The owner summary

Going live is where the running cost starts. Name one person to check outputs on a rhythm that matches the risk, record what people change and why, and agree in advance what would make you pause. Consequential bookkeeping or financial changes need a person's approval before they take effect, because sampling does not replace that approval. If nobody can do those things, the workflow is not ready to run unattended.

The demo covers the easy part

A demonstration shows the first afternoon: clean inputs, a tidy result and an owner who is pleased with it. Live work is messier. Customers write in new ways, staff meet cases nobody planned for, and the tool you bought can behave differently in situations the trial never tested. The Australian Government's guidance on AI adoption makes this point directly. It says risks often emerge from how an AI system behaves in different situations and use cases, "rather than only from software updates".

Imagine a hypothetical bookkeeping office that uses an AI assistant to sort incoming supplier invoices into categories. The trial went well. Six months later a supplier starts sending scanned invoices with handwritten notes, and the assistant misfiles some of them. Nobody notices quickly, because the person who used to read every invoice now trusts the sorting. The cost shows up as reconciliation work several months later, when someone finally asks why the totals do not match.

The example is hypothetical, not a client result. It shows a pattern worth planning for.

What keeps appearing after launch

Four kinds of work tend to continue once a workflow is live. Each one needs an owner.

Checking time. Someone still reviews outputs. The first week may tell you little about later months, so plan the check as an ongoing task with a time allowance, not as a launch activity.

Correction records. The Centre's workflow guidance recommends capturing changes: record which outputs people changed and why. That record can work as an early signal. A rising number of corrections in one category may point to a problem, even when no single error looks serious.

Product change. The Office of the Australian Information Commissioner says due diligence for AI products should not amount to a "set and forget" approach. It recommends regular reviews of the product's performance, staff training and monitoring throughout the product's life. Plan reviews with the possibility of change in mind, including model, product and terms changes.

Exceptions. Some cases the tool cannot handle will reach a person. Someone must own those cases, have the authority to pause the workflow and be able to do it without asking permission first.

Write the running rules before go-live

Write these rules on one page before the workflow is switched on. Keep it short enough that someone will read it in a year.

  • Name one accountable owner and one backup. The essential practices for AI adoption ask organisations first to decide who is accountable for the outcome.
  • Choose the oversight pattern. A person can approve each output before it is used, which suits high-impact work, or a person can watch the work as it happens and step in when needed. The Centre's redesign guidance describes both patterns.
  • Set a sampling rhythm suited to the risk. Lower-risk outputs can be sampled on a schedule and reviewed against the original requirement rather than against how polished the draft looks. Consequential bookkeeping or financial changes need a person's approval before they take effect. Sampling does not replace that approval, and no fixed sample size is suitable for every workflow.
  • Keep a correction log. Record the output, what changed, why it changed and whether the case was new.
  • Agree the pause rule in advance. Write down which error or pattern stops the workflow and who can switch it off.
  • Schedule a review when anything material changes. That includes the types of input, the product's terms or features, the people involved and the rules that apply to the information.

If the rules feel heavy for a single small task, that is a useful finding. It may mean the task is not yet a good candidate for a running AI workflow.

What to count each week

The running cost includes the time and risk that stay with the business, beyond the subscription line. Three simple counts will show whether the saving is still there.

First, minutes spent checking. Second, corrections logged in the period, grouped by type. Third, escalations: cases the tool could not handle, and any complaint or error you can trace back to the workflow. A simple weekly line for each count is one practical starting point. If checking time rises while corrections fall, the workflow may be asking more of people than it saves. If corrections rise, possible causes include a change in inputs, a change in the product or model, or a process change that the rules did not anticipate.

These counts also give you the evidence for a decision. Continue when, for the risk involved, the checks are manageable, the corrections are understood and the owner still signs off the output. Stop or redesign when the corrections are rising, the sample keeps failing or nobody has the time left to review properly.

Where this fits

This article covers the period after a decision to adopt. Two earlier Insights cover the steps before it. Before you buy another AI tool sets out the full cost and the stop condition before a subscription starts. Did your AI pilot create value, or just move the work? describes how to judge a trial. Read them in that order if you are still deciding. If a trial has already passed, use the running rules above before the workflow becomes part of daily work.

Sources and limitations

The running rules, the weekly counts and the hypothetical example are NAVE's editorial synthesis. The example is not a client result. This is general business guidance. Suitable controls and privacy obligations depend on the business and on the use.

A relevant next step

Compare the running and review costs of several possible AI uses before you choose one. NAVE's AI Opportunity Audit is designed to help owners weigh rough value cases, information needs and risks before choosing a roadmap. It does not promise a result.

Explore the AI Opportunity Audit

A free 30-minute fit call is available if you want to talk the question through. It is optional. You do not need it to use this article.