A peer-reviewed review of 69 studies on AI in the upstream forest products supply chain found plenty of working models and very few in operation. Its authors list seven reasons. One is about the algorithms. The rest are about the data and who holds it.
Source: Subedi, Gautam and LeBel, Artificial intelligence in the forest products supply chain: current applications and open challenges, Forestry 99(2), published 27 February 2026, doi.org/10.1093/forestry/cpag010. Open access under CC BY 4.0. All quotations are from the paper. The grouping of its seven barriers into model, data and institutional is ours.
What the paper is
A team at Université Laval and the FORAC research consortium reviewed 69 documents on AI in the forest products supply chain, 55 of them original research articles, published between 2019 and 2025 with the largest number in 2024. The scope is fibre supply, forest operations, storage and transportation. It stops at the mill gate. Transformation and sales are explicitly out of scope, which matters for what follows.
Most of the 69 are what you would expect AI in forestry to be. Convolutional networks classifying species and defects from images. Inventory from LiDAR and satellite. Reinforcement learning for spatial planning. Autonomous vehicles in mill yards. By the paper's account the models mostly perform, and the number of published studies has grown sharply since 2019, peaking in 2024.
Why so little of it is in operation
The authors list seven barriers. The second, verbatim: "Forest data are distributed across multiple stakeholders and are often stored in incompatible formats or lacking metadata." The third is inconsistent data quality, "collected under differing protocols and are often outdated". The fourth is that models trained on one site's data do not transfer to the next. The seventh is interpretability, deep learning "perceived as a 'black box approach'".
Read the list with the algorithms taken out. Cultural resistance, fragmented data, poor data quality, models that do not transfer, no computing infrastructure, regulatory rigidity, black boxes. One of the seven is a property of the models. Three are the data. Three are the organisations around it.
The paper's own summary of where progress is happening: "Sectors handling abundant, high-quality standardized data are advancing fastest towards real-world deployment."
The fix they propose is not an algorithm
The review offers two structural remedies. The first is the decision theatre, a shared space where the parties to a supply chain work on the same data at the same time, "where diverse datasets can be harmonized through a unified decision interface". The second is what the authors call a system integrator:
a neutral third-party organization responsible for coordinating collaborative, transparent, and data-driven planning among mills and agencies sharing a common procurement area
Its first job, in their description, is to act as "an informational hub integrating heterogeneous data streams". Both remedies are data-integration institutions. Neither is a model. A review whose subject was machine learning ended by recommending a data layer with a governance structure around it.
The example that carries the argument
The one application the paper puts closest to operational use is harvester productivity forecasting. The reason it gives is StanForD, the standard format that harvester computers write to. One reviewed study trained on 2,794 machine hours from 22 harvesters, all in that format, and compared 17 algorithms. Where the data arrives standardised, the model is nearly ready.
Where it does not, the paper's phrase is "far from full operationalization", and it applies that to forest management, traceability and routing. Same families of algorithm. Different data.
Where it touches ground an operation already stands on
Three of the open challenges the paper names sit on data a mill or a forest manager already holds.
Log yard inventory. Open challenge S2. Buffer policies at roadside landings and terminal yards are "often static and rely heavily on manual estimates, which can result in overstocking, material degradation, or supply interruptions". Vision systems can now count the pile. Their integration into a responsive inventory system, in the paper's words, "remains limited". Joining the count to the harvest schedule, the haul plan and the mill's consumption is a data problem rather than a vision problem.
Price as a model input. The reviewed price-forecasting studies work on market price series for timber and lumber. What a price move does to one operation's margin depends on that operation's own sales ledger, which sits on the far side of the mill gate and outside the paper's scope. Forecasting the market and knowing what the market did to you are different jobs.
Contracts. The authors note that machine learning and natural language processing "can aid in automating cost estimation, and contract management", and in the same paragraph that wood procurement "is not a purely economic optimization problem" because long-standing relationships carry weight. Reading the contract and setting its terms beside the price series is the part a machine does well. Deciding is still the buyer's.
What we cannot take from it
The scope ends at the mill gate, and most of the operations we work with sit on both sides of it. The 69 documents are research, not adoption figures. Nobody counted how many mills actually run any of this. And we read the review, not its 69 sources.
For an operation weighing where the next dollar of technology spend goes, the paper's own ordering is worth taking seriously. The models exist. The data they would run on, in most operations, still does not sit in one place. That is the part we build.