high-risk systems with conformity obligations, transparency duties for certain AI interactions and AI-generated content, and documentation duties for general-purpose model providers. A web app, an API or an internal tool that you built with the help of a coding agent is usually not itself an AI system under the Act, and delivering one does not typically make you a provider of one.
What changed is the buyer’s side. Enterprises inside the Act’s reach are building AI governance programmes, and those programmes push documentation duties down the supply chain through procurement. Security questionnaires that asked about dependencies and data handling now ask whether AI was used in development, to what extent, under what oversight, and with what verification. A contract that once required a licence inventory now requires an account of AI involvement. The question arrives whether or not the Act formally applies to you, because your client’s compliance team does not price that distinction finely. Answering it with a shrug reads as risk; answering it with a signed-off paragraph of prose reads as an assurance nobody can check. The useful answer is a provenance record.
A provenance record for AI-assisted software states, from evidence rather than memory, how the software came to exist: which parts of the build have a recorded coding-agent trail, what the agents claimed and which of those claims held up, what a human confirmed on the record, what was independently executed and verified, and what remains open. The essential property is that every statement is checkable. A record that says “approximately 40% AI-generated” with no way to verify the number is marketing; a record that says “20 decisions recovered from coding-agent session history, each with source references, and 5 of 5 recorded checks independently re-executed” is evidence.
Honesty about direction matters too. Agent session history is a floor, not a ceiling: it proves AI involvement where it exists, and its absence proves nothing. A credible record says exactly that, because a buyer’s auditor will say it if the record does not.
In practice the AI question arrives in three shapes, and Mantyl’s export command produces all three from a generated project passport. The AI provenance record itself is a machine-readable JSON document with a human-readable rendering: AI involvement counted from session-evidenced claims and decisions, contradictions listed rather than hidden, human oversight and acceptance on record, verification results with commands and log digests, and the meaning of every evidence label defined in place. The CycloneDX 1.6 AIBOM carries the same trail in the SBOM format supply-chain tooling already ingests, down to a serial number derived from the passport digest so even the identifier is reproducible. And the procurement answer sheet fills in the questions security reviews ask in the order they ask them, each answer traceable back to the passport.
The whole flow runs locally, needs no account, and uploads nothing. From the project root:
Scan reads the repository and the local coding-agent history (Claude Code, plus Cursor and Codex in beta) through fail-closed secret redaction. Verify executes the project’s own install, test and lint steps in an isolated Docker sandbox, for Node and TypeScript projects and, in beta, Python. Generate assembles the digest-stamped passport, and export maps it into the three documents. Ship them beside the delivery; the recipient can independently recheck everything they summarise with one command.
It is not a conformity assessment, it is not a certification, and it is not legal advice, and Mantyl’s exports state this on their own face. If you are actually shipping an AI system into a high-risk category, you need counsel and a notified-body conversation, not a CLI. What the record does is narrower and more common: it answers the transparency question that now sits in ordinary software procurement, with evidence a reviewer can check instead of prose they must take on trust. A document that names its own limits is the one worth attaching, which is the same reason the passport labels a contradicted claim as contradicted instead of deleting it.
The evidence vocabulary behind every statement is documented in the truth model, the delivery practice the record belongs to is covered in AI code handover, and if the delivery deserves proof that rests on neither party, an independent signed verification is one credit, £29.