What you owe a client is not the automation itself but everything that lets them cope when it fails and you are not there.
- Write a plain handover document. What it does, what it depends on, what it costs to run, how to turn it off, and what wrong output looks like.
- Disclose AI use in writing. In most cases the law does not require it, but the alternative is the client finding out from an output.
- Plan for silent failure. Automations tend to keep running after something they depend on changes, so give the client a check they can see.
- Agree the limits before you start. Say which parts you did not verify, and put what you will not do, such as indefinite free maintenance, in writing.
You automated something for a client. It works. What you owe them is not the automation — it is everything that lets them survive it failing when you are not there.
You know what it does, what it assumes, and where it is fragile. The client knows none of that, and will discover it at the worst possible moment. Everything below is about closing that gap before it opens.
What the handover document contains
- What it does, in one paragraph, in their language. Not the architecture — the outcome.
- What it assumes. Every dependency: a file in a location, a field name, a page layout, an API. These are what break, and they break silently.
- What it costs to run. Monthly, with the API line separate. See the repo analysis — the API meter comes out of the margin, so show it to the client rather than leave them to find it.
- How to turn it off. One instruction, written for someone who is panicking.
- What "wrong" looks like. The output when it is working, so they can recognise when it is not.
- Who to contact, and what you will and will not do.
Disclosure
If AI is in the pipeline, say so in writing — not because you must in most cases, but because the alternative is them finding out from an output.
Under the EU AI Act, Article 50(4) requires deployers to disclose AI-generated text published “with the purpose of informing the public on matters of public interest” — and not where the text has had human review or editorial control and someone holds editorial responsibility for it. It applies from 2 August 2026.Regulation (EU) 2024/1689, Article 50, read at source 17 Sep 2026: “Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated.” The obligation does not apply “where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.” · Commission Guidelines on transparency obligations, adopted 20 Jul 2026, read at source 11 Sep 2026 — the obligations began to apply on 2 August 2026, and outputs generated before that date need not be marked retroactively. If your client publishes unreviewed output on matters of public interest, that obligation is theirs — and they need to know it exists.
The clause that matters most
Automations do not usually stop. They keep running against something that changed, producing confident output that is now wrong — a field renamed, a layout altered, a source moved.
A chain that fails loudly costs an afternoon. A chain that fails quietly costs whatever it produced in the meantime, multiplied by how long nobody noticed.
So the handover needs one thing beyond documentation: an assertion the client can see. A count that should be roughly twenty. A file that should appear daily. Something whose absence is visible without reading a log. Automations covers the design side.
What you owe, and what you do not
- You owe: the documentation above, a working off switch, honest cost figures, and disclosure.
- You owe: telling them which parts you did not verify. If you did not read the generated code, say so — published vulnerability rates in AI-generated code run roughly 40% to 70%.Pearce et al., Asleep at the Keyboard?, 2022: “we found approximately 40% to be vulnerable”. Veracode, 2025 GenAI Code Security Report: “45% of code samples failed security tests”, with Java at “a 72% security failure rate across tasks”. Both as read at source for Auditing AI Work, 11 Sep 2026.
- You do not owe: indefinite free maintenance, or liability for a dependency the client changed.
- Both of those need to be in writing before you start, not negotiated after something breaks.