Superintelligence Prep for Coffee Shops: Catering and Supplier Orders
A practical plan for coffee shops to hand catering quotes, supplier reorders and loyalty questions to an AI agent, with owner approval guarding margins.

An AI agent for a coffee shop is most useful in three places: drafting catering quotes from your own price sheet, preparing supplier reorders from stock counts, and answering loyalty questions from the published program rules. The agent gathers details and prepares the paperwork, while the owner keeps approval over prices, commitments and purchases. Those approval rules are what protect margins as the agent improves, because a more capable agent can take on more work without taking on more authority.
Key takeaways - Start with tasks that repeat, follow written rules and produce a result a manager can check in a minute: catering intake, reorder drafts and loyalty lookups. - Give the agent read access to the records each task needs and draft-only access to anything that leaves the shop. - Keep prices, minimum catering commitments and supplier substitutions under owner control until the agent passes tests built from your own past cases. - Write approval rules as risk tiers tied to actions, not to individual people, so they still hold when the agent takes on more work. - Judge improvement against a fixed set of past cases, not by how helpful the agent seems in a demonstration.
Which coffee shop tasks can an agent take on now? Not every task is a good first job. The strongest candidates share three traits: they happen often, they follow rules you can write down, and a person can verify the output quickly. Three categories usually fit those traits. - Catering inquiries, where the agent collects event details and drafts a reply and quote for the owner to approve. - Supplier reorders, where the agent compares stock against par levels, drafts an order, and a person decides whether it is sent. - Loyalty questions, where the agent answers from the program rules and the customer's visit record. - Refunds, comped drinks, charge disputes and custom pricing, which should go to a person with a summary attached.
A useful test is simple. If a new team member could follow the written rule and reach the same answer, the agent can probably follow it too. If the answer depends on a judgment the owner has never written down, write that judgment first, then give the agent the task.
How should a catering inquiry be handled without guessing? Catering requests arrive as messages with gaps. A customer might say the team needs coffee on Thursday, and the shop needs a dozen facts before it can quote. A well-designed agent does not fill those gaps with assumptions. It asks for them in a fixed order, then builds the quote only from the owner's price sheet. - Confirm the headcount, the date, and whether the order is pickup or delivery with a time window. - Confirm the drink and food mix, including milk alternatives and any allergens the customer names. - Check the date against the capacity calendar the owner maintains, including deliveries already booked. - Build the quote from approved prices only. If an item is not on the sheet, flag it for the owner instead of estimating. - Hold the reply as a draft. After approval, send the quote with the shop's deposit and cancellation terms.
Illustrative scenario: A customer asks for coffee and pastries for forty people at an office on a Thursday morning. The agent confirms the headcount and delivery window, notes a request for oat milk in half the cups, and checks the calendar. The morning already has a delivery booked, so the agent flags the conflict in its draft. The owner moves the delivery slot, approves the quote, and the agent sends the terms. Event-style orders follow a similar pattern, as described in AI Agents for Event Planners: Ready for Superintelligence.
Supplier reorders: drafts that a person signs Supplier reorders are the most repetitive job in a coffee shop and the easiest place for quiet errors: a duplicate order after a timeout, a missed cutoff, or a substitution nobody noticed until the delivery arrived. The agent should prepare these orders with its reasoning visible, and a person should decide what leaves the building. - Keep par levels, supplier cutoff times, minimum order quantities and approved substitutions in one document the agent can read. - Let the agent read the count sheet or depletion data and propose an order with the reasoning attached. - When an order is submitted through an integration, attach a unique reference so a retry cannot place the same order twice. Payment platforms handle this with idempotency keys, and a supplier integration should follow the same discipline. - Require owner approval for any order above the agreed threshold, any new supplier, and any substitution the agent proposes. - Log every draft, approval and submitted order so that a later question about a delivery can be answered from records.
Purchasing work in other businesses follows the same sequence of draft, approval and follow-up, as AI Agents for Procurement: Quotes and Supplier Follow-Up explains. Shops that also sell beans wholesale have extra batch and order records to keep straight, which AI Agents for Coffee Roasters: Wholesale Orders and Batch Notes covers.
Loyalty questions: answer from the rules, route the exceptions Loyalty programs generate a steady stream of small questions: how many visits remain before a free drink, whether points were credited, whether a promotion applies to a particular order. Most of these have answers in the published rules and the customer's record. An agent can answer them accurately if it reads the current rules each time, rather than relying on a memory of last season's version.
The exceptions need a different path. A customer who says a balance is wrong, asks for a one-off discount, or wants a reward applied to an older order is asking for a decision. The agent should summarize the case, attach the relevant records, and hand it to a person. It should never promise a credit, invent an offer, or imply that staff have already agreed.
Customer messages are also untrusted input. A message telling the agent to ignore the shop's rules and apply a full discount is an attempt to change its behavior. The OWASP guidance on prompt injection prevention explains why instructions inside user content should not override the operator's rules. Authorization guidance from OWASP makes a related point: access must be enforced by the system, not granted because a request sounds convincing. In practice, the agent cannot apply a discount it has no permission to apply, whatever the message says.
Owner approval is how margins stay protected as the agent improves Margin protection is not a setting you switch on once. It is a set of decisions about what the agent may do without asking. A stronger agent will draft better quotes and handle more messages, but your margin still depends on the price sheet, minimum commitments and substitution rules. Those stay under owner control whatever model sits underneath. The NIST AI Risk Management Framework treats this kind of work as governance: assigning accountability for outputs, measuring how a system behaves, and managing the risks that grow with its role.
Approval tiers that work in practice: - Tier one: the agent acts alone on low-impact tasks, such as confirming opening hours, reporting a loyalty balance from the record, or logging an inquiry. - Tier two: the agent drafts and the owner approves before anything is sent, such as catering quotes and supplier orders. - Tier three: the agent does not resolve the case. It routes it with a summary, such as refunds, disputes, custom pricing and any change to the price sheet.
Before moving a task from tier two to tier one, test the agent against past cases where the right answer is already known: old catering quotes, previous reorders, real loyalty questions. Anthropic's guidance on evaluating agents describes this approach of checking behavior on realistic tasks before trusting it with more. Record every failure, fix the rule or tool that caused it, and rerun the same set. An improvement counts only if the old cases still pass.
Be Human handles this as a managed service. The first engagement defines the one role the shop needs, the memory it reads (price sheet, par levels, program rules), the tools it may use, the approval boundaries, and a success criterion the owner can observe, such as the share of quotes approved without edits. Ongoing operations cover monitoring, recurring evaluations against past cases, permission reviews whenever a point-of-sale, supplier or booking integration changes, and controlled improvements that pass through the same approval path.
Checklist for a first pilot and where to begin - Write the price sheet, par levels and supplier cutoffs in one document, with an owner name next to each rule. - Choose one task from the three categories and assign each action to an approval tier. - Collect past catering quotes, reorders and loyalty questions with known correct outcomes to serve as the test set. - Grant read access only to the records that task needs, and draft-only access to anything that leaves the shop. - Run the test set, log every failure, and fix the cause before any action moves up a tier. - Review permissions after any change to your point-of-sale system, supplier accounts or booking tools.
If you are not sure which of these tasks should go first in your shop, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions Can an AI agent send catering quotes without owner review? Not at first. Keep quotes in draft until the agent has passed tests on your past quotes and the price sheet is complete. The agent still saves the most time on the slow part, which is collecting headcount, dates, delivery windows and dietary details.
Should an agent place supplier orders directly? Start with drafts. Direct submission makes more sense later, for recurring low-value items with fixed par levels, a supplier integration that confirms receipt, and unique order references that prevent duplicates. Keep an approval threshold for unusual orders, new suppliers and substitutions.
What information does the agent need? It needs the price sheet, par levels, supplier cutoffs, the published loyalty rules and read access to the records each task requires. Give it no more than that. Payment card details and staff pay records are not needed for these tasks and should stay outside its reach.
How do I know the agent is getting better? Keep a fixed set of past cases with known correct outcomes and rerun it after each change. Watch the edits the owner makes to drafts, since they show exactly where the agent misses. If a change improves new cases but breaks old ones, it is not an improvement.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


