Superintelligence for Coffee Roasters: Lots, Subscriptions, Brew Help

Superintelligence for Coffee Roasters: Lots, Subscriptions, Brew Help

How coffee roasters can log roast lots, handle subscription orders and answer brewing questions with AI, while the roaster keeps quality decisions.

•

Want to apply this in your company?

Book a call with Juan →

Superintelligence for coffee roasters, in practice, is not a system that decides the roast profile. What a roastery can run safely are narrow agents and automations: logging lots with origin and roast date, answering recurring subscription requests, and answering brewing questions from recipe cards the roaster has approved. Quality decisions stay with the person who cups the coffee. AI handles recording, alerts and handoffs, with context.

Key takeaways

  • The lot is the unit of control. Origin, process, roast date and profile must be recorded before any order ships.
  • Subscriptions need written rules for cycle, pause, coffee swaps and delivery failures, so the agent applies them without improvising.
  • Brewing questions can be answered by AI only from approved recipe cards, with a handoff when the case falls outside the script.
  • Releasing a lot, changing a roast curve and pausing a coffee are human decisions, with a record of who decided and why.
  • The sensible path is to start with one flow, with an agreed metric and logging, and expand only afterward.

Why the lot is the right unit of record

When a subscriber says this month's coffee tasted more acidic than the last one, the useful question is which lot they received, when it was roasted and on which curve. Without that answer, the conversation becomes guesswork. That is why the lot record should be the starting point for any automation, not a side detail of inventory.

A minimum record usually includes:

  • lot and harvest identification, with the origin reported by the supplier;
  • green coffee process and the date it arrived at the warehouse;
  • roast date, the person who roasted it, and the profile or recipe used;
  • cupping notes recorded by the roaster, including what was approved and what remained in doubt;
  • lot status: available, resting, reserved for subscription, or suspended.

At the start, this record can be a well-structured spreadsheet. What matters is that every order and every message can point to a specific lot. If the support agent only knows that a customer receives the subscription coffee, it cannot give a useful answer.

How to build the record step by step

  • Define the required lot fields and write down what each one means, so the people roasting and the people answering customers use the same language.
  • Create a unique identifier for each lot and make it appear on the bag and the order.
  • Decide who can change a lot's status, and record every change with a date and a name.
  • Link the subscription order to a released lot, so the system does not send a suspended coffee.
  • Document what the agent can read (status, recipe cards, subscriber history) and what it must never touch (cupping notes, recipe changes).
  • Review the record once a week with the roaster and the support team, noting what was missing when a customer needed an answer.

Recurring orders: what can be automated

A subscription is a process of dates and rules. Delivery cycle, pauses, coffee swaps, address and card updates all change what leaves the roastery and what logistics has to do. The deterministic part, such as calculating the next shipment, blocking a suspended lot or announcing the dispatch date, should live in a fixed flow with written rules. AI comes in to understand requests in natural language, such as pausing only next month or switching to a lighter coffee, and to ask for whatever is missing.

Any action that changes billing, plan or stock needs explicit confirmation and, where appropriate, human approval. The agent should not promise refunds, discounts or lot swaps on its own. Proactive shipping notices depend on customer permission and approved message templates, as explained in our article on WhatsApp outbound messages.

Brewing questions without improvising

Brewing questions are the most repeated ones, and they often reveal product problems. Which grind for a French press, why the cup tasted bitter, how much coffee to use for two cups. The answer should come from a brewing card approved by the roaster, separated by profile and by method: grind size, dose, approximate water temperature, extraction time and a short paragraph on what the cup should taste like.

The agent answers with the card and does not invent recommendations. When the customer describes something outside the script, such as a strange taste, a defective bean or a bad reaction, the agent collects the basics (lot, method, equipment and what was noticed) and passes the conversation to a person with a full summary. The customer does not have to repeat the story.

Illustrative scenario: a complaint about taste

Illustrative scenario: a subscriber replies on WhatsApp that this month's coffee came out more acidic than usual and asks whether they should adjust the grind. The agent identifies the lot from the order, confirms the roast date on record and sends the brewing card for the profile and the method the customer mentions, without promising anything about the lot. It asks about the method and equipment, logs the conversation and creates a handoff to the roaster with the lot, the report and the card that was sent. If several subscribers complain about the same lot in the same week, the system flags the person responsible for release, who decides whether the lot stays in circulation, whether the grind recommendation changes, or whether the lot leaves the shipping queue.

Where the roaster's decision must stay

Some decisions should not be delegated to an agent:

  • releasing or suspending a lot after cupping;
  • changing the roast curve or the recipe of a blend;
  • deciding what changes in a brewing card;
  • deciding compensation, replacement or reshipment for a defect;
  • classifying a coffee as available for subscription.

The system can suggest, gather evidence and remind the team of deadlines, but the decision stays with a person and is recorded in the history. Subscriber data, such as name, address, phone number and purchase history, is personal data. Limit the agent's access to what each task requires, and follow the guidance of Brazil's data protection authority, ANPD. When an automated decision affects a customer, our article on automated decisions and the right to review helps define the path.

How to apply the method in a roastery

The method we use has four steps. First, map the current flow, showing how a lot becomes an order today and where information gets lost. Second, limit data and actions: the agent reads status and cards but does not change billing or recipes. Third, pilot one real flow, such as brewing questions for a single coffee line, with every conversation logged and an agreed metric defined before starting. Finally, move what worked into the team's routine and decide whether the next step is to expand, adjust or stop. When the operation needs ongoing support, maintenance can be handled by Be Human. For the leadership team, our superintelligence glossary helps separate what is a promise from what is an operation.

Frequently asked questions

Can superintelligence decide the roast profile for a roastery? No. Profile, lot release and recipe are decisions for the person who roasts and cups the coffee. AI systems can record data, organize notes and flag patterns, but responsibility for what ends up in the cup remains human.

Where should a roastery start if it still uses spreadsheets? Start with lot logging and one small flow, such as brewing questions for a specific coffee. Define the metric before the pilot, for example the time until an answer with an approved card is sent and the share of correct handoffs, and log every conversation. Only then decide whether it is worth extending the flow to subscriptions.

Can the agent promise a replacement or refund when a coffee arrives defective? It should not. Replacements, refunds and compensation depend on a written policy and a human decision. The agent collects the lot, date, report and photos, and passes the case on with a summary, which protects both the customer and the team.

Do you need a sophisticated platform to start? Not necessarily. Many operations begin with structured records, simple forms and the messaging channel they already use, expanding the infrastructure only when the pilot shows what needs to be automated.

Sources

coffee roastingAI agentssubscriptionsWhatsApplot tracking