Matthew McLaughlin's estimate comes out to something specific once you run the numbers. Thirty-two weeks of review for one day of readings [4] works out to 224 analyst-days per day of data [16], which is another way of saying the mill was generating analytical work about 224 times faster than one engineer could clear it. Spread across the fleet, that is roughly twelve hours of human attention per sensor per day of output [17]. A gap of that shape closes when the question of who looks, and how often, gets answered differently, not when the model underneath gets better.
The capability Domtar needed was already on the contract. Waites' full-service tier included an analyst who could learn the facility, its pain points, its personnel and its staffing levels, then tune the sensors' recommendations to them [6]. McLaughlin's contribution was to spend an entitlement the mill already held, and the extra instrument feeds he routed into it came from equipment the plant already owned [8]. His stated logic was that machine learning teaches you nothing if you teach it nothing [20].
Pair the age ceiling with the meeting cadence and the design becomes legible. Thirty days is about four weekly cycles [18], so an item can be raised, missed, raised again and still close inside the window, and an item that expires has been visibly passed over four times. What Domtar built is administrative rather than mechanical: a clock on attention, not on machinery.
What the record grants is a timeline and one mechanism. McLaughlin said improvements became noticeable within three months of the more active approach [9], and the savings channel described in any detail is lubrication, where the analytics helped set a lower viscosity grade so motors use less horsepower and draw fewer amps [14]. Neither comes with a magnitude. That is enough to justify copying the method, and short of what a finance team would need to underwrite a payback case.
This is one mill, one engineer, and a vendor chief with an obvious interest in saying his product needs a devoted internal champion, or in Rob Ratterman's phrase, someone who becomes a lighthouse for the whole company to follow [11]. That framing holds up. The mechanism is less clean, though: the workflow Ratterman describes has the AI doing the primary analysis while diagnosis still runs through his analyst, on-site vibration specialists and Domtar's own maintenance staff [10]. McLaughlin's own account points the same way: he treated the installation as an advanced system that could not be administered like a legacy predictive maintenance program [19].
The trade-off comes down to this: sensors get added by purchase order, but triage capacity has to be added by roster, and those two things do not move on the same schedule. A plant that buys the first without staffing the second ends up with a longer queue and the same failures. This quarter the decision is whether to fund the standing hour and name its owner, and next quarter's test is whether the 30-day ceiling holds now that the engineer who set it has been promoted to reliability superintendent [1].