Coffee Roasters and AI Agents: Wholesale Reorders and Roast Records

An AI agent for a coffee roaster is most useful when it answers wholesale reorder questions from the order history and roast records you already keep, not from memory or guesswork. That requires a defined role, consistent records of accounts, blends and roast batches, and clear limits on what the agent may promise or change. Build those records first, and a stronger agent later can spot reorder patterns that today only a experienced person notices by chance.

Key takeaways

• A reorder answer is only as reliable as the order history behind it, so start by making account names, blend names and quantities consistent.
• Order history records what left the building, while the roast log records what was produced. An agent needs both to judge stock and timing.
• Reorder patterns should be calculated from records and shown with their source, not inferred from a general feeling about a customer.
• Give the agent a narrow role that drafts replies and proposals, with named people approving prices, substitutions and delivery promises.
• Choose one account group and one blend for the first test, and define a success check you can observe in plain terms.

Why wholesale reorders are a memory problem

Most roasters who sell to cafés, offices and shops know their steady accounts well. They remember that one café goes through a house espresso every few weeks, or that a restaurant switches to decaf in the colder months. That knowledge usually lives in the head of the person who answers the phone or the email thread. When that person is busy, away or new to the business, every reorder question becomes slower and easier to get wrong. Wholesale makes this worse than retail because each account has its own cadence, grind, bag size and sometimes a standing agreement about delivery days.

An AI agent does not solve this by being clever. It helps only when the facts it needs are stored in a form it can read and check. Without those facts, it will fill the gaps with plausible text. A plausible answer about a coffee reorder is really an answer about inventory, cash flow and whether a customer's bar will be stocked on Saturday morning, so the cost of a confident guess is higher than it looks.

What a reorder question actually needs

A message such as "can you send the usual to the harbor café?" looks simple. Answering it properly requires several facts at once:

• The account, its legal name, its delivery address and the contact who is authorized to place orders.
• The usual itself: blend or single origin, grind, bag size, and any substitution the customer has already accepted.
• The order history for that account, with dates and quantities, so the normal interval is visible rather than remembered.
• Current stock of that roast, including whether the batch is still inside the window you are willing to sell it.
• Any open quote, credit hold, standing delivery day or unresolved complaint on the account.

If one of these is missing, the agent should say so and route the question to a person. It should not fill the gap. The same discipline applies when a roaster buys green coffee and tracks supplier confirmations, a workflow covered in AI Agents for Procurement: Quotes and Supplier Follow-Up.

Roast logs: the record of what was actually made

Order history tells you what was sold. The roast log tells you what went into the drum and what came out. The difference matters. A batch planned for one blend may be split, partly sent as a sample, or rebagged for a different account. Weight loss during roasting changes with the green lot and the profile, so the weight you bagged is the number that matters for stock, not the number on the plan.

These fields are worth recording on every batch, from the next roast onward:

• Batch identifier, roast date and the person who ran the roast.
• Green lot, origin and the green weight loaded into the roaster.
• Roast profile name, with a version number when the profile changes, and the roasted weight out.
• Roast level reading if you take one, and cupping notes written in plain words.
• Where the batch went: a wholesale account, retail, a sample request, or waste.

A roast log is a reliable source only when people fill it in at the roaster, not afterwards from memory. An agent reading months of inconsistent entries will produce confident summaries of noise. The habit of recording deviations when they happen, and keeping corrections visible instead of overwriting history, is the same one used in AI for Quality Management: Nonconformances and Audit Prep. It also makes the agent's reasoning easier to inspect later. The OpenTelemetry guidance on observability concepts describes this kind of traceable record, in which you can see what happened, in what order, and with which inputs.

A plan to structure your records before an agent reads them

This is the work to do now, whichever tool you choose later. It is mostly administrative, and it pays off even if you never deploy an agent:

• Step one: Write one master list of blends and single origins with a stable name and code. When a blend is retired, mark it retired instead of renaming it, so old orders still point to the right thing.
• Step two: Create one record per wholesale account with the legal name, billing contact, ordering contact, delivery days and the agreed usual.
• Step three: Export the full order history you have, check it for duplicate accounts and mismatched blend names, and correct the most obvious errors.
• Step four: Add the roast log fields described above to every batch, and mark any entries you have to reconstruct from memory so they are not treated as fact.
• Step five: Decide which reorder signals you trust: the usual interval since last order, a standing schedule, or a note from the account owner. Write down which one wins when they disagree.
• Step six: Define approval boundaries before any agent drafts a reply. Prices, credit terms, substitutions and ship dates stay with a named person.
• Step seven: Choose one account group and one blend as the test, and write the success check in plain words, such as the proposed reorder date matching what the owner confirms in most cases you review.

Where the agent stops and a person decides

An agent that reads records and drafts replies can be very helpful without being trusted with commitments. It can compare the last order date with the usual interval, flag a conflict between a planned roast and a sold-out batch, and prepare a reply for review. It should not change a price, offer a discount, promise a delivery date, substitute a blend the account has not accepted, or release a batch that failed a cupping check.

The agent's access should follow least privilege. It needs read access to the account records and roast log, write access only to drafts, and no access to send messages or edit orders directly. Customer and supplier messages should be treated as content to read, not as instructions to obey, because an email can contain text that tries to change what the agent does. The principle of granting only the permissions a task requires, and checking them again as staff and responsibilities change, is set out in the OWASP authorization guidance. Review those permissions on a schedule, not only after something goes wrong.

Illustrative scenario: a reorder that looks routine

Illustrative scenario: A small roaster supplies a neighborhood café with a house espresso. The café usually orders every few weeks, and a message arrives asking for the usual in a larger bag size this time. The agent checks the order history and sees that the interval has shortened. It then reads the roast log and finds the latest batch of that blend is nearly sold through. The agent drafts a reply that offers the next roast date, asks the owner to confirm the larger bag size, and marks any price change for approval. The owner reviews the draft and approves it. The agent promised nothing, and the record shows exactly why it made the suggestion.

The value in this example is not speed alone. A person who already knows the café would have reached the same answer, but the agent makes the reasoning visible, so the next person covering the phone can follow it without reconstructing the history.

How Be Human approaches this, and how to choose the first task

Be Human starts with one useful role, in this case a reorder assistant for a defined group of wholesale accounts. The first engagement defines what memory the agent reads, which tools it may use to look up orders and roast batches, where approval is required, and one observable success criterion. Managed operations then keep those connections working, review permissions when people join or leave, run recurring evaluations that compare the agent's suggestions with what the team actually did, and handle incidents when a record is wrong. The goal is a system the roaster can inspect and adjust, not a black box that answers customers on its own.

If you are unsure which of your roaster's records or reorder questions should become the first task for an agent, take the free Be Human AI diagnostic to find it.

Frequently asked questions

Can an AI agent answer wholesale reorder questions for a coffee roaster today? Yes, for a narrow set of questions, as long as account records, order history and batch stock are readable. Start with drafts and suggestions that a person approves, rather than with an agent that sends orders or confirms delivery dates. The first useful step is usually reliable account data, not a more advanced model.

Do I need specialized roasting software before an agent is useful? No specific product is required to begin. Consistent blend names, one record per wholesale account and a roast log kept in any structured form are enough for a first test. Dedicated roasting or inventory software may simplify the work later, but it does not replace clean records.

How is a roast log different from inventory? Inventory shows how much of each product is on hand, while a roast log shows how each batch was produced and where it went. The roast log explains the difference between the weight planned and the weight bagged, and it records the green lot, profile and any deviation. An agent needs both to judge whether a reorder can be filled from existing stock.

What should stay with a person when an agent handles reorders? Prices, discounts, credit terms, substitutions, delivery promises and release decisions for batches should remain with a named person. The agent can calculate timing, flag conflicts and draft replies for review. Keeping those commitments with people protects the customer relationship and makes every decision traceable to someone accountable.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.