Most machine-learning tools sold into a plant come with a project attached: a data scientist to build the model, a notebook to tune it, a few months of back-and-forth before the first useful output. Spall’s AI has no setup screen because it needs none, but “no setup” doesn’t mean “no decisions.” It means the decisions that matter got moved from a modeling exercise to two questions a plant manager already knows how to answer.
What Spall configures on its own
Nobody picks a model type. Every signal a machine reports gets matched to the model built for that kind of signal, a run/stop flag becomes an availability model, a part counter becomes a rate model, by a rule that runs the moment enough history exists, not a person choosing from a dropdown. The choice is recorded, not guessed: every fitted model carries the reason it was selected, visible on the machine’s own page, so the selection is inspectable rather than a black box even though no human made it.
Nobody sets a training schedule. Every model retrains automatically on an hourly cycle and scores new readings every five minutes, for as long as the machine keeps reporting. Nobody decides which signals are worth watching, either. A numeric reading that clears the sample floor gets its own drift detector without being asked for. A downtime reason that recurs enough times becomes a risk estimate. A cycle history long enough to have a fastest sustained stretch gets a golden-run reference. None of that waits on a configuration step, it happens because the data exists.
What a person still has to decide
A cost rate. A dollar figure only means something when it’s priced against a real rate: dollars per hour for a stopped machine, dollars per part for a scrap loss. Spall resolves that rate from whatever’s configured, a bill-of-materials figure, a level set higher up the hierarchy, or the platform’s own default, and states which one it used right next to the number. A rate nobody has ever set anywhere still produces a labeled default, not a blank, but a more specific rate always beats a generic one, and only a person on the floor knows what a stopped press actually costs this plant.
A reason code. Downtime risk, precursor patterns, and the failure-risk board all depend on stops being coded with an actual reason instead of left blank. Spall doesn’t invent reason codes or guess at them from context, an operator or a supervisor assigns them, the same way they always have. The AI’s job starts once that coding exists, not before it.
Whether AI answers are turned on. Ask your factory and the automated explanation layer are gated per tenant. Turning that on, and who gets to see it, is an account-level decision, not something Spall assumes. Every other part of the platform, the models, the insights, the dollar figures, works the same with or without it.
Which finding to act on. An insight is a comparison, not an instruction. Spall doesn’t decide that a drifting sensor means a bearing needs replacing, or that a recurring downtime reason is worth a maintenance ticket this week instead of next month. A person reads the evidence and makes that call, the same call they’d make reading a chart, just with the chart already built and the pattern already found.
Which proposed rule to create. A precursor pattern offers a one-click rule, a threshold or an alarm-repeat trigger, prefilled from what the data showed. Nothing gets created automatically. Someone still reviews the offer and decides it’s worth wiring up as a live alert.
An example, from a real first week
A plant connects a stamping press with nothing beyond the standard gateway setup: a run signal and a part counter, nothing else configured. Within a day, an availability model and a production-rate model exist for it, both reported against the platform’s default cost rate, since nothing more specific has been set. A week in, a supervisor starts coding stops with real reasons instead of leaving most of them blank. Downtime-risk and precursor models start becoming possible for the first time, since both need actual reason-coded stops to learn from, not just raw downtime minutes. A month in, someone sets the press’s real cost rate from what it actually runs, forty five dollars an hour of parts against the assembly line downstream. Every existing insight that used to carry a default-rate figure now prices against the real one, without anyone touching the models themselves, since the rate is resolved fresh every time a number gets shown, not baked in at fit time.
Nothing in that sequence required a project plan. Each step is one decision, made by someone who already knew the answer, layered on top of a system that was already working before any of them happened.
Why the split lands where it does
The line between these two lists isn’t arbitrary. Everything on the automatic side is a question the data itself can answer: which model fits this shape of signal, how many samples is enough, whether a pattern clears a real statistical bar. Everything on the decided side is a question only a person with context about this specific plant can answer: what a stopped machine actually costs here, what a stop was actually caused by, whether a finding is worth acting on today. Spall’s job is to do all of the first kind automatically and surface the second kind clearly enough that deciding it takes a minute, not a meeting.
That split is also why more configuration never gates entry. A machine with nothing set beyond a run-signal connection still gets an availability model, still gets a golden-run reference once it has enough cycles, still gets an Ask your factory answer about its uptime. Adding a cost rate, coding stops with real reasons, wiring up an alarm code as its own signal, each one buys a more precise number or a new kind of finding on top of what already works. None of it is required before the AI does anything useful at all.
Quick recap
- Model selection, retraining, signal discovery, and feature exclusion all run automatically, with the reasoning behind each fit visible, not a black box.
- A person still sets the cost rate a dollar figure prices against, and Spall states which rate it used, configured, resolved, or default, right next to the number.
- Reason codes come from operators and supervisors, the same as always. The AI’s downtime, precursor, and failure-risk models depend on that coding existing.
- Whether AI answers are enabled at all is an account-level decision. Everything else on the platform works the same either way.
- Acting on a finding, and choosing whether to wire up a proposed rule, stays a person’s call. The evidence is built. The decision isn’t automated.