Where Perishable Boxes Fail

    What a Rutgers team found when they ordered like customers, and what operators have published since.

    Aston Scott

    I've spent 22 years in the gel pack business and run a cold-chain packaging company in Grapevine, TX.

    Published: November 16, 2026 · Updated: November 16, 2026

    One vendor told customers, "Your bison meat may be thawed by the time it gets to you. Touch the meat and if it is cool to the touch your order is in good condition."

    On the last day of the 2017 Food Safety Summit in Chicago, in a session called Home Delivery, a Rutgers professor named William Hallman had a different way to check.

    He ordered the food.

    His team, working with researchers at Tennessee State, placed 169 orders the way a customer would. Inside those boxes were 684 items: 271 meat, 235 seafood, 133 game, and 39 poultry.

    As Food Safety News reported on that study, nearly 47% of those items arrived with a surface temperature above 40°F. The researchers also found the packages were likely to sit outside for eight hours or more before anyone opened them, and only about 5% of the deliveries required a signature.

    That 47% is the number people quote. What they found inside the boxes is the part that stayed with me.

    Surface temperatures jumped around within a single shipment. They even changed from one spot on the same item to another. Hallman later explained why fish is a hard one to ship: a fillet is thick in the center and thin at the edges, so two parts of the same piece can arrive in different condition. Items that went out frozen, he said, had a better chance of arriving in good condition.

    A lot of the packing had a plainer flaw. The boxes often lacked dunnage, so nothing held the food against the coolant. The porch problem above made that worse: if the box sits outside, packing that already isn't holding the food against the coolant has even less margin.

    They had also interviewed 1,002 consumers. Ninety-five percent thought this kind of food was safe.

    Hallman was blunt about the bison advice. "Cool to the touch," he said, is not a food safety standard.

    He also said something that usually gets left out of the retelling. Many of the packages did arrive at an acceptable temperature. "So it's possible to ship safely."

    And this, which I think every perishable operator should read twice:

    "The problems that we observed in our study are things that could be relatively easily resolved with the right packaging materials and the right know-how."

    The right packaging materials are part of what he was pointing at. The rest is know-how.

    That study is almost ten years old. I still see boxes fail in the same four places: the forecast, a single pack-out, the weather, and the porch. A few operators have published what they're doing about each one. Some of it is AI. Some of it is a rule you could write this week.

    The waste starts before anyone tapes the box

    The cheapest spoiled product is the product you never bought.

    HelloFresh says its production facilities wasted less than 1% of purchased ingredients in 2019, and credits a weekly machine-learning forecast tied to menu planning, fulfillment, and customer-care feedback. In a separate newsroom piece, it describes an in-house algorithm that predicts how often each recipe will be picked, so ingredients can be ordered weeks ahead without over-buying. (Company-reported.)

    The most useful breakdown I found came from Gousto, the UK recipe-box company, in a 2019 engineering post by data scientist Irene Iriarte Carretero. They forecast at three levels: how many boxes go out on a given day, how many portions of each recipe, then what to tell suppliers.

    For weeks already open for ordering, they use live basket data and look at which customers, by signup period and acquisition channel, are still likely to order. For weeks further out, they used Facebook's Prophet tool for order counts and a random-forest model for recipe popularity. New recipes borrow the history of similar ones.

    She states the trade-off plainly. They could avoid every sell-out by over-forecasting, but "that would produce a lot of waste."

    Gousto's founder, Timo Boldt, wrote in 2018 that forecasting "was one of the first places that we used AI," and that the goal was simple: "more accuracy = less food waste."

    You don't need a data science team to copy the structure. Forecast orders, then units, then ingredients. Treat a new product differently, by borrowing the history of something like it. And decide, on purpose, where you want to sit between sell-outs and waste.

    1. 1

      Orders

    2. 2

      Recipes or units

    3. 3

      Ingredients or SKUs

    Gousto's public forecast runs at three levels: boxes, recipes, then ingredients.

    One pack-out, every lane

    Most DTC perishable brands I talk to have one or two pack-outs. Their lanes, seasons, and carriers vary far more than that.

    The money version of this story is Fish Fixe, the portioned-seafood brand from Shark Tank, as told in a case study by its fulfillment partner, Grip. Shipping was eating 11% of revenue as the customer mix moved from mostly local one-day delivery to coast-to-coast. After the switch, Grip reports that Fish Fixe cut shipping costs by more than 25% and was sending 48% of orders nationwide within one day. (Vendor-reported.)

    The most detailed look at the method is Cometeer, which ships flash-frozen coffee, also through Grip. Grip describes an "AI-driven" engine that picks, for each order, the box, insulation, refrigerant, carrier, and route, based on inventory, destination, and weather. It reports a 25% reduction in refrigerant use for Cometeer and more than 30% expansion of next-day coverage. Cometeer CEO Matthew Roberts is quoted saying it "gives us confidence that all our hard work… makes it to your freezer frosty and uncompromised." (Vendor-reported case study.)

    I'm not endorsing a 3PL. The point is the method. The pack-out becomes a decision made for each order, based on lane, season, and carrier, instead of a fixed kit. You can start that with a lookup table long before you need a model.

    A weather rule beats a guess

    Some failures are decided before the label prints.

    The cleanest public example I found isn't food. Kate McLeod sells solid moisturizers that melt in the heat. The brand worked with the agency Myntr to build checkout logic that pulls the forecast for the delivery address. If the local temperature is above 85°F, the order automatically gets "Warm Weather Shipping." There's separate logic for 1-packs and 2-packs, which melt at different points, and for subscription orders.

    That's a weather rule, not AI, and a perishable brand can copy it this week.

    Ready-made options exist too. Shopify apps such as ThermoGuard and PerishShield can check forecasts along a route and flag or hold an order before the label prints. I haven't tested them, so this is only to say they exist, not a recommendation. Grip says it also watches for heatwaves, blizzards, and wildfires, and alerts brands so they can hold, reroute, or warn the customer.

    The last eight hours

    The Rutgers finding I still think about is those eight hours or more on the porch. A lot of perishable shipments are gifts, so the person at the door isn't even expecting them.

    In late October 2026, after a pilot in the Netherlands, HelloFresh Benelux said it was rolling Samsara telematics out to 500+ electric delivery vans across 10+ hubs, including temperature monitoring "to protect food quality." Most DTC brands ship by parcel and don't run vans, but the takeaway still applies: measure the last mile instead of assuming it.

    In 2020 and 2021, HelloFresh also tested Keep-It time-temperature indicator labels with more than 3,000 customers in the Netherlands and Germany, in a study with Wageningen University and the Thünen Institute. It reports "a 15% food waste reduction potential," and notes that wider use depends on EU regulatory change. (Company-reported.)

    If you ship by parcel, the cheap version is still a conversation. Send the delivery-day email or text that says this is perishable, and bring it in. Make that obvious on the outside of the box too.

    If I ran ops this quarter

    This is the list I would run:

    1. Pull 90 days of "warm / melted / thawed / damaged" tickets and tag each one by ship day, destination region, carrier service, and pack-out. Find your one worst lane.
    2. Forecast in three layers, Gousto-style: orders, then units, then SKUs or ingredients. A spreadsheet is enough to start.
    3. Write down your sell-out versus waste target, instead of letting it drift.
    4. Make pack-out a decision table by lane and season, not one kit for everything.
    5. Add one weather rule, hold, upgrade, or notify, at a temperature threshold you have validated for your own product.
    6. Check your dunnage. In the Rutgers study, product that wasn't held against the coolant was a recurring problem.
    7. Tell the recipient it's perishable on delivery day, especially when the box is a gift.

    None of that needs a big AI project. The models in these cases sit on top of disciplined data about orders, lanes, and outcomes. Get that data right first.

    Hallman said the problems could be resolved with the right packaging materials and the right know-how. The list above is the know-how. If you want to see the gel packs and shippers we make for meal-kit and perishable brands, they're here. That's one of the materials. It doesn't replace a forecast, a pack-out table, or a plan for the porch.

    Question for operators: which of the four costs you the most right now, the forecast, the pack-out, the weather, or the doorstep? I would like to hear what has actually worked for you.


    Sources

    1. Food Safety News, "Research shows food safety gaps in home-delivery meal kits" (May 12, 2017). Hallman at the 2017 Food Safety Summit, Chicago, "Home Delivery" session: 169 shipments; 684 items (271 meat, 235 seafood, 133 game, 39 poultry); ~47% above 40°F surface temperature; variation within shipments and on the same item; lack of dunnage; 8+ hours outside; ~5% signature required; 1,002 consumer interviews, 95% thought the products were safe; bison "cool to the touch" vendor guidance, which Hallman said is not a food safety standard. https://www.foodsafetynews.com/2017/05/research-shows-food-safety-gaps-in-home-delivery-meal-kits/
    2. Supermarket News, "Freshness remains elusive for seafood delivery" (Aug 22, 2017). Hallman quote ("relatively easily resolved with the right packaging materials and the right know-how"); "So it's possible to ship safely"; many packages arrived at an acceptable temperature; frozen items had a higher chance of arriving in good condition; fish fillets vary in thickness from center to edge. https://www.supermarketnews.com/seafood/freshness-remains-elusive-for-seafood-delivery
    3. HelloFresh Group newsroom, "Fighting Food Waste: How HelloFresh helps tackle a global issue." <1% of purchased ingredients wasted in production facilities (2019); weekly dynamic ML forecasting. (Company-reported.) https://www.hellofreshgroup.com/en/newsroom/stories/122cb0f3-05ac-5e32-9f60-737b60002d88/
    4. HelloFresh Group newsroom, "The Data River – How Data Flows Through the HelloFresh Production Process." In-house recipe-demand forecasting algorithm. (Company-reported.) https://www.hellofreshgroup.com/en/newsroom/stories/the-data-river-how-data-flows-through-the-hellofresh-production-process/
    5. Irene Iriarte Carretero, Gousto Engineering & Data, "Reducing Food Waste Through Data Science – Forecasting at Gousto Part 1" (Apr 30, 2019). https://medium.com/gousto-engineering-techbrunch/reducing-food-waste-through-data-science-e18e6fee56cf
    6. Timo Boldt (Gousto founder), "Powered by Artificial Intelligence – Understanding Gousto's AI," LinkedIn (Jun 12, 2018). https://www.linkedin.com/pulse/powered-artificial-intelligence-understanding-goustos-timo-boldt
    7. ReFED, "Three Ways AI Is Driving Reductions in Food Loss and Waste" (Oct 22, 2024). Summarizes the 2022 Pacific Coast Food Waste Commitment pilot (Shelf Engine + Afresh, 14.8% average food-waste reduction per store). https://refed.org/articles/three-ways-ai-is-driving-reductions-in-food-loss-and-waste/
    8. Afresh, "Using artificial intelligence to reduce food waste in grocery retail" (PCFWC case study page; results measurable as early as eight weeks). https://www.afresh.com/resources/guides/using-artificial-intelligence-to-reduce-food-waste-in-grocery-retail-pcfwc-case-study
    9. Grip, "How Cometeer Delivers Frozen Coffee at Scale with Grip" (Jul 11, 2025). Vendor case study: per-order engine, 25% refrigerant reduction, 30%+ next-day coverage, weather alerts, Matthew Roberts quote. (Vendor-reported.) https://gripshipping.com/blogs/how-cometeer-delivers-frozen-coffee-at-scale-with-grip
    10. Grip, "How Shark Tank's Fish Fixe overcame their expensive shipping issues with Grip" (Aug 27, 2024). Vendor case study: 11% of revenue on shipping; >25% shipping cost cut; 48% of orders within one day. (Vendor-reported.) https://gripshipping.com/blogs/how-shark-tanks-fish-fixe-overcame-their-expensive-shipping-issues-with-grip
    11. Myntr Commerce, "Kate McLeod Cool Shipping" case study. Weather-based checkout logic, 85°F threshold, separate 1-/2-pack logic, Recharge. https://myntr.io/blogs/case-studies/kate-mcleod-cool-shipping
    12. Shopify App Store listings: ThermoGuard (https://apps.shopify.com/thermoguard) and PerishShield (https://apps.shopify.com/perishshield). Cited only to show that weather-hold apps exist; not tested or endorsed.
    13. Samsara press release, "HelloFresh Selects Samsara to Power its Electric Fleet in Benelux" (Oct 28, 2026). https://www.samsara.com/uk/company/news/press-releases/samsara-hellofresh
    14. HelloFresh Group newsroom, "Every second person throws away food once or twice a month…" (Jun 28, 2023). Keep-It TTI test with 3,000+ customers, Wageningen/Thünen study, "15% food waste reduction potential." (Company-reported.) https://hellofreshgroup.com/en/newsroom/stories/every-second-person-throws-away-food-once-or-twice-a-month-how-innovative/