When only one person can maintain your AI tool

An employee builds an AI tool that turns approved supplier documents into a comparison for the purchasing team. Colleagues use it every week. Before the employee takes leave, someone asks for a short explanation and the shared folder. Both arrive. Whether another person can maintain the tool remains an open question.

This example is fictional. The application is working; the decision concerns who can take responsibility for its continued use. As the founder of aiomics, I consider that distinction relevant whenever a useful experiment becomes something colleagues depend on. The proposed handover below is an organisational method, with no claim that its effectiveness has been experimentally established.

Define what the successor will take over

Start with the work that must continue during the creator’s absence. For our purchasing example, that could mean producing the weekly comparison from an approved document collection, investigating incomplete results and proposing limited changes to the instructions. Replacing the model or connecting a new repository could remain outside that person’s authority.

Write down the accepted use, permitted maintenance actions, decisions requiring approval and the person who can give that approval. Include the time available for maintenance. A colleague who can run an existing task may still need engineering support for a change. A handover can therefore be valid for a narrow scope without transferring every technical responsibility.

Google’s 2016 practitioner book describes a progressive transfer of service responsibilities, supported by training and temporary assistance from the developers. This is operational experience, without a controlled comparison of handover methods. [1]

For an internal AI tool, I would translate that principle into a specific acceptance decision: which work can this person undertake, with which support, from which date? The creator and successor should both be able to explain the boundary.

Ask the successor to establish the current state

Give the successor an ordinary request with artificial supplier documents in an isolated test environment. Before producing a comparison, ask them to identify the approved instructions, connected document collection, available model identifier and expected review. They should also find an accepted example showing which information must survive into the output.

The purpose is to make discrepancies visible. Perhaps the shared folder contains an old instruction, while the creator has been using a corrected version elsewhere. Perhaps the comparison looks complete, but the accepted example shows that conditions attached to a specification must appear beside it. Record what the successor found and where it differs from the intended configuration.

Keep credentials in the organisation’s approved access system. The successor needs their own authorised access. A handover document can point to the access procedure without containing passwords or secret keys. If only the creator can approve that access, assign another authorised approver before concluding the handover.

Observe a decision that requires judgement

Next, introduce one announced exercise with an undisclosed detail: one artificial supplier document omits the condition under which a stated performance is achieved. Ask the successor to prepare the comparison and explain what can be passed on. Agree beforehand that unsupported conditions must remain unresolved and be referred to the responsible specialist.

Observe whether the person notices the missing information, checks the source and respects that boundary. If the AI supplies a plausible condition, the task remains to establish its basis. A polished comparison alone cannot show whether the successor understood the uncertainty.

Google also describes experienced staff observing a newcomer who takes the lead, while remaining available to help. That account supplies a practical reference for observing readiness. [2]

In the proposed exercise, record each intervention from the creator: locating a file, explaining a rule or correcting a judgement. Assistance is useful learning evidence. It also identifies which tasks the successor has not yet demonstrated without help. Repeat the affected task with a different artificial example after the missing explanation or access has been repaired.

Rehearse one permitted maintenance change

Ask the successor to make a small, approved change in the test copy: for example, require each extracted condition to retain its document reference. Preserve the previous instructions and compare the results on examples already agreed with the specialist. The successor should explain what changed, what remained acceptable and which result still needs review.

Then have them return the test copy to its earlier configuration and establish that the intended version is active. This examines a maintenance action within the agreed scope. It establishes no ability to restore an unavailable external model or recover the entire service. The existing continuity guide addresses those separate questions.

A successful rehearsal covers only the person, tasks and configuration observed. It does not establish clinical suitability, authorise new uses or guarantee future reliability. Where a platform makes a version unavailable or a change impossible to isolate, record the limitation and narrow the maintenance authority accordingly.

Make a scoped acceptance decision

Retain a short record: scope accepted, tasks observed, help required, unresolved dependencies, available support and the next reason to review the arrangement. Useful triggers include a change in responsibilities, the document collection or the tool’s permitted actions.

If essential work still depends on the creator, leaders have concrete options: fund further preparation, arrange qualified support or restrict the tool’s use during the absence. The effort belongs in the decision about whether to keep and expand the application. The handover is complete when the receiving person and the accountable leader understand what has been demonstrated, what support remains necessary and where their authority ends.