AI in agroexports: essential knowledge to empower teams and the company (part II)
Column published in Spanish on Agraria.pe on 5 August 2026.
"I'm glad I was wrong. I thought that by now there would have been a bigger impact on entry-level office jobs." Sam Altman (CommBank Accelerate AI, 26 May 2026)
Part I explained, in simple terms, what lies behind generative AI, how it relates to other forms of automation and why it can be wrong even when it answers with apparent confidence. Now we will explain how Claude is used and walk through a practical case.
How Claude is used: chat, skill, artifact and agent
Understanding how the model works is only half the story; the other half is how to put it to work. A loose chat (ask a question, get an answer) is barely the entry door. The real value appears when you give it context, data and method, and when you let it not only answer but execute.
To show this concretely, Claude serves as the example, although the essentials apply equally to ChatGPT, Gemini or other tools. The path goes from the simplest to the most powerful: from chat to skills and agents.
Chat. This is the entry point and the one almost everyone already knows. The person writes a question or an instruction (a prompt) and Claude answers. It is useful for writing an email to a client, summarising a season report or explaining a regulation in a destination market. The recommended structure is to define the role you want it to play, explain the context it needs to know, specify what it must achieve, indicate how it should do it, point out which information it should analyse and, finally, set how it should present the answer.
Giving it context. The next step is to stop copying and pasting. Depending on the product, the plan, the connectors enabled and the permissions granted, the tool can access desktop folders, Dropbox or Google Drive and work with several files at once. Instead of uploading them one by one, you can point it to a directory with forty agrochemical trials and ask it to synthesise the conclusions on one page. A permanent context can also be configured, from the company's activity to its definitions, indicators and manuals, so that it answers on the basis of the organisation's reality rather than generic information from the internet.
The skill. In Claude, a skill is a reusable folder or package that can contain instructions, scripts, templates and other resources. It can incorporate, for instance, the protocol used to evaluate a variety, refined over five seasons; the exact way the budget is prepared; or the procedure for preparing the documentation of an ISO certification or calculating the carbon footprint. The skill does not guarantee that Claude will repeat the process identically, because the behaviour remains probabilistic, but it allows the procedure to be applied more consistently. A well-designed skill can also flag what information is missing. It is also a way to extract judgement from a single person's head and turn it into institutional memory.
The agent. An agent combines an AI model, tools, data, permissions and procedures, possibly including one or more skills, to execute a sequence of steps towards a goal. It can review what it obtained and correct the process before concluding. Scheduled or continuous execution also requires infrastructure, triggers and specific permissions. Its autonomy is its greatest power and its main risk, so its scope is limited and its output is audited. When several agents coordinate under an automated supervisor, a multi-agent system is configured. For example, an agent that was programmed and runs independently can search the news at 8 a.m. every day for the impact of El Niño on Peruvian export crops and send an email with the summary and links of interest.
Artifacts. Much of what Claude builds takes the form of artifacts: pieces that appear in a panel next to the conversation, such as documents, charts, calculators or dashboards, which the user can review and adjust without touching code. For a manager, one of the most useful cases is a dynamic dashboard. By sharing yield, cost and fruit-size data by block, an interactive board can be generated in which a filter or an assumption changes every result; for example, a contribution calculator by block that identifies negative values, even over an image of the farm when a KML or KMZ file is available. When a figure must be exact, the tool can write and run code to make the calculation more traceable and reproducible, although the logic, the formulas and the result must still be validated. What used to be a request to the IT department can today emerge from a conversation with AI, in which a person or a team takes part in the co-creation and launch of its own products.
Specialised and integrated environments. Claude no longer works only as a chat, but as a set of applications and environments adapted to different types of work. Claude Cowork lets you delegate multi-stage tasks across files, documents and applications; Claude Code is aimed at software development and can analyse repositories, write and test code, run commands and fix errors. Added to these are Claude Design, for creating designs, prototypes and presentations; Claude Science, for scientific research; and Claude Security, for detecting and fixing vulnerabilities in code. Claude for Small Business brings pre-configured connectors and flows into Cowork for finance, sales, marketing and operations tasks. Through Claude Tag, it can join selected Slack channels, accumulate context and receive tasks when mentioned as @Claude. It also works inside Word, Excel, PowerPoint and Outlook, keeping context and moving information between these applications. It can browse web pages through its integration with Google Chrome; for example, searching avocado prices at the main global retailers and organising the results.
Where the gain is: freeing and sharpening judgement
It is worth recovering a distinction once made by Sequoia Capital, one of the most important venture capital firms in the world. Intelligence consists of applying rules, cross-checking data and recognising patterns; it is a complex task, but largely procedural. Judgement, on the other hand, means deciding when a case does not fit the rule, interpreting the context and prioritising with incomplete information. That is why the most intelligence-intensive tasks tend to be automated first, while many of those that today depend on judgement could also become automatable. Even so, changes of context will keep demanding new and better decision criteria.
The main constraint in a farming company is time. Professionals in positions that require the most judgement usually spend a good part of the day on tasks that do not require it. Consider a nutrition specialist who spends hours making sure fertilisers are ordered on time, squaring schedules and filling in forms. None of that fully uses what makes them valuable, and much of that load can be taken on by AI. Their greatest contribution lies in judgement, from reading phenological monitoring and interpreting a deviation to adjusting the plan for a season that is never the same. That capability remains harder to automate.
There is a second effect. AI does not only free up time for judgement; it can also sharpen it. Faced with a relevant decision, such as an expansion, it helps structure it, order the alternatives, run scenarios and sensitivity analyses, and write down the assumptions. It can also act as devil's advocate: you can ask it to critique a fertilisation plan or a season strategy and expose risks, contingencies and second-order effects that had not been considered. It does not make the decision; it forces you to think it through better before adopting it.
AI in action: the cost benchmark case
To show how these technologies combine in practice, here is a conceptual architecture based on a real cost benchmarking process developed by ACM and currently being automated. The greatest value still lies in judgement, backed in this case by more than two decades of experience in this type of analysis. The process requires extracting foreign trade data, harmonising accounts with different charts of accounts or cost structures, and cross-referencing the field context that explains the spending. Instead of one person consolidating spreadsheets for weeks, it can be organised as a multi-agent system, in which an orchestrator coordinates and each component uses the corresponding technique. It is currently estimated that AI is directly involved in around 70% of the activities in the process, although the scope is still being refined and human supervision remains necessary, both for quality control and because of the system's stage of maturity.
Price extraction (automation). A tool similar to an RPA automatically accesses customs portals every week (ACM uses Veritrade), downloads the crop's export data and leaves it cleaned on the servers, without anyone having to download the files. It also generates charts and tables of volumes, prices and FOB values by variety, company, port of shipment and commercial window, among other dimensions.
Cost harmonisation (language and common rules). The system receives accounting information from different ERPs and, drawing on language understanding capabilities, interprets the meaning of each account and proposes its classification under a common standard. To do so it uses historical data and previously classified examples. For instance, it can recognise that "labour transport", "staff mobility" and other descriptions belong to the same category.
The tool does not assign accounts definitively. It formulates proposals based on previous benchmarks, reference examples and established rules, and can even present different alternatives so that an expert confirms the most appropriate classification. Although the result could be shown in a visually sophisticated dashboard, the preference has been an Excel file that details step by step how each classification was made and explains its rationale. This way, a specialist can audit the process, review formulas, reconcile amounts, analyse deviations and validate the criteria used. When a specialist changes a classification, the correction can be stored as a new rule or reference example for the system to consider in later analyses.
To reduce the risk of bias and reinforce confidentiality, the system does not use the identity of the company analysed during classification. In general, the protection of information in these models depends on the provider, the type of plan, the privacy settings, the permissions granted and the contractual conditions. For sensitive information, the organisation must use approved enterprise solutions and expressly verify the policies on the processing, storage and use of information.
Field context (speech to text and language). It processes unstructured information, such as handwritten notes, logbooks and audio recordings the team makes internally after a visit to summarise the main critical points. It transcribes the audio, interprets its meaning and can relate a difference in yield to a specific factor.
Analysis and writing (generative AI). With the cleaned numbers and the context, it writes the executive report in fluent Spanish, explains the causes of the deviations and points out the critical issues.
Orchestrator and auditor. It controls the order, checks that averages are weighted by area rather than simple, and audits the statements in the report. If it detects an inconsistency, it stops the process, orders a recalculation, validates the result and only then gives its approval.
The lesson is clear. None of this works without orderly information and defined procedures; a generic chat given the same task can return inconsistent results. Moreover, because AI expresses itself fluently and generates attractive charts, it can convince you that a result is robust when it is wrong. In one ACM experience, an NDVI dashboard was generated and the procedure was requested to be documented in Excel so that every calculation could be audited. The review showed that the farm average had been obtained as a simple average, not weighted by area. The dashboard looked impeccable, but the result was incorrect. That is why the automated auditor and human judgement are not optional. It is also essential to know the variables and sources used in depth. For example, there are dashboards and infographics built with NASA POWER information on temperatures along the Peruvian coast and in other countries. However, the main risk is not necessarily in the AI, but in not knowing the characteristics of the source. NASA POWER provides meteorological variables at a spatial resolution of 0.5° latitude by 0.625° longitude, roughly 55 by 65 to 70 km on the Peruvian coast, while solar parameters have a resolution of 1° by 1°. In certain locations, a single cell can integrate sea surfaces, coastal valleys and areas with major changes in altitude towards the Andes. If this limitation is not considered, the analysis may attribute to the point under evaluation conditions that do not adequately represent its local reality and lead to wrong conclusions.
Strategy for senior management
Adopting this ecosystem with judgement requires several changes in the decision map.
Document expert judgement. The judgement of your best engineers lives in their heads today and leaves when they move on. Turning that judgement into skills (which signals a farm manager reads, how they interpret a deviation, how they solved past exceptions) is building institutional memory where before there was only individual memory. It is also the foundation without which no AI tool truly performs.
Teams denser in judgement. As routine tasks are automated, the difference between one company and another will increasingly depend on the quality of the decisions taken in exceptional situations. This may lead to more compact teams, but with more experience, a cross-cutting view and the ability to connect dimensions that used to be analysed separately, such as field and finance, production and commercial, or sustainability and compliance. However, a higher density of experience should not mean no longer developing young talent. AI also allows less experienced professionals to learn faster, raise the quality of their analyses and take on responsibilities that used to require several years of experience.
The fallacy of perfect clean-up. It is said that you have to spend years tidying up the ERP before looking at AI. In Peruvian agriculture, waiting for perfection is a recipe for paralysis, and the reality is usually the opposite: adoption is the catalyst that forces the data to be put in order. The sensible strategy is concurrent adoption, with limited pilots whose urgency to work forces the organisation to organise and integrate its foundations.
Eliminate before automating. Before incorporating AI, it is worth critically reviewing the existing process. Many routine procedures, such as inherited controls, warehouse forms or rarely used PDF reports, do not need automation but elimination. Automating an unnecessary activity does not create efficiency; it only increases cost and speeds up the execution of a process that remains deficient. The right sequence is to question every requirement, eliminate what is redundant, simplify what remains and only then automate. A useful question for any organisation would be: how many reports could be eliminated today without affecting decision-making or the operation?
The return to human judgement and systems thinking
Technology will keep evolving and lowering its costs, while many of today's tools will be replaced in a short time. So sustainable competitive advantage will not depend solely on the software a company buys, but on the quality of its data infrastructure, the operational knowledge it manages to structure and put to use, and its ability to connect the different areas of the organisation. In this process, it is essential to think in whole systems and not only in isolated tasks or activities.
AI adoption should not aim at cutting positions or indiscriminately replacing the most experienced professionals. Well governed, integrated and audited, it can take on much of the execution load (filling in forms, number crunching and data consolidation) and free up time for engineers, managers and directors to interpret scenarios, manage exceptions and apply expert judgement to a changing and complex business such as agriculture. Some of these activities may also be progressively automated, but final validation, responsibility and accountability must remain clearly assigned. So the first question should not be which model or tool to use, but which decision we want to improve and which system of information, knowledge and responsibilities supports it. Just as when choosing a blueberry variety, it is not enough to define which genetics to use. The decision depends on the project's location, the climate conditions, the commercial capacity to place the fruit and the productive, financial and operational resources available.
