AI in agroexports: essential knowledge to empower teams and the company (part I)
Column published in Spanish on Agraria.pe on 15 July 2026.
For years we were promised that technology would transform agriculture. Management software arrived, then satellite imagery and precision agriculture, then Internet of Things sensors and drones; yet adoption has been slow. To a large extent, Silicon Valley specialised in agtech products but struggled to understand such a complex business, where every crop, area, variety and climate behaves differently and no season repeats itself exactly.
Is artificial intelligence (AI) just one more word on that list? No, for a concrete reason. Its greatest advantage is not that of a new device, but that it puts in the hands of farming teams themselves the ability to assess and solve many of the problems they face. Perhaps more importantly, it can become the interface that finally brings together three worlds that almost never understand each other: those who build the technology, those who use it and those who need it.
The real risk for management is not being left out, but not understanding the technology well enough to direct it. The manager does not need to execute it, but does need to know what it can do, what its limits are and how to train teams, both to strengthen them and to automate the routine so they have more time to apply their judgement.
In that sense, management must prioritise, because AI makes more initiatives viable than can be tackled at once and, without clear direction, efforts scatter. It must also govern, protecting the confidentiality of data and ensuring traceable processes, so that decisions rest on verified results and not on answers that sound convincing but are wrong. This matters even more when many employees are already using AI on their own.
AI today is a label that mixes everything together, and much of what is sold as AI is actually automation. That is why, before deciding what to use or what to build, it helps to see the full automation map of a company and locate where AI really sits.
How a company and AI interact
For years, companies have used traditional software, such as ERPs and irrigation and fertigation controllers. On the administrative side there are also RPAs (Robotic Process Automation) and software bots, which repeat fixed tasks, such as downloading electronic invoices from the tax authority's portal every day and attaching them to the system, or taking the XML of those invoices and loading the fields (tax ID, amount, date and description) into the ERP. These tools operate with exact rules and formulas written by people; they do not learn on their own.
On top of this base comes artificial intelligence, which, unlike the above, can learn patterns from data. Machine learning finds relationships and generates estimates. For example, it can anticipate the flowering of a block from climate variables. What makes it machine learning, and not a simple growing-degree-day formula, is that the relationship is not fully defined in advance; the model estimates it from multiple seasons. To incorporate a new season, its data must be integrated and the model retrained or recalibrated.
Deep learning is a subfield of machine learning based on deep neural networks. It can process structured and unstructured data, although it is especially useful for images, audio and text. In the packhouse, both size measurement and defect classification can be solved with different computer vision techniques; when deep neural networks are used, that is deep learning. These systems can classify, fruit by fruit and at high speed, defects such as bruising, rot, sunburn or softness. Another example is the model trained by Peru's Ministry of Agriculture to identify 11 crops from satellite imagery through their spectral signatures.
Generative AI is based, in most current systems, on deep learning models capable of generating content (text, reports or code) and operating in natural language. Its ability to interpret colloquial expressions comes from training on large volumes of text from many sources, not only chat conversations. This category includes Claude, ChatGPT, Gemini and Copilot, among others, and it is the gateway through which many other capabilities are used today.
Above those capabilities there is not necessarily a higher level of intelligence, but an orchestration and agentic execution layer. While a generative model can write a report, an agent combines the model with tools, instructions, memory, data and permissions to execute a sequence of steps towards a goal. It can decide which action to take, which tool to use and when to review or correct a result, without asking for authorisation at every iteration. That autonomy is its greatest power and, at the same time, its main risk. That is why its scope is limited, boundaries are set and, in sensitive decisions, a human control point is kept. When several agents coordinate under a supervisor, which may itself be automated, we speak of a multi-agent system.
The foundations of generative AI
Until recently, each of these layers required specialists: ERPs and bots needed programmers; models and simulations, professionals with a command of statistics and mathematics; reports and visual pieces, analysts and designers. Generative AI now allows many of those capabilities to be gathered at a single point of access: natural language, without needing a programming language or writing code. Today, a person with no technical training can describe what they need, in writing or by voice, and obtain results that used to require an expert team. It is worth understanding how it works inside, because that is where both its power and its limitations come from.
Large language models (LLMs) are among the most advanced families of AI, and this is where developments such as Claude, ChatGPT, Gemini, DeepSeek and Grok sit. One of the central tasks used to pre-train the model underlying OpenAI's ChatGPT is predicting the next token, which may be a word or part of one. It is similar to a phone's predictive keyboard. When you type "the market rewards size…", the phone may suggest "jumbo" or "large". It does not guess; it calculates, from millions of examples, which continuation is statistically most likely. After predicting one token it predicts the next, and the next, until it builds an answer. After pre-training, these models usually go through additional tuning stages to follow instructions and produce more useful answers. They also consider not only the previous term but the relevant elements of the context. That mechanism is called attention.
The same fruit-size example helps explain how words are represented. Each term can be assigned a vector, a list of numbers learned from large volumes of text. The values that follow are illustrative. At the start, a naive encoding represents no semantic relationship between words. Mathematically, each word has a vector with three components: jumbo (1, 0, 0), extra large (0, 1, 0) and tractor (0, 0, 1). To relate jumbo and extra large you calculate their product: (1, 0, 0) x (0, 1, 0). In Excel this is the SUMPRODUCT function, which multiplies the corresponding components and adds the results: 1×0 + 0×1 + 0×0 = 0. In this initial encoding, every pair yields zero. When the vectors are standardised, greater similarity is usually reflected in values close to 1.
In that initial encoding, "jumbo" is as far from "extra large" as from "tractor". During training, the model processes large volumes of text from different sources and learns that the first two terms appear in similar contexts. It therefore brings their vector representations closer, while "tractor" stays further away.
In the learned representation, the model reflects those relationships through values such as these: jumbo (0.94, 0.31, 0.02), extra large (0.85, 0.52, 0.10) and tractor (-0.26, 0.90, 0.34). These values are also illustrative. Jumbo and extra large end up closer to each other than to tractor. Calculating the product, jumbo and extra large yield approximately 0.96 (related terms), while jumbo and tractor produce a value close to zero (unrelated terms). Real models apply this principle in hundreds or thousands of dimensions and adjust a huge number of parameters through large volumes of data, multiple iterations and later alignment stages.
The model spreads probability across the possible tokens and, depending on its configuration, selects the most likely one or samples among several alternatives. In the blueberry example it may choose jumbo, because that size usually gets a better price. However, the answer might not be valid for avocado. If it lacks sufficient context, the model repeats the most frequent pattern it found in the texts. It optimises the statistical plausibility of the answer; it does not verify by itself that it is true.
That is the limit every manager must remember. AI can express itself with total confidence, use impeccable wording and still be wrong; the industry calls this phenomenon "hallucination". So when an exact figure or a statement supports a decision, one should not rely solely on the tool's memory. It is necessary to provide the data, guide the analysis and request additional calculations, validations and checks. Although computational capabilities are increasingly astonishing, interaction with AI happens through natural language, and there is therefore a gap between what a person considers an obvious instruction and how the tool interprets or executes it.
One example is forwarding emails from one account to another. For a person, copying an email address correctly is a basic task. Yet even after being given the correct recipient, an AI tool can introduce errors when writing it. In one concrete experience, this happened roughly once every 10 or 12 emails sent. The problem was only detected because Gmail bounced the messages. It had been assumed that copying and pasting an address was a task where AI could not fail, but the tool does not always follow the same reasoning or procedure a person would use.
Another example appears when preparing a report with AI support. It may be necessary to explicitly tell it to use the latest version of the document available in the folder, especially when manual adjustments were made afterwards. The tool may tend to work with the file it previously generated or saved, or with the version it keeps in the temporary context of the session, even if it is not the most recent. This can happen even when it has been instructed to always take the latest available version. As with any skill, only frequent use makes it possible to understand these limitations, improve the instructions and progressively master the tool.
Understanding these foundations is essential, but not sufficient. The next question is how to turn these capabilities into a concrete way of working inside an agroexport company. Part II explains, using Claude as an example, how to move from a simple conversation to the use of context, reusable skills, artifacts and agents, and presents an applied case of cost benchmarking.
