How is Domtar using AI in its sawmills?

To grade lumber at its planer mills, according to its own account published in August 2025. Over two years, Domtar has used AI modelling and machine learning to improve product quality at seven mills it names; its fact sheet lists about 60 locations in all. Boards "are scanned by sensors and photographed from multiple angles". AI models check each board against each customer specification, flagging defects such as separations, discolouration or insect holes. The account is candid about what keeps it working. Technicians label defects to train the models, supervisors field-test each model regularly, and false positives are corrected with the scanner supplier. It gives no figures for accuracy or waste reduced.

What the usual answer says

The top results are trade and owner-site versions of Domtar's article. They report the scanning, and the human oversight too: labelling, field tests, supplier corrections. What they don't do is say what that work should leave behind in a mill's records.

What Domtar says

Domtar's article of 11 August 2025 describes the process. At the planer mill, the scanner "analyzes the wood's composition, grading it visually and identifying inconsistencies that are undetectable to the human eye". The task was once "labor-intensive", could be hazardous, and was "prone to human error". "Now, it takes seconds and is highly accurate", the company says.

Specifications drive the design. Programmers "create AI models to assess wood for each individual specification", at a level of detail the company says people could not reach. One mill "makes around 15 unique lumber products, and each one needs to meet customer specifications", the company's quality lead for its planer mills explains. The same quality lead sums up the benefit: "AI allows us to be more precise and consistent," and "there's a lot less waste".

The article then turns to the people. "Process engineers and technicians train the AI by tagging pixels of the defect", it explains. Supervisors "perform regular field tests on each AI model to ensure the machine's assessments are aligned with their specifications". Errors happen. Sometimes a model flags "an irregularity that is not actually a defect", marking a sound board as unfit. Then "mill employees work with the machine supplier to adjust the model, reducing waste".

The scale behind it

Domtar's 2026 fact sheet gives annual production capacity of about 2.6 billion board feet of lumber and other wood products, and more than 13,000 employees. Domtar is privately owned and files no public annual report. Its own articles are the most detailed public record of its technology; the 2025 sustainability document we could open doesn't mention AI.

What it means for other sawmills

In Domtar's set-up, each specification has its own model, trained on labelled examples of the defects that matter for it. Other systems use one model with grade rules per specification. Either way, in our view, a model trained on one wood supply can misjudge another, which is why regular field tests against experienced graders matter. And every false positive has a cost: a sound board downgraded or rejected.

That makes AI grading a data maintenance job. The useful records are the labelled defect images, the field test results per model, and the false positives with the grade and value they cost. A mill that keeps those can see which models are drifting and what that is worth, and can make a case to its supplier with evidence.

A worked check

Say a planer mill runs 15 specification models and field-tests each monthly against a grader on 100 boards. One model starts flagging 8% of boards the grader passes, up from 2%. The cost per shift is the extra 6% times the boards that model grades per shift, times the price gap between the grade the board deserved and the grade it got. With that number, the mill can prioritise that model for relabelling ahead of the others. The figures are illustrative.

What to watch

The useful next disclosure would be a figure: the share of boards graded without human review, the false positive rate, or the value recovered. A private company need not publish those. Its own mills will have them if the field tests are recorded as described.

When it doesn't apply

Mills that grade only commodity products, or grade by hand, face a different trade-off. Domtar's account describes planer mill grading, not sawing optimisation or log scanning. And as a private company, Domtar may use AI elsewhere without publishing it.

Quarri for sawmills is built around how a sawmill runs, from log intake to shipped order.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.