Definition
What artefacts should an AI handover include?
A complete AI handover has six artefacts: the working system, versioned prompts, eval set, frozen failure cases, decision record and runbook.
By the Foxnut team · Updated
What artefacts should a complete AI handover include?
AI handover artefacts are the things that must physically change hands for your team to run an AI system without the people who built it. Practitioners who actually hand these systems over converge on six: the working system in your own infrastructure, the prompts under version control, the eval set, the frozen failure cases, the decision record and the operating runbook.
Foxnut Studios is a senior-only practice in Bengaluru and Paris, and handing over is how our AI work ends. There’s a link to reach us at the bottom of the page. Let’s first make sure you know what to ask for.
The logic of the list fits in one sentence: an AI system can be broken without touching a line of code, so the handover must transfer the means of changing it safely along with the system. This page walks the list artefact by artefact: what each one is, the form it should arrive in, and the question that tests whether you actually received it. The position behind the list sits with the handover method at Foxnut Studios.
How do you test whether each artefact is real?
You test each artefact by what it lets your team do, not by whether it exists. A handoff is easy to claim and easy to fake, because a folder of documentation satisfies the word without transferring the capability. Each row of the table below ends in a question you can ask on the closing call. It can be answered in minutes if the artefact is real, and not at all if it is not.
| Artefact | The form it should arrive in | The question that tests it |
|---|---|---|
| The working system | Running in your accounts, your billing, your access controls | Can we revoke the builder’s access today and still operate? |
| The prompts, versioned | A repository with history, not a document with copies | Can we see what changed in production last month, and who changed it? |
| The eval set | Real cases with expected outcomes, runnable on demand | If we edit a prompt tomorrow, what tells us whether it got worse? |
| The frozen failure cases | The actual inputs that broke the system during the build, kept as a fixed set | What has already gone wrong, and would we catch it if it recurred? |
| The decision record | Written reasons for the choices that are not visible in the code | Why is this threshold set where it is, and what was rejected? |
| The operating runbook | Named owners, checks on a schedule, and steps for when a check fails | Who notices when it breaks silently, and how fast? |
Why do the other artefacts depend on the eval set?
They depend on it because without an eval set nobody can tell whether a change broke something. The strongest position among practitioners is that the eval set, more than the prompts, is the real deliverable, because prompts without that check are just text the next person is too scared to touch.
An eval set is a collection of real inputs with known-good outcomes that can be re-run after any change. Its value is what it converts: “did my edit make this worse?” stops being a feeling and becomes a checkable answer.
Without it, every other artefact decays. The versioned prompts become a history nobody dares extend, and the working system becomes untouchable the day the builders leave. A handoff that includes prompts but no evals has handed over the system’s text and kept its safety.
What are frozen failure cases?
Frozen failure cases are the actual inputs that broke the system during the build, kept as a fixed set. Every AI system that reaches production broke repeatedly on the way there, on specific ugly inputs: the malformed invoice, the customer message in two languages, the edge case nobody designed for.
Practitioners argue for handing over exactly those inputs, because they are the institutional memory of the build: where this system, on this data, actually fails. That memory otherwise walks out with the builder.
Confusion to avoid: the frozen cases do not duplicate the eval set. Evals encode what good looks like, and the failure archive encodes what bad has already looked like. A slide back toward an old failure is then caught by replaying history, and not rediscovered by a customer.
What is a decision record in an AI handover?
A decision record is a plain written trail of the choices and tradeoffs behind the system. An AI system’s code shows what was built, and it does not show why. Why retrieval is configured this way, why a confidence threshold sits at this value, which cheaper model was tried and rejected on quality: each of these was learned at some cost, and none of it is visible in the system itself.
Practitioners who have inherited these systems report caring less about the initial prompts over time and more about preserving the engineering decisions, because an undocumented decision gets re-litigated by the next person and re-broken the same way.
The record also has a second life. For a company operating AI in a regulated market, the record of how the system was built and why is the raw material every explanation starts from, which is why this page sits next to the EU AI Act read from the operator’s side.
What does the operating runbook do?
The operating runbook turns a build into an operation: named owners for each workflow, checks that run on a schedule and not on suspicion, and written steps for what to do when a check fails. A conventional application fails loudly, with errors, crashes and red tests. An AI workflow’s characteristic failure is silence: it keeps running and quietly produces worse output, or stops and nothing notices.
Practitioners tell a canonical story about this: an automation that stopped working, unnoticed for days, at the cost of the client relationship. The runbook is the least glamorous item on the list and the one whose absence is most expensive, because the other five artefacts describe the system and this one keeps it alive.
How do you check a handover in one conversation?
Ask the six questions in the table above as a closing checklist: revoke access, show the change history, run the evals, replay a failure, explain a threshold, name the owner. A vendor with a real handoff can answer all six in under an hour, on screen, using your accounts.
That is also the honest form of an AI readiness assessment for a team inheriting a system. Readiness here is not a maturity score. It is whether your team can change the system on Monday and know by Tuesday whether the change was safe.
What does Foxnut Studios hand over?
We hand over a running system, because we build AI systems into commercial operations and our stated deliverable is operational independence: a system your team runs without us. We do not present the category list above as a description of our own practice, because our practice follows the scope of each engagement.
Which artefacts ship on a given engagement follows that scope. A pure consulting engagement is worked with your team, scoped, structured and implemented together, and the artefacts live with that team from day one.
Where we also own implementation, we build end to end, including guiding you through buying and setting up every tool the system needs, and we run what we built under a BOOT model (build, own, operate, transfer) through the engagement and its support period.
Where we differ from the category list is the destination and not the checklist. By the end of the support period the handover is total, you run the system alone, and our continuing role is troubleshooting or short consulting on new requirements, on request and not on retainer.
One dated account of a full handover, told in the founder’s own words, names what actually changed hands: a system running on the client’s own servers and API keys, a troubleshooting framework, a maintenance cadence the client’s own team was trained to run, and training for everyone the system touches, not just the people who operate it day to day.
We are a one-tier practice, so the people on your closing call are the people who built the system, with nobody in between to go and check. We launched our own products before we advised anyone else on theirs. Your first step is a free audit with no pitch deck, and our consulting page says what else we do.
Run the six questions on whoever built your current AI vendor’s system. If any one of them stalls, tell us which one and we will tell you honestly how bad that is.
Foxnut Studios takes on work like this from Bengaluru and Paris. If you want to see how we do it before you talk to anyone, here is how an AI engagement is scoped and priced.