What should an audit recommend when your team shows that a planned AI project does not have the data it needs?
The recording
Still shown64 seconds in · untouchedThe audit simulator presents three automation projects and lets the CEO approve any of them. When predictive maintenance is selected, only 0.3 of its estimated 5 hours of weekly savings counts as achievable because the company has no recorded maintenance history for a model to learn from. See it full size.
What changed it
The plant supervisor reported that failures were fixed and never recorded, leaving no incident history for a model to learn from.
Interview finding · Plant supervisor · Step 03
The plant supervisor's evidence changes the recommendation. The CEO's doubts are not treated as evidence.
The artifact
DeployedThis walkthrough is public.
This walkthrough is public. Its first six steps present the audit evidence. In the seventh, you can select predictive maintenance yourself and see why the audit still recommends “not now.”
Biokehä Oy is a fictional company. The people, interviews, quotes, and figures in the walkthrough were created for this demonstration. The walkthrough needs no account and has no backend, so nothing you enter leaves your browser. Its final step opens a separate artifact hosted elsewhere.
The record
Four questionsWhat it used
The audit starts with public-record research and a kickoff brief, then tests those assumptions in four fictional interviews: the CEO, plant supervisor, reporting lead, and waste-reception lead. Each conclusion in the walkthrough names the stage it came from.
What it checked
The Explore step multiplies each weekly-hours estimate by its readiness. Full predictive maintenance starts at 5 hours a week but is only 5% ready, so the walkthrough displays 0.3 hours a week as currently realizable.
The hours and readiness figures are invented estimates for this fictional company. The walkthrough performs arithmetic on them. They are not measured savings, a forecast, or the output of an audit-scoring rubric.
Where it stopped
When full predictive maintenance is selected, the walkthrough responds: “Greenlighting this does not make it possible.” Its verdict is “not now (do not automate yet)” because no maintenance history exists for a model to learn from.
It proposes a lightweight maintenance incident log first, owned by the plant supervisor. Predictive maintenance can be reconsidered after roughly a year of incidents, parts, and downtime have been recorded. If that later work goes ahead, the audit assigns it to a specialist industrial machine-learning partner.
What this does not prove
No real company, client engagement, or interview evidence was used in this experiment. It does not show that the same conclusion would survive real interviews or operating data. The blueprint gives enough direction to accept or refuse the work. It is not a technical specification and cannot be built without further questions.
Reading
Author's noteWhat I take from it
A useful audit should be able to advise against the project that led the kickoff. Here, the recommendation changes when the plant supervisor reveals that the required history was never recorded. The audit shows the chain from the original idea to the missing evidence, then names a smaller first move and the right type of specialist for the later work.