Lessons From Our First Full Commercial Packing Season

At the South Chungcheong pilot facility, the final peach run ended in late September after a 2024-2025 packing season built around three unanswered questions: could AI routing withstand a live facility, could coating supplies move from sample kits to full-season volumes, and would packing-house staff act on the model's recommendations? We finished with partial answers to all three. Some exceeded expectations. Others did not.

This account comes before we redesign anything for the 2025-2026 season. Recording the friction while it remains fresh is more useful than polishing the story afterward.

Where AI Routing Outperformed Our Expectations

Across the four pilot facilities, batch-prioritization accuracy was the clearest result. At three sites, the model's urgency rankings closely matched the high-risk crates experienced packers identified in post-shift reviews. At the South Chungcheong stone-fruit facility, we compared the daily model queue with a line supervisor's independent list, made before she saw the output. The lists overlapped substantially.

That match showed us the model was not offering a separate signal from experienced judgment. It was delivering a faster, more consistent version of much the same signal. During a 400-crate morning shift, triage takes the model under three minutes. Even a skilled supervisor cannot cover that many crates without sampling.

The AI remained effective mid-season, when mixed batches and alternating varieties made the work less uniform. Climacteric scores for peaches and nectarines stayed consistent across stone-fruit varieties, a result we had approached cautiously. Internal testing had weighted training data toward one dominant cultivar, yet the model generalized to related varieties on the sorting line better than pre-season validation suggested.

The Coating Logistics We Underestimated

The coating chemistry delivered the expected results. Treated and untreated comparison batches showed consistent transpiration reduction and firmness retention against our internal data. KFD compliance documents held up at every facility without incident, and retailers reported no coating-related quality complaints during the season.

Supply sequencing was the weak point. Packing seasons shift in ways that are hard to plan remotely: a cold snap can move harvest forward two weeks, while a dry spell can delay it. We scheduled coating shipments around historical harvest-window averages, which proved too blunt for real-season variation.

Three times, facilities called to report an early variety run and low coating stock. Expedited delivery solved each case, but the added cost and pressure should not be routine. During the July Jeju mandarin peak, two facilities called about stock levels within the same week.

We have moved to four weeks of buffer stock at each facility. Dashboard alerts now appear when stock drops below a two-week threshold calculated from recent throughput. The buffer uses some facility working capital, but removes the race against harvest, making the trade-off worthwhile.

Growers Responded With Careful Interest

During the pilot, our direct contacts were packing-house operators rather than growers. Growers still heard about the coating trials through their packers, and some brought questions of their own. They were good questions.

The common question concerned efficacy least of all. Growers quickly accepted that a shelf-life-extending coating was plausible. The idea is familiar, although our formulation differs from standard carnauba wax. Their question was: "Who told you this is food-safe for my export market?"

That is exactly the question to ask, and we want to state the limit of our answer clearly. For domestic Korean sales and most East Asian export markets, our ingredients are documented and cleared under applicable food standards. When the pilot ran, our regulatory mapping for this protein-wax hybrid formulation was unfinished for EU and US export. We said so directly instead of offering vague reassurance, and the documentation work continues. Growers with active EU or US export programs should ask about current status before using the system for those shipments.

Growers who stayed engaged through the pilot were not put off by that boundary. The specific answer reassured them. Several showed interest in the 2025-2026 cohort once documentation for their target export markets is confirmed.

The Fourth Facility: Failure and Its Cause

Three of our four pilot facilities generated satisfactory data. The fourth did not, and the reason matters.

That facility used an older sorting line with color-index sensors that produced inconsistent readings. The hardware had not been recalibrated for several seasons, and the resulting noise made confident model scores difficult. We missed the issue during onboarding because our pre-season sensor checks were not rigorous enough to catch gradual calibration drift.

Through the seasonal peak, the facility's batches received lower-confidence urgency scores than their condition justified. The line team sensibly relied more on visual inspection than on the model. That response was rational, but it left us with three functional pilot facilities rather than four.

The lesson was direct: routing performance depends on sensor data quality, and calibration cannot be assumed. We are adding a mandatory two-week validation period during onboarding. Data will be ingested, but treatment recommendations will wait until quality is confirmed within an acceptable range. The validation results will be visible to the facility, allowing both sides to address sensor problems before the season begins.

Changes for the 2025-2026 Season

Three operational changes:

Sensor validation gate at onboarding. As described above, no facility enters the active routing system without a two-week validation phase confirming acceptable data quality. Facilities with sensor calibration issues will be supported in fixing them before go-live, not during peak season.

Four-week coating buffer with automated alerts. All commercial partner facilities shift to a buffer stock model. Low-stock dashboard alerts trigger at a threshold calculated from the facility's own throughput rate over the prior four weeks, adjusted for the crop type's harvest window pattern.

Mid-season review calls. Last season we scheduled end-of-season debrief calls with each facility. The most useful feedback came from those calls, but it came too late to correct problems in-season. For 2025-2026, we are adding a mid-season call at the four-week mark, explicitly structured to identify operational friction before peak volume hits.

What Remains Uncertain

A season-end article could make the first commercial season sound like a tidy proof of concept. Ours was only partial, and we want to state that plainly.

We still do not know how the AI model will perform across a wider crop mix. Pilot data is concentrated in stone fruit and some citrus. Pome fruit and tropical crops are included in the model but have less training data. Accuracy for those crops may remain lower until more seasons of facility data accumulate.

Nor do we know whether the observed performance will hold with many more facilities. Four facilities over one season is a real-world test, not a lab test, but it differs from 15 or 20 facilities running concurrently on shared infrastructure under load.

For the 2025-2026 season, we plan to add four to six facilities, strengthen sensor onboarding, and broaden training data enough to make firmer public claims about model behavior on crop types not yet tested commercially at scale. Until then, we will distinguish carefully between what we can claim and what we have actually observed.