Every AI system has a second birthday nobody plans for: the day the people who built it stop being the people who run it. The original developers move on, the agency engagement ends, the champion inside the company changes roles — and the system keeps answering customers, making decisions, and spending money, now maintained by people who never saw it built.
This is where most AI systems quietly start to rot. Not because the model degrades, but because the knowledge around it was never written down. Why was that prompt phrased that way? Which of those three tools is load-bearing and which is a leftover experiment? What does the system do when the vendor changes the model — and who decided that was acceptable?
The fix is boring and cheap: treat handover as a deliverable, not a farewell meeting. Before the builders leave, the system needs three things. First, a runbook: what the system does, what it must never do, and the five most common failures with their fixes. Second, a decision log: the trade-offs that were made and why — the model choice, the data sources that were rejected, the guardrails that were argued over. Third, a named owner with the authority to say no to the next feature request.
What a good handover actually contains
A runbook is not an architecture diagram. It is the answer to "it broke at 2am, what do I do?" — the dashboards to check, the switches to flip, the vendor status pages to watch, and the exact point at which to escalate to a human. If the runbook is a 40-page PDF nobody reads, it is not a runbook; it is a souvenir.
The decision log matters more than most teams expect. AI systems are full of choices that look arbitrary six months later — a temperature setting, a retrieval cutoff, a refusal phrase. Without the reasoning recorded, the next maintainer "fixes" the thing that was deliberately that way, and a working system starts failing in ways nobody can explain.
The handover test
The simplest test costs an afternoon: hand the system to someone who did not build it and ask them to handle a realistic incident — a wrong answer reaching a customer, a sudden cost spike, a vendor outage — using only what is written down. Whatever they cannot do is the handover gap, and it is better found in a rehearsal than at 2am.
This is also the moment to check the contracts. Does the new owner have access to every account the system depends on — the model provider, the vector store, the monitoring? Access that lives in one person's inbox is not access; it is a single point of failure wearing a password.
We build AI systems with the handover in mind from the first week — documented decisions, a runbook written as we go, and an owner named before launch — because the systems that last are the ones someone can actually run. Read how we scope AI builds for the long term — the cheapest reliability work you will ever do is writing down what you already know.
Field note: On one handover, the incoming team discovered the agent's escalation rule lived only in a prompt comment the original developer had written and never mentioned. It worked perfectly — until a routine prompt tidy-up deleted the comment, and escalations silently stopped for three weeks. The rule now lives in code, in the runbook, and in the decision log. Three copies sounds like overkill until you have needed the second one.
Related service AI products



