76 problems facing farming today.
AI will solve them by 2030.

What’s brokenWe're feeding plants yesterday's recipe.

Why it hasn’t been solvedFarms dose a fixed recipe and top up to a target salt level, so as daily demand shifts with growth stage and climate they oversupply some ions and undersupply others.

Why AI changes the equationSimulation and optimization can predict each day's real nutrient demand from crop stage and climate, then steer dosing to match it instead of holding a fixed recipe.

What’s still requiredAccurate demand models for each crop and stage, dosing controls that can act on the predicted demand, and validation across full crop cycles.

The payoffDemand-matched dosing modeled roughly 51% lower fertilizer costs, about double fertilizer-use efficiency, and up to sixfold water-use efficiency versus a fixed recipe in a 2026 hydroponic-lettuce simulation; a field grapefruit trial reported up to ~40% savings in a different setting. Modeled, not farm-measured: a direction, not a promised saving.

What’s brokenWe know how much fertilizer is in the water. Not which nutrients are missing.

Why it hasn’t been solvedElectrical Conductivity (EC) tells you the total salt concentration, the overall fertilizer level, but not the concentration of any single ion. No sensor today reads nitrate, potassium, calcium, magnesium, phosphorus, sodium and micronutrients continuously in-line at farm-grade accuracy for a small grower, so a missing nutrient hides inside a normal-looking EC reading.

Why AI changes the equationAI cannot invent the missing measurement; the real blocker is accurate, affordable ion hardware. Once a signal exists, models can combine partial readings with other water and crop data to estimate the rest.

What’s still requiredAccurate, affordable in-line ion hardware to provide a reliable signal, and validation that estimates hold across a real, changing solution.

The payoffKnowing every nutrient in the water in real time would cut fertilizer waste and prevent the hidden shortages and excesses that cost yield and quality.

What’s brokenThe sensors that read nutrients don't stay honest.

Why it hasn’t been solvedIon-selective electrodes, the only affordable per-ion option, drift, react to interfering ions, and foul over a crop cycle, with no turn-key compensation keeping them accurate without manual recalibration.

Why AI changes the equationA model can learn each sensor's drift and cross-interference from repeated farm measurements and recalibration records, then correct readings between cleanings instead of trusting the raw number.

What’s still requiredLong-term drift and interference data across a full crop cycle, and validation that the corrections stay accurate without frequent manual recalibration.

The payoffSensors that stay honest between cleanings would prevent both over- and under-feeding and make per-ion monitoring practical on a small farm.

What’s brokenThe meter says the salt is fine. The plant is starving.

Why it hasn’t been solvedAn EC meter estimates overall dissolved-ion concentration and a pH meter measures acidity; neither identifies the concentrations of individual nutrients, so holding a target total drifts into imbalance as fast-uptake nutrients deplete while slow-uptake ones accumulate.

Why AI changes the equationAn inference model can read hidden ionic balance from the history of salt and acidity readings, crop stage, and occasional water tests, flagging the imbalance the single number hides.

What’s still requiredPaired historical water-test and continuous-reading data, and validation that inferred balance matches lab analysis.

The payoffSurfacing hidden imbalance early would stop a farm feeding a plant a mix it cannot use, protecting yield and quality while wasting less fertilizer.

What’s brokenWe can see some nutrients in the water. Not the ones that matter most.

Why it hasn’t been solvedCheap optical chips can infer some ions without wet chemistry, but they fail on phosphorus, magnesium, and sulfur in real, changing solution matrices.

Why AI changes the equationWith enough calibration data, model fitting can turn weak optical signals into usable soft sensors that estimate the ions raw optics cannot read directly.

What’s still requiredCalibration datasets spanning real water chemistry, and validation that soft-sensor estimates hold as the solution changes.

The payoffCheap, continuous ion estimates would let small farms dose precisely without lab tests, cutting fertilizer waste and protecting crop quality.

What’s brokenMinerals quietly clog the pipes and steal the food.

Why it hasn’t been solvedCalcium and sodium compounds build up in a closed loop: scale strips calcium and phosphorus out of solution and clogs flow, while sodium and chloride accumulate with every top-up, and growers only react after it forms.

Why AI changes the equationPrecipitation risk can be modeled from solution chemistry and operating history, so the system warns before scale forms instead of after it clogs.

What’s still requiredSolution chemistry and operating records across full crop cycles, and validation that predicted precipitation matches what actually forms.

The payoffPreventing scale and salt buildup protects nutrient availability, keeps flow open, and avoids blind discharge of water.

What’s brokenPlants can suffocate underwater. And we might never see it coming.

Why it hasn’t been solvedRoots take oxygen both from the flowing nutrient solution and from the air inside the channel. In a thin-film system with a dense root mat, oxygen at the root surface can fall well below the reservoir reading, so what matters is oxygen where the roots actually sit, and that is exactly what conventional monitoring misses.

Why AI changes the equationAI could combine limited oxygen measurements, water-flow data, temperature, and root imagery to estimate where oxygen shortages develop without requiring a sensor at every root. Those estimates would need validation against actual root-zone oxygen measurements.

What’s still requiredRoot-zone oxygen measurements to validate predictive models, multimodal inputs (flow, temperature, imagery), and testing in operating farms.

The payoffDetecting oxygen shortages before they damage roots could cut crop losses, improve growth consistency, and enable more efficient oxygen management.

What’s brokenThe chiller runs like a light switch while the bill climbs.

Why it hasn’t been solvedHolding water at the right temperature means running a chiller that is often most of the nutrient-delivery subsystem's power, switched on and off with no smart control.

Why AI changes the equationThis is a standard control problem, not a frontier-AI problem: a controller can run the chiller and pumps to hold temperature with far less on/off cycling, using known thermal behavior.

What’s still requiredTemperature and load histories, plus a controller that can modulate cooling instead of simply switching it.

The payoffSmarter cooling cuts the nutrient-delivery subsystem's power use, and because water cooling is a large share of that load, it directly lowers operating cost.

What’s brokenPlants poison their own water for the next crop.

Why it hasn’t been solvedRecirculating solutions accumulate allelochemical exudates that suppress the next crop even when the nutrient test looks perfect, so the test gives no warning.

Why AI changes the equationThis needs controlled wet-biology experiments, not compute alone. AI can help design those experiments and find which grow conditions correlate with the buildup, but the biology has to be measured.

What’s still requiredControlled experiments isolating the compounds, long crop-cycle validation, and sensors that can detect buildup before it hurts the crop.

The payoffCatching autotoxicity early could recover a reported yield loss of around 30 percent in lettuce and let recirculating systems run longer without hidden decline.

What’s brokenAlgae steals the food and the air, then comes back.

Why it hasn’t been solvedAlgae consumes nutrients and oxygen, and its surface growth supports biofilm that can harbor plant pathogens, and it regrows wherever light reaches the solution.

Why AI changes the equationImage and environment signals can automate algae detection and decide when to intervene, timing a response to the growth rather than reacting after it spreads.

What’s still requiredImaging and environmental data tied to algae growth, and validation that detection triggers the right intervention at the right time.

The payoffCatching algae early protects nutrients and oxygen, reduces pathogen risk, and keeps food and water cleaner.

What’s brokenEvery way to clean the water has a side effect.

Why it hasn’t been solvedEvery disinfection option has a documented downside: ultraviolet light weakens with cloudiness and organic load and breaks down iron chelates, ozone oxidizes micronutrients, and peroxide risks harming the crop, so growers have no shared standard for matching a dose to a given water.

Why AI changes the equationEfficacy and residual safety require physical and food-safety validation, so AI's role is to model dose against water quality and residual risk, not to replace the treatment itself.

What’s still requiredWater-quality and dose-response data, residual measurements, and food-safety validation before any dose rule can be trusted.

The payoffMatching disinfection dose to water quality would protect micronutrients (ozone can destroy most of the chelated iron and much of the manganese), keep the crop safe, and stop growers from guessing.

What’s brokenRoot rot kills from below, days before anything looks wrong up top.

Why it hasn’t been solvedPythium and Fusarium attack the hidden root zone and only surface on the leaves once the damage is already done, so a camera pointed at the canopy sees nothing until it is too late. The infection is established well before it is visible.

Why AI changes the equationVision trained on the canopy and root zone, combined with the room's environment history, could learn the faint leading indicators that come before root disease, catching the change in the days when treatment can still save the crop.

What’s still requiredConsistent imaging of roots and canopy, clean long-term environment logs, and validation in operating farms where outbreaks can be confirmed by lab tests.

The payoffCatching root rot days earlier would let growers intervene while the crop is still salvageable, cutting crop loss and the cost of culling whole channels.

What’s brokenCameras only spot a hungry plant once it already looks sick.

Why it hasn’t been solvedNutrient-deficiency symptoms visible to the eye often appear only once a plant is already badly damaged, so a camera sees the shortage at roughly the same late point a person would. In a 5,648-image sugar-beet study, telling deficiency types apart from ordinary RGB images was hard: phosphorus deficiency was the most difficult macronutrient deficiency to recognize, liming omission was confused with the unfertilized control, and a comparable RGB study across seven nutrient treatments reached only 65.4% test accuracy. Catching a shortage in the days before symptoms appear is still an open problem.

Why AI changes the equationVision models can be trained to read the earliest stress signatures in leaf appearance, changes too subtle for the eye and too early for today's models, turning ordinary imagery into an early-warning signal.

What’s still requiredRepeated imaging of the same plants over time paired with confirmed nutrient measurements as ground truth, plus validation across growth stages and varieties.

The payoffFeeding a crop days earlier could protect yield and reduce the fertilizer wasted on a plant that has already stalled.

What’s brokenA plant can have plenty of calcium in its water and still develop tipburn.

Why it hasn’t been solvedCalcium and magnesium are secondary macronutrients, not micronutrients. Tipburn is driven mainly by a localized calcium shortage in fast-growing leaf edges: calcium moves with water in the transpiration stream, so a leaf edge expanding faster than calcium can reach it can burn even when the solution is well supplied. Magnesium deficiency is a different disorder with different symptoms. Both are hard to see before damage shows.

Why AI changes the equationAI could combine growth rate, transpiration-related climate data, nutrient supply and imagery to estimate tipburn risk before symptoms appear. Such predictions would need validation across cultivars and operating conditions.

What’s still requiredMultimodal sensors sensitive to weak signals, repeated non-destructive measurements, and lab ground truth tying those signals to actual tissue nutrient levels.

The payoffCatching calcium and magnesium shortages early could cut the tipburn and leaf disorders that downgrade marketable heads and waste already-grown crop.

What’s brokenThe cameras that see trouble earliest cost many times more than ordinary ones.

Why it hasn’t been solvedThe sensing that gives the earliest warning, spectral imaging, costs far more than an ordinary camera (one review describes hyperspectral systems as costly, often several thousand US dollars), and it produces far more data than the computer at the edge of a farm can process. No device yet closes the gap between cheap, spectrally rich, and on-device.

Why AI changes the equationAI can compress and interpret spectral data on low-cost edge hardware and help keep a cheap sensor calibrated, but it cannot make the sensor itself cheaper. The blocker is affordable spectral hardware, not model compute.

What’s still requiredAffordable spectral hardware with reliable calibration, edge processors able to handle the data volume, and field validation that low-cost units match laboratory instruments.

The payoffGetting early-warning sensing onto a farm-level budget would put disease and stress detection within reach of ordinary growers, not just well-funded labs.

What’s brokenA model trained on one farm can fall apart on the next.

Why it hasn’t been solvedA stress model tuned to one variety and one grow room does not transfer. In a 2026 field study of virus detection, average accuracy fell from 92.3% under cross-validation to 67.4% under external validation, and the MCC dropped from 84.1% to 32.2%, so every new deployment becomes a fresh data-collection project.

Why AI changes the equationDomain adaptation and farm-specific calibration let a model adjust to a new variety or room from a small amount of local data instead of being retrained from scratch.

What’s still requiredDiverse multi-farm and multi-cultivar datasets, methods that express uncertainty when a model meets new conditions, and on-farm validation that transfer actually holds.

The payoffModels that generalize would let a working system reach the next farm without a full re-collection, cutting the cost and time of every new deployment.

What’s brokenBacteria hide inside the leaf before anything shows on the surface.

Why it hasn’t been solvedBacterial pathogens colonize a leaf internally, and visible lesions only appear once the infection is advanced. In one comparison, spectral imaging caught soft-rot several days after inoculation while ordinary cameras took roughly three weeks, far too late for many interventions.

Why AI changes the equationImage-based anomaly detection could flag the internal bacterial stress a normal camera misses, days before lesions, so a grower can confirm and act while the crop can still be protected.

What’s still requiredSpectral or multispectral imaging at farm scale, laboratory-confirmed infection labels, and validation that early flags match real outbreak timing.

The payoffEarlier detection lets growers remove or treat affected plants before bacteria spread through a shared water system, protecting the rest of the crop.

What’s brokenA virus can spread through a crop for weeks without a single sign.

Why it hasn’t been solvedViruses replicate for weeks before any color change and travel mechanically through shared water. Detection built for one specific virus reaches high accuracy, but no cheap, general test can simply declare that a virus is present across types.

Why AI changes the equationComputer vision paired with biological assays could learn the weak spectral signatures different viruses leave early, while lab work supplies the ground truth that vision alone cannot. The biology, not the model, is the gate.

What’s still requiredConfirmed laboratory assays as reference, hyperspectral imaging sensitive enough for latent infection, and multi-virus datasets to test whether one model generalizes.

The payoffCatching a virus while it is still latent could stop it from spreading through an entire recirculating system, avoiding the loss of a whole planting.

What’s brokenThe pests are smaller than a grain of rice and they hide under the leaves.

Why it hasn’t been solvedAphids, thrips and mites are tiny and tuck themselves under leaves and into canopy shadows, so a camera watching a whole tray sees little until the population has already become an economic threat.

Why AI changes the equationComputer vision can scan trays at scale and pick out pests too small and too hidden for a person walking the rows, flagging hot spots early for targeted treatment instead of blanket spraying.

What’s still requiredHigh-resolution imaging that reaches under the canopy, labeled pest imagery collected at farm scale, and validation against human scouting counts on operating farms.

The payoffFinding a pest population while it is still small means treating a few plants instead of the whole room, cutting both crop loss and chemical use.

What’s brokenHeat and drought show up in the leaves before the plant ever looks bad.

Why it hasn’t been solvedTipburn and heat or drought stress appear first as a rise in canopy temperature or a small dead patch at the leaf tip, signals that blend into ordinary variation and that a single camera struggles to separate from the normal day-to-day change of a crop.

Why AI changes the equationFusing thermal imagery with ordinary imagery over time, alongside climate data, lets a model separate genuine early stress from normal variation. In research, thermal-plus-visual fusion has flagged water stress up to 84 hours ahead of conventional soil-moisture sensors, though that lead is from a single controlled lettuce study.

What’s still requiredThermal and visual cameras deployed together, continuous time-series data, and validation that early flags translate into saved yield on real farms.

The payoffCatching heat or drought stress hours earlier gives a grower time to adjust the climate before the crop stalls, protecting both yield and quality.

What’s brokenThe plant photos we train on don't look like a real farm.

Why it hasn’t been solvedThe big public plant-disease datasets are single detached leaves photographed on plain backgrounds, not whole canopies of greens growing under grow lights. Models trained on them do not transfer to real conditions: a Purdue field trial on corn reported its best generalization accuracy at 81.6 percent (DenseNet169, with image backgrounds removed), and that work is on corn rather than greenhouse greens under grow lights.

Why AI changes the equationBuilding large, labeled, open datasets of whole canopies under real grow conditions is the enabling move. AI then trains on data that matches the farm instead of the lab, and can help label and curate that data at scale.

What’s still requiredShared data standards, images captured under real grow lights, expert labeling, and a common benchmark every study reports against.

The payoffOpen datasets that match reality would let every farm build on shared work instead of repeating the same collection, accelerating reliable detection across the whole industry.

What’s brokenA model can tell you a plant is stressed. It can't tell you why.

Why it hasn’t been solvedToday's models report that a plant is stressed but not that Tuesday's humidity spike caused it. Fusing vision with environment data improves detection, but trustworthy cause-and-effect demands designed experiments and causal methods, not just more data.

Why AI changes the equationCausal modeling on top of fused vision and environment data could point to the actual driver of a stress signal rather than the symptom, turning a detection into an explanation a grower can act on. The experiments that prove causality still have to be designed and run.

What’s still requiredControlled experiments that vary one factor at a time, dense environment and imagery logs, and causal methods validated against causes that are already known.

The payoffKnowing the cause, not just the symptom, lets a grower fix the real problem instead of treating the leaf, cutting repeat losses and wasted interventions.

What’s brokenThe part of the plant that matters most is the part no one ever looks at.

Why it hasn’t been solvedRoots are the earliest warning for nearly every disease, yet they are the least-observed tissue in a farm: white, matted, low-contrast, and hidden inside the channel where no camera naturally looks, so the canary dies out of sight.

Why AI changes the equationComputer vision can turn root images into repeatable health scores, and prediction models can link those scores to the crop outcome weeks later, replacing a passing glance with a number a grower can actually track over a cycle.

What’s still requiredStandardized imaging inside the channel, a labeled set of scored root images tied to measured outcomes, and validation across crops and systems.

The payoffA repeatable root-health score would catch disease and stress before the leaves show it, reducing crop loss and giving a grower a leading indicator instead of a post-mortem.

What’s brokenWe're trying to grow healthy roots without knowing what healthy even looks like.

Why it hasn’t been solvedNo validated reference profile exists for a healthy hydroponic root microbial community, not the expected organisms, not a diversity target, not a marker of disease suppression, so growers have no baseline to monitor or steer toward, and a beneficial product cannot be judged against a standard.

Why AI changes the equationScientific discovery can mine the scattered sampling and sequencing studies to assemble a candidate baseline community, then flag which organisms and diversity metrics actually track a healthy crop rather than merely co-occurring with one.

What’s still requiredExtensive biological sampling and culture work to build the reference dataset, plus validation that the proposed markers predict real crop health and not just laboratory conditions.

The payoffA quantitative healthy baseline would let growers monitor the root microbiome, judge beneficial products against a standard, and begin steering the community instead of guessing at it.

What’s brokenThe good bacteria we buy for water farms were designed for soil.

Why it hasn’t been solvedThere is no reliable inoculation protocol for Bacillus, Trichoderma, or plant-growth-promoting rhizobacteria in recirculating systems: all three evolved for soil, often fail to establish in flowing water, and some marketed beneficials measurably depress growth instead of helping it.

Why AI changes the equationScientific discovery can mine trial and sequencing data to identify which strains and conditions actually establish and help in recirculating water, and autonomous research can run that protocol search at a scale no single lab can.

What’s still requiredWet-biology trials in operating recirculating systems, survival and colonization measurements, and ground-truth crop outcomes; compute cannot prove microbial efficacy on its own.

The payoffA working inoculation protocol would let water farms use biological disease suppression instead of chemistry, improving root health and cutting input cost.

What’s brokenWe can detect the pathogen. We just can't agree on when to act.

Why it hasn’t been solvedDetection methods disagree in both sensitivity and what they measure: culture plating, baiting and lateral-flow tests each respond at different pathogen levels, and a DNA test detects far lower levels but cannot tell a live pathogen from a dead one, so no actionable 'act now' concentration exists.

Why AI changes the equationPrediction models can estimate a practical action threshold from longitudinal water tests, symptom onset, and intervention records, turning a set of disagreeing assays into a single risk signal a grower can act on.

What’s still requiredLongitudinal water-test series linked to symptoms and interventions, and validation that a proposed threshold holds across systems and seasons.

The payoffA defensible action threshold would let growers treat at the right moment, early enough to matter, late enough not to waste inputs, reducing both crop loss and needless treatment.

What’s brokenDisease can sweep a whole farm before the test comes back.

Why it hasn’t been solvedEnvironmental DNA sequencing reads the whole loop community but through a routine lab it is too slow to act on: a state plant-pathology lab quotes about 1-2 weeks for a PCR identification and 2-4 weeks for sequencing, and a commercial molecular plant-testing lab quotes about 2 weeks. The literature notes that sending samples to a laboratory and obtaining results often requires several days. Rapid on-site nanopore and portable qPCR workflows can compress this to hours: one published workflow reports roughly 6 hours end to end, but they are not yet routine on a working farm, and the fast targeted tests only detect pathogens they were designed to find. Circulating nutrient solution is a principal route by which Pythium propagules disperse, so disease can move while the sample is in transit.

Why AI changes the equationPrediction models can forecast outbreak risk from fast, cheap signals so growers act during the days the lab is still running; the sequencing itself remains a laboratory cost and latency problem, not a prediction one.

What’s still requiredFaster or field-deployable assays, cheaper sequencing, and validated early signals that predict an outbreak before the results arrive.

The payoffClosing the gap between an outbreak and its answer would let a grower isolate and treat in time, averting the loop-wide loss a week-long wait makes nearly certain.

What’s brokenThere's a layer of slime growing inside the pipes, and no one can see it.

Why it hasn’t been solvedBiofilms on channel walls and drains harbor pathogens and consume dissolved oxygen, yet no practical sensor measures biofilm load in place inside a working channel without stopping the flow or taking the tier apart, so the problem is only found after it has already grown.

Why AI changes the equationComputer vision and prediction can estimate biofilm buildup from the images and flow signals a farm already collects, flagging where slime is thickening before it chokes a line or seeds disease.

What’s still requiredNew in-situ sensing hardware, biological ground truth for what the sensor is actually seeing, and validation inside operating channels.

The payoffSeeing biofilm before it blocks flow or shelters pathogens would protect dissolved oxygen, cut cleaning surprises, and shrink the disease reservoir in a shared loop.

What’s brokenA blocked drain can kill the roots before a leaf ever looks sick.

Why it hasn’t been solvedA clogged drain or failed pump stops the nutrient film, the roots begin to dry, and the damage is done before any shoot symptom appears, so a top-down check of the plants cannot catch the failure in time to save them.

Why AI changes the equationPrediction models can learn the sound, level, pressure, and flow signatures that precede a clog or pump failure and warn in minutes instead of days.

What’s still requiredInstrumentation for sound, level, pressure, and flow on small systems, labeled failure events, and validation that early warnings fire before crop damage occurs.

The payoffMinutes of warning instead of days would turn a crop-killing failure into a maintenance stop, reducing crop loss and downtime on every rack.

What’s brokenOne bad drop of water can infect the whole farm. We can't yet trace where it started.

Why it hasn’t been solvedOne shared reservoir and manifold means a single pathogen introduction reaches every tier, and circulating nutrient solution is a principal route by which Pythium propagules disperse. What is missing is a routine field tool that reconstructs where an introduction entered or how fast it moved through the loop, so a grower cannot isolate the source before the whole loop is infected.

Why AI changes the equationPrediction and discovery methods can combine pathogen sampling across tiers with flow timing to estimate the entry point and the spread path, turning scattered positives into a source hypothesis a grower can test.

What’s still requiredLoop-wide pathogen sampling, flow and timing records, and biological validation that the inferred source matches the real one.

The payoffLocalizing the source would let a grower isolate one tier or inlet instead of treating the entire loop, cutting both crop loss and the cost of a full-system response.

What’s brokenWe run a whole room like it's one plant.

Why it hasn’t been solvedClimate and lighting are steered from a single room-level sensor average, but conditions differ measurably by tier and position along the rack; no small-farm study quantifies what per-tier control actually buys, so growers pay room-level prices for tier-level outcomes.

Why AI changes the equationIntelligent automation and a digital twin can model how conditions vary tier by tier and drive control only where it changes the crop, instead of averaging the whole room into one setpoint.

What’s still requiredPer-tier sensing, a validated model of how conditions vary across a rack, and trials that measure what finer control actually returns.

The payoffControl aimed at the tier rather than the room would reduce wasted light and climate energy and even out crop quality across a rack.

What’s brokenWe can change how food tastes with light. We just don't know the recipe.

Why it hasn’t been solvedThere is no turn-key recipe for which ultraviolet, far-red, or blue-to-red mix moves a target compound, and published effects contradict across dose, background conditions, and cultivar, so the recipe stays guesswork.

Why AI changes the equationAutonomous research can run high-throughput light trials and prediction models can map dose and spectrum to a target compound, searching a space far too large to test by hand.

What’s still requiredControlled multi-factor trials, compound assays for ground truth, and validation across cultivars and growing conditions.

The payoffA repeatable light recipe for flavor and nutrition would let a grower dial a target taste and quality instead of accepting whatever the crop happens to deliver.

What’s brokenRun the lights when power is cheap and the bill drops. Nobody sells that controller.

Why it hasn’t been solvedOptimized price-aware lighting controls cut lighting energy cost by about 4 to 65 percent versus fixed schedules in studies and simulations, but no turn-key controller exists for a small rack, so the savings stay in the lab.

Why AI changes the equationIntelligent automation can solve the constrained scheduling problem, shifting light to cheap hours while respecting crop response, climate, and daily light needs, and ship it as a controller.

What’s still requiredReliable power-price and crop-response models, plus validation on real small-rack systems.

The payoffLoad-shifting light to cheaper hours would cut electricity cost for the single largest power draw in a lit farm while keeping the crop on target.

What’s brokenWe steer the grow room by a number built on sensors that drift.

Why it hasn’t been solvedVapor pressure deficit is only as good as its temperature and humidity inputs, and cheap sensors drift and compound their error worst at the high-humidity end, exactly where condensation and disease risk live.

Why AI changes the equationPrediction models can detect and correct sensor drift from the redundancy of many readings over time, keeping the steered number honest without constant manual recalibration.

What’s still requiredSensor redundancy or reference instruments, drift histories, and validation that corrected readings match a trusted standard.

The payoffHonest humidity control would avoid the condensation and disease that hide at the humid end, and stop control from chasing a number that is quietly wrong.

What’s brokenWe can feed plants more carbon dioxide. Keeping it profitable is the hard part.

Why it hasn’t been solvedAdding carbon dioxide can raise photosynthesis and growth in both sealed and ventilated growing spaces, depending on concentration and conditions. The hard part is economics: in a ventilated space, air exchange for ventilation, humidity and temperature control lets the added CO2 escape, so enrichment competes for energy and dollars with the systems that keep the crop healthy.

Why AI changes the equationIntelligent automation and a digital twin can weigh enrichment against ventilation, dehumidification, cooling and electricity cost in real time and dose CO2 only when it actually pays.

What’s still requiredMeasured growth and yield response to enrichment, sealed-room energy data, and a validated model tying dose to outcome.

The payoffDosing enrichment only when it pays would capture any growth gain without letting the sealing and moisture costs erase it.

What’s brokenThe top shelf bakes while the bottom shelf stays wet.

Why it hasn’t been solvedStacked racks create a real vertical gradient, the top tier runs hot and dry under the lights, the bottom cool and wet, and airflow that fixes the temperature can wreck humidity removal, so one adjustment worsens the other.

Why AI changes the equationComputer vision and a digital twin can map the airflow and thermal pattern across a rack, and intelligent automation can target control where the gradient actually hurts the crop.

What’s still requiredSensor and imaging maps of airflow and temperature, a validated model of the rack microclimate, and trials of targeted control.

The payoffEvening out the rack microclimate would reduce hot spots and damp zones, improving crop uniformity and cutting the energy spent fighting the gradient.

What’s brokenThe plants sweat out more than half our climate energy.

Why it hasn’t been solvedIn a transpiring room the latent moisture load is often more than half of the climate energy, and the common reheat and condensing strategies routinely waste much of it, so the biggest climate cost is also the least optimized.

Why AI changes the equationIntelligent automation can coordinate dehumidification, cooling, and reheat so moisture is removed with the least energy, instead of by a fixed strategy that ignores what the room is doing.

What’s still requiredReal-time moisture-load sensing, equipment performance data, and a validated control model, software can optimize the system but cannot remove the dehumidification hardware.

The payoffRemoving moisture at lower energy cost would cut the largest share of climate consumption, lowering both power use and operating cost.

What’s brokenOur climate system runs on last week's assumptions about plants that change minute to minute.

Why it hasn’t been solvedPlants are the fastest-changing humidity source in the room, but no farm measures transpiration live, lysimeters are expensive and stem gauges invasive, so climate control runs on static assumptions instead of the load the plants are actually adding.

Why AI changes the equationPrediction models can forecast transpiration from climate and crop signals a farm already collects, giving climate control a live estimate of the load it is actually fighting.

What’s still requiredCrop and climate signals tied to measured transpiration ground truth, and validation that the forecast improves real control decisions.

The payoffMatching climate control to real transpiration would cut both the energy spent over-drying and the crop risk from under-drying.

What’s brokenAI can run the whole farm in simulation. It can't run a real one.

Why it hasn’t been solvedModel-predictive and reinforcement-learning control beats fixed setpoints in simulation but nothing ships at small-farm scale: there is no calibrated digital twin to plan against, sensors drift, and models do not transfer across facilities.

Why AI changes the equationA digital twin combined with reinforcement-learning control can search the coupled climate, light, and water decisions together, but it needs a twin that matches the real farm, which is the part that remains unsolved.

What’s still requiredA calibrated digital twin for the facility, drift-resistant sensing, and validation that learned control transfers to a real rack without losses.

The payoffControl that learns the coupled loop would beat fixed rules on energy, water, and yield at the same time, a gain simulation keeps promising and the field has not yet delivered.

What’s brokenIndoor farming's energy cost is set by physics, not by scale.

Why it hasn’t been solvedThe energy to light a crop per kilogram is dominated by fixture physics: roughly 10-18 kWh/kg for stacked vertical growing versus about 5.4 for a greenhouse. That is not fixed; it improves with higher fixture efficacy and higher yield per square meter, but scale alone does not change the physics of lighting a canopy. Rising energy prices contributed to CEA failures, though the reported causes were mostly multi-causal.

Why AI changes the equationA digital twin can model where the energy actually goes per kilogram and test operating changes against that floor, but no model can remove the physical energy a crop requires to grow.

What’s still requiredMetered energy data tied to crop and yield, a validated per-kilogram model, and operating trials; this is a structural economic limit, not a computation to be run.

The payoffKnowing the true energy floor per kilogram would let a grower price honestly, target the loads that can move, and avoid bets that physics has already lost.

What’s brokenA farm can promise you a harvest date. It cannot yet guarantee it.

Why it hasn’t been solvedThe pieces exist separately: harvest-date forecasts carry error of around two to three days, and honest confidence intervals can be built, but no system joins a calibrated forecast, a schedule that acts on it, and a rule for who absorbs the shortfall when the crop lands outside the promised range.

Why AI changes the equationAI forecasting can carry a calibrated range instead of a single number, and scheduling optimization can plan harvests against that range rather than against one fixed yield figure, turning a guess into a commitment with stated odds.

What’s still requiredValidated yield forecasts, properly calibrated prediction intervals, and an integrated scheduling and shortfall-allocation workflow tested in a working operation.

The payoffReliable delivery commitments instead of missed dates, less product stranded when supply and demand are mismatched, and confidence intervals customers can actually plan around.

What’s brokenA camera can weigh a plant in a lab. It cannot yet do it in a packed greenhouse.

Why it hasn’t been solvedImage-based weight estimation works in controlled conditions, a 3D-reconstruction model estimated lettuce fresh weight with a mean absolute error of about two and a half grams, but accuracy collapses once leaves overlap in a real crop, where one multi-modal study reports error climbing into the tens of grams and varying widely across varieties. Stacked racks compound the problem because the confined multi-layer space leaves no room for a fixed overhead camera.

Why AI changes the equationComputer vision paired with sensor fusion can estimate biomass from multiple views or a moving camera, and models can be trained to stay accurate when leaves occlude one another, something a single fixed-camera measurement cannot do.

What’s still requiredMulti-view or mobile imaging hardware that fits confined stacked racks, occlusion-robust models, and validation against destructive harvest weights in real crops.

The payoffContinuous, non-destructive yield estimates that feed forecasts and scheduling, without sacrificing plants to measurement.

What’s brokenThe best public crop model was fitted to someone else's greenhouse.

Why it hasn’t been solvedNiCoLet, the best-known public plant-growth model published between 1998 and 2004, was fitted to a specific greenhouse context. Re-fitting it to a new variety or a new room still means going back and re-measuring the plants by hand.

Why AI changes the equationParameter inference and transfer learning can recalibrate an existing mechanistic model from data a farm already generates, replacing a fresh round of manual measurements with computation against logged climate and growth records.

What’s still requiredReliable farm data streams, parameter-estimation methods that hold up across varieties and rooms, and validation against measured growth.

The payoffGrowth models that actually fit the farm they run on, so yield forecasts start from a model that matches local conditions.

What’s brokenWe can predict how big a plant gets. Not what is inside it.

Why it hasn’t been solvedHybrid models that combine machine learning with plant physics predict overall growth well, but they remain weak on the details that matter most, nutrient levels and internal composition, and none has been proven across a full multi-week crop cycle.

Why AI changes the equationPhysics-informed machine learning constrains a learned model with known plant mechanism, so the model can be tested against repeated crop cycles and pushed toward the internal quantities, like nutrient status, where pure data-driven models fail.

What’s still requiredMulti-week crop-cycle datasets with internal nutrient and quality measurements, hybrid models validated over a whole cycle, and replication across crops.

The payoffForecasts that track not just bulk growth but the internal quality of the crop, enabling earlier intervention before a cycle goes wrong.

What’s brokenA forecast can look certain and still miss the date you promised.

Why it hasn’t been solvedA forecast can score well overall and still claim far more certainty than it has: on crop-yield data, uncalibrated prediction intervals covered only around 40 percent of actual outcomes, and a simple conformal correction lifted that to roughly 81 percent. Naive forecasts overstate their confidence, and growers are not getting the fix.

Why AI changes the equationConformal prediction and calibrated uncertainty quantification wrap a forecast in intervals that reflect how often the model is actually right, giving an honest range instead of an overconfident point estimate.

What’s still requiredCalibration methods deployed on real forecast pipelines, held-out validation tests against recorded outcomes, and adoption by growers.

The payoffCommitments backed by honest ranges, so a farm knows when a forecast can be trusted and when to hedge.

What’s brokenA model trained on early data quietly stops matching reality.

Why it hasn’t been solvedA forecast trained only on early-crop data drifts out of date as sensors drift, logs get skipped, and predictors that never truly drove the outcome get weight anyway, so the model silently decouples from what the crop is actually doing.

Why AI changes the equationOnline and continual learning can re-estimate the model as new data arrives and as inputs degrade, distinguishing genuine signal from drift and re-weighting or dropping predictors that stop mattering.

What’s still requiredDrift-aware monitoring, an online update pipeline that tolerates missing logs, and validation that continuous updates improve rather than destabilize forecasts.

The payoffForecasts that stay accurate through the whole season instead of decaying after the first few weeks.

What’s brokenMost yield models are built for lettuce. Real farms grow a mix.

Why it hasn’t been solvedAlmost all public yield studies model a single lettuce crop, and the leading models make no cross-crop generalization claim, so a model that spans a real mix of greens is rare even though mixed production is what many farms actually run.

Why AI changes the equationRepresentation learning and transfer learning can learn shared structure across crops and adapt to a new green from limited data, instead of requiring a separate model per species.

What’s still requiredMulti-crop labeled datasets, transfer methods that hold across greens, and validation on real mixed plantings.

The payoffOne forecasting approach that covers an entire mixed-greens operation rather than a single species.

What’s brokenCompanies promise digital-twin savings nobody has independently checked.

Why it hasn’t been solvedVendors advertise 25 percent energy savings and 30 percent less crop loss with no independent audit, while a peer-reviewed review of the field found that only about 9 percent of the articles it examined actually virtualize the plants themselves. The missing piece is not the model, it is a way to prove the model works.

Why AI changes the equationA digital twin can be made auditable by versioning its inputs, predictions, and observed outcomes, so any claimed saving can be traced back to the data and model version that produced it.

What’s still requiredStandardized audit protocols, versioned data and prediction pipelines, and independent validation against a real operation's measured results.

The payoffDigital twins whose savings can be verified rather than asserted, so growers can tell a real deployment from marketing.

What’s brokenHarvest schedules run on a single number that is always wrong.

Why it hasn’t been solvedThe mathematics of scheduling is solved: mixed-integer linear programming, constraint programming, and reinforcement learning all work, but most published plant-factory schedulers still optimize against a fixed yield number rather than a calibrated probabilistic forecast, so they plan against a figure that is always wrong.

Why AI changes the equationOptimization can be run directly against a probabilistic forecast, so a schedule is planned across the range of likely yields and hedged for the tail outcomes instead of betting on one number.

What’s still requiredIntegration of calibrated forecasts into scheduling engines, plus validation that probabilistic plans beat fixed-number plans in operation.

The payoffFewer missed orders and less waste, because the schedule already accounts for the uncertainty in the crop.

What’s brokenWe know the environment changes how food tastes. Nobody has mapped it.

Why it hasn’t been solvedLight, conductivity, temperature, and stress each shift a crop's volatiles, sugars, and phytonutrients, but there is no public model that predicts a cultivar's chemotype from the climate recipe it was grown under, so flavor cannot be designed in advance.

Why AI changes the equationAI can learn the environment-to-chemotype mapping from controlled-condition trials and assay results, working backward from a desired taste to the climate recipe that produces it, effectively a search engine over growing conditions.

What’s still requiredPaired datasets of controlled conditions and measured chemotypes, models validated across cultivars, and assay throughput to generate the training data.

The payoffProduce grown deliberately for flavor and nutrition rather than for yield alone, opening a quality premium that can be designed and repeated.

What’s brokenA new food-safety law expects lot records from a farm that never harvests in lots.

Why it hasn’t been solvedThe FSMA Section 204 traceability rule requires lot-level records at critical tracking events. The original compliance date was January 20, 2026; FDA has proposed, not finalized, a 30-month extension to July 20, 2028, and a separate congressional measure restricts enforcement before that same date. Standard tooling still assumes a field harvest with discrete lots, not the continuous harvest of a recirculating system.

Why AI changes the equationData modeling and record linkage can represent continuous harvest as traceable lots, tying each event back through the shared water and growth history so records stay complete and auditable.

What’s still requiredData models built for continuous harvest, integration with on-farm records, and validation that generated records satisfy the rule's critical tracking events.

The payoffCompliance without rebuilding the farm around field-style lots, and faster, more precise recalls when something goes wrong.

What’s brokenWe promise every batch tastes the same. Nothing measures the taste.

Why it hasn’t been solvedControlled-environment farms sell consistency, but the closed-loop control that exists regulates nutrients and pH, not flavor chemistry, so batch-to-batch taste is not measured and not controlled from one cycle to the next.

Why AI changes the equationAI can predict flavor from environment and harvest records and close the loop on measurable flavor proxies, so a farm steers toward a target taste instead of hoping the recipe delivers it.

What’s still requiredPer-batch flavor measurements, models linking environment to taste, and control logic that can act on those predictions.

The payoffConsistent taste batch after batch, less rejected product, and a consistency claim backed by measurement rather than assumption.

What’s brokenThe big 'more nutritious' claim compares two different things.

Why it hasn’t been solvedThe widely quoted '4 to 40 times more nutrients' claim comes from USDA-funded work comparing microgreens with mature leaves of the same species, not soil with hydroponics. The two USDA write-ups, one at 4 to 40 times, one around 5 times, are two framings of that single dataset, a range across species and nutrients versus an average. The soil-versus-hydroponic question is separate, with mixed results.

Why AI changes the equationAI tied to real assay measurements can mine the nutrition literature and cross-reference claims against their underlying datasets, so a marketing number is matched to the comparison it actually supports.

What’s still requiredStandardized nutrition assays, cross-study data with matched comparison conditions, and models validated against measured nutrient content.

The payoffNutrition claims that survive scrutiny, protecting consumer trust and letting genuine quality differences be stated honestly.

What’s brokenHow long greens stay fresh is decided while they grow. Nothing predicts it.

Why it hasn’t been solvedShelf life is largely determined before harvest, by nitrogen, light integral, and water status, but no model predicts a batch's days-to-degradation from the conditions it was grown under, so freshness potential is unknowable until the product is already failing.

Why AI changes the equationAI can learn the link between pre-harvest conditions and post-harvest outcomes, turning grow records into a predicted shelf life for each batch before it ships.

What’s still requiredPaired pre-harvest and post-harvest datasets, models validated against measured degradation, and integration with packing and logistics decisions.

The payoffBatches matched to the distance and time they can survive, reducing spoilage and letting farms promise realistic freshness.

What’s brokenOne contaminated water loop can threaten an entire crop.

Why it hasn’t been solvedIn a recirculating system a single contamination event can reach every plant on the shared loop. Research shows treatment technologies such as UV-C can reduce human bacterial pathogens in nutrient solution under experimental conditions, but a demonstrated reduction is not a guarantee of elimination in every operating system, and any treatment must avoid damaging plants or disrupting nutrient chemistry.

Why AI changes the equationAI cannot replace sanitation. It can combine microbial monitoring and water-system data to flag contamination risk earlier, target which parts of the loop are most likely affected, and optimize when treatment runs.

What’s still requiredEffective, validated water-treatment technology, food-safety validation and regulatory evidence, plus detection systems feeding the risk models.

The payoffSafe produce from recirculating systems, fewer recalls, and confidence that a contaminated loop can be remediated rather than cleared out.

What’s brokenWashing cannot reach a pathogen that is already inside the leaf.

Why it hasn’t been solvedPathogens internalize from contaminated solution into edible tissue, and washing acts only on the surface. What is inside the leaf stays there, and a reduction on the surface is not elimination, so a wash-free guarantee needs proof that internalization never happened.

Why AI changes the equationImaging and risk models can help detect and flag tissue likely to harbor internalized pathogens, but the safety proof itself depends on microbiology and food-safety evidence, not on AI.

What’s still requiredDetection methods for internalized pathogens, microbiological and food-safety validation, and regulatory evidence to support wash-free claims.

The payoffWash-free products justified by evidence rather than assumption, protecting consumers and upholding the safety promise of indoor growing.

What’s brokenMicrogreens are premium because of hand labor, and hand labor is why they cannot scale.

Why it hasn’t been solvedMicrogreens turn roughly 27 crop cycles a year on a budgeted 13-and-a-half-day cycle, but labor is the dominant input: extension budgets put it at about 15 minutes per tray, with other named-task studies ranging lower, so the premium only survives at hand-labor scale.

Why AI changes the equationTask-specific automation combines computer vision to locate and grade delicate shoots with robotics to handle them, targeting the fiddly steps that make microgreens labor-heavy rather than treating them like head crops.

What’s still requiredRobots and end-effectors gentle enough for microgreens, validated automation of the highest-labor tasks, and an economic case for each deployed step.

The payoffMicrogreens that scale beyond hand labor, lowering the labor cost that keeps a premium crop small.

What’s brokenWe can measure flavor in a lab. We cannot grade every harvest for it.

Why it hasn’t been solvedBrix, dry matter, and volatiles are the measurable proxies for flavor, but measuring them today means lab instruments, mass spectrometry for volatiles, refractometry for Brix, gravimetric or near-infrared methods for dry matter, or a human taste panel, neither of which runs fast or cheap enough to grade an entire harvest.

Why AI changes the equationFast sensing paired with calibration and prediction can estimate these quality proxies at line speed, using spectral measurements and models trained against lab reference values so every batch, not a sampled few, gets a quality read.

What’s still requiredFast spectral or near-infrared sensors, calibration datasets tied to lab measurements, and models validated to grade at harvest speed.

The payoffQuality graded on every harvest rather than a sample, enabling sorting by flavor, consistent product, and quality claims backed by measurement.

What’s brokenNobody can tell you what farm automation actually saves.

Why it hasn’t been solvedThe labor figures behind automation payback models are estimates that rarely get checked against measured work. No standard tool measures pounds-per-labor-hour per activity, so a farm cannot compute the payback of automating this step versus that one from its own numbers.

Why AI changes the equationAI could turn routine production data into measured labor intensity per activity and compute payback step by step, replacing the farm-specific assumptions that CEA cost models currently rest on. It is operations analytics, not a biological fix.

What’s still requiredStandardized, per-activity labor-time data collected across farms and validated cost models that hold beyond a single operation.

The payoffFarms would know which automation actually pays back before they buy it, avoiding capital spent on steps that save little and directing it to the steps that save the most labor.

What’s brokenWe plant a seed and do not know if it will grow until weeks later.

Why it hasn’t been solvedA seed lot is a gamble on germination that is only revealed after the space is already spent. Standard tests score first counts at about 4 to 7 days and final counts out to about 21 days depending on species, so weak lots occupy channel slots before anyone knows they are weak.

Why AI changes the equationComputer vision on repeatable seedling images could score germination and vigor far earlier than a grow-out, flagging poor lots before they take up a slot. It shifts the check from a destructive multi-week test to an image taken in place.

What’s still requiredLabeled seedling image sets paired with ground-truth germination outcomes across species, plus scoring models validated against standard test results.

The payoffWeak lots get caught before they consume channel space and labor, improving crop uniformity and cutting the waste of growing out seed that was never going to perform.

What’s brokenThe recipe for a perfect crop exists only inside private companies.

Why it hasn’t been solvedThousands of cultivars exist, but no open database maps cultivar plus spectrum plus EC, pH, and temperature to time, yield, and flavor. That knowledge is held privately, so every operation re-derives the same recipes from scratch.

Why AI changes the equationA searchable database could learn cultivar-by-environment responses from trials and predict recipes for untested combinations instead of running every one blind. It generalizes from recorded outcomes rather than replaying fixed recipes.

What’s still requiredShared trial data across cultivars and environments, standardized records of conditions and results, and models trained across many farms so predictions transfer.

The payoffFaster recipe development and less duplicated trial-and-error across the industry, so a cultivar matched to the right environment reaches yield and flavor sooner.

What’s brokenAutomated transplanting exists. A cheap retrofit for a small rack does not.

Why it hasn’t been solvedAutomated transplanters exist in large purpose-built facilities, but no affordable machine retrofits onto a small NFT rack, so on a small farm every seedling is still moved by hand. The task is small, repetitive, and repeated for every single plant.

Why AI changes the equationPerception and motion planning could let a retrofit arm find each plug and place it at spacing, adapting to a rack that was never designed for a machine. It does not remove the need for a robot that can physically handle plugs; that hardware is the blocker.

What’s still requiredA low-cost retrofit manipulator that can grip and place plugs reliably, plus real-world validation on operating racks.

The payoffRemoves a repeated manual touch from every plant's life, cutting labor per plant and freeing hand labor for work machines cannot yet do.

What’s brokenOne dirty channel can infect an entire farm. Cleaning it is still by hand.

Why it hasn’t been solvedBetween cycles, channels carry root mats and biofilm, and because the loop is shared, one dirty channel risks the whole system. Hand-washing is slow and leaves no proof that the channel is actually clean.

Why AI changes the equationVision could grade a channel after cleaning and confirm that no root debris or biofilm remains, turning an unverifiable manual step into an auditable check. It verifies cleanliness; the physical removal of debris still needs a mechanism.

What’s still requiredImaging hardware that works inside channels, labeled clean and dirty reference images, validated detection models, and a cleaning mechanism for the physical scrubbing.

The payoffFewer loop-wide infection events, less manual scrubbing labor, and auditable proof of clean lines before the next cycle starts.

What’s brokenHarvest robots work in giant farms. Retrofitting one anywhere else is unsolved.

Why it hasn’t been solvedWhole-facility systems demonstrate harvest automation: SPREAD's Techno Farm Keihanna was designed for 30,000 heads per day and reported a 99% operating rate as of January 2021 (heads harvested against maximum capacity, not uptime), and SPREAD transferred farm operations to a Tokyu Land subsidiary in April 2025. Iron Ox laid off about half its staff in November 2022 and rebranded to Inevitable Tech in 2023. What is missing is not harvest robots in general but a reliable, affordable retrofit for a small rack.

Why AI changes the equationPerception and planning could let a retrofit cutter locate and sever whole heads at the right point. But harvest automation is a physical integration problem: a machine must fit a small rack, and AI does not remove that hardware gap.

What’s still requiredA retrofit mechanical harvester and end-effector hardware that fit a small rack, plus extensive real-world testing on operating farms.

The payoffRemoves the single largest manual labor step in a leafy-green operation and enables harvest without a hand crew, at a scale far smaller than the purpose-built facilities that exist today.

What’s brokenPlants get carried up and down the racks by hand, over and over.

Why it hasn’t been solvedEvery plant is moved several times, and on a rack most of that means lifting and carrying. Autonomous mobile robot navigation is commodity and open-source systems exist, but the gap is retrofit integration with an in-house 4-tier rack, not the navigation itself.

Why AI changes the equationPerception and motion planning could let a retrofit lifter or mover place trays into a rack whose geometry was never standardized for machines, adapting to each rack instead of requiring a purpose-built facility.

What’s still requiredA retrofit lifting and handling mechanism, integration hardware for the rack, and validation on real operating racks.

The payoffRemoves repeated lifting labor from every plant's life and reduces both labor hours and the injury risk that comes with carrying trays up and down stacks.

What’s brokenTo breed a better indoor crop, you measure every plant by hand.

Why it hasn’t been solvedIndoor breeding is meaningless without measuring how each cultivar responds to a light recipe, and that phenotyping is manual or research-only. It does not scale inside a working farm.

Why AI changes the equationAutomated imaging plus models could measure cultivar response to environment during normal production, turning routine growing into phenotyping data. It replaces hand measurement with repeated, consistent capture.

What’s still requiredCalibrated imaging, standardized measurement protocols, and labeled datasets linking genetics and environment to observed outcomes.

The payoffFaster selection of cultivars matched to indoor systems, so better-suited varieties reach yield and consistency without years of manual measurement.

What’s brokenPicking one leaf at a time without bruising it stumps every robot.

Why it hasn’t been solvedSelectively picking individual marketable leaves without damaging the crown is where robotics fails completely. It combines fine manipulation, occlusion of leaves by other leaves, and a damage tolerance that no fielded machine yet meets.

Why AI changes the equationVision could locate and rank leaves by maturity through occlusion and planning could sequence the picks, solving where to cut. The physical act of grasping one wet leaf without crushing it is the wall, and that is hardware, not perception.

What’s still requiredDexterous end-effectors that can handle wet leaves without damage, occlusion-tolerant perception, and testing on real crops.

The payoffHarvest only what is ready, raising marketable yield, cutting waste from damaged leaves, and reducing hand labor at the most delicate step.

What’s brokenRobots crush wet, fragile leaves. The gentle hands do not exist yet.

Why it hasn’t been solvedHolding a wet, fragile leaf without crushing it is a materials-and-control problem that sits upstream of harvest automation. Prototype grippers now handle leafy greens in trials, but none reach commercial reliability on a farm.

Why AI changes the equationSimulation and iterative control could tune grip force and contact for each leaf, letting soft grippers adapt instead of applying a fixed squeeze. It improves control and tuning, but the gripper's physical materials remain the constraint.

What’s still requiredBetter soft materials, force and tactile sensing on the gripper, and validation across real harvest conditions rather than lab trials alone.

The payoffUnlocks harvest automation for delicate greens, reducing crop loss from handling damage and the hand labor needed for gentle harvesting.

What’s brokenCloning the best plants would fix inconsistency. It is still hand labor.

Why it hasn’t been solvedFor premium slow-from-seed herbs, clean clonal propagation would fix genetic inconsistency, but micropropagation is labor-dominated and automation is hard to scale. Tissue culture is primarily a biological production process, not an AI-compute problem.

Why AI changes the equationVision could grade and select explants and monitor cultures for contamination and growth, cleaning up the manual inspection steps. It assists the process but does not change that the core work is biological handling under sterile conditions.

What’s still requiredAutomated sterile handling hardware, culture-monitoring sensors, and validation at production scale.

The payoffGenetic consistency across crops and a uniform premium product, without the labor intensity that keeps clonal propagation out of reach at scale.

What’s brokenBuilding a vertical farm costs up to ten times what a greenhouse does. Running one costs up to five times as much.

Why it hasn’t been solvedA centralized vertical farm carries up to ten times a high-tech greenhouse's capital cost per square meter of growing space, and up to five times its operating cost.

Why AI changes the equationCapital intensity is a structural business constraint, not a computation problem. AI can squeeze operating cost and yield from a built farm, but it cannot remove the capital required to build one.

What’s still requiredValidated operating models showing which automation and control actually pay back, and financing structures matched to the real payback period.

The payoffLower capital per unit of food would mean more farms can be built and more of them survive, which is the difference between a niche and an industry.

What’s brokenWe can grow lettuce almost anywhere. Making it affordable is another problem.

Why it hasn’t been solvedAs of 2025, a lit stacked farm's cost per head is above a field or greenhouse head for commodity greens, driven by electricity, capital and labor. That gap varies with energy prices, capital cost, utilization and market, so it is a current disadvantage, not a permanent impossibility; ventures that sold purely on price lost.

Why AI changes the equationThe price gap is structural and AI cannot close it. What AI can do is protect the premium: consistent quality, flavor, and freshness that justify a higher price.

What’s still requiredQuality and freshness models tied to real batch data, and market validation that the premium holds.

The payoffWinning on flavor, freshness, and locality instead of price is what lets indoor greens survive against field lettuce.

What’s brokenWe grow lettuce brilliantly. We still can't grow calories.

Why it hasn’t been solvedModeled indoor lettuce yields are approaching a near-term ceiling of about 230-234 kg per square meter of growing-layer area per year, reached by optimizing plant density rather than by adding light (Meeuws et al. 2025). That is a model-based target under stated assumptions, not a proven biological maximum. Beyond leafy greens, microgreens and strawberries, no calorie crop has demonstrated vertical economics.

Why AI changes the equationDigital twins and automated high-throughput trials can test how new crops and genetics respond to indoor conditions far faster than manual research, narrowing the space to what is worth building.

What’s still requiredValidated crop models, automated trial systems, and testing across full crop cycles for crops beyond lettuce.

The payoffBreaking the lettuce-only ceiling would let indoor farms grow crops that actually feed people, instead of only premium greens.

What’s brokenEvery farm records data its own way, so nothing compares.

Why it hasn’t been solvedThere is no agreed schema for recording crop, sensor or yield data, so farms aren't comparable and everyone rebuilds analytics from scratch. The relevant standard, ISO/AWI 34713-1, Controlled environment agriculture systems, Part 1: General concepts and control framework, is a new work item still under development under ISO/TC 347, not yet published or adopted.

Why AI changes the equationShared standards are an enabling layer: once data is recorded the same way, models can be trained across many farms instead of one, and results become comparable.

What’s still requiredAgreement on a common data schema, adoption across farms, and the reference datasets that make cross-farm models possible.

The payoffA common language for farm data would let every farm benefit from what others learn, instead of each one starting over.