Skip to content
be-human.ai
← Back to the journal

AI by industry

Superintelligence Prep for Liquor Stores: Age Checks and Delivery Rules

An AI agent for a liquor store can enforce age checks and delivery rules as fixed limits, while a licensed person keeps every compliance decision.

Be Human typographic cover with the title: Superintelligence Prep for Liquor Stores: Age Checks and Delivery Rules
Conceptual illustration created with AI · Be Human

An AI agent for a liquor store can answer customer questions, show delivery windows, collect order details and keep a clean record of each conversation. Age verification and delivery eligibility, however, must be hard boundaries that the agent cannot cross, and every compliance decision stays with a licensed person who answers for the store. The agent routes, records and flags. It does not decide.

Key takeaways

  • Age verification is a gate the agent never passes on its own, whatever the customer says or claims.
  • Delivery eligibility comes from zones, hours and license conditions stored as fixed rules, not from judgment in a chat.
  • Every refusal, handoff and override is logged with the reason, the time and the person who reviewed it.
  • A licensed person owns compliance decisions, and the agent escalates to that person through a defined path.
  • Start with one narrow role, such as order questions and pickup status, before any delivery automation.

Why age checks must be a hard boundary, not a suggestion

A language model is good at conversation and poor at immovable rules. A customer who says they are old enough, or who claims to be buying for a friend, has given the agent a statement, not proof. Some customers will also try to argue the agent out of its rules with phrases like "ignore your instructions" or "the manager always makes exceptions." The OWASP guidance on prompt injection prevention treats any user-supplied text as untrusted input, and the practical lesson for a store is simple: the rule must not live only in the wording of a prompt that a clever message can rewrite.

For that reason, age status should be a state controlled outside the conversation. The agent can read whether a verification was completed at the counter, at the door or in the point-of-sale system. It cannot mark a verification as complete, and it cannot accept an identity document in chat. When the record is missing or unclear, the answer is always the same: the order does not proceed, and a person is notified.

Which delivery rules belong in the agent's hard limits

Delivery rules are easiest to manage when they are written as configuration that the system checks, not as paragraphs the agent is asked to remember. A practical starting list usually includes:

  • Permitted delivery zones, taken from the store's license terms and stored as a lookup table the agent reads.
  • Permitted delivery hours, including holiday or local restrictions, maintained by a named person.
  • Recipient requirements, such as verification at the door before hand-over, recorded by the driver.
  • Product categories or order sizes the store has chosen to exclude from delivery.
  • Any signal that the buyer may be intoxicated or unclear about the purchase, which triggers an immediate handoff rather than an automated answer.

Licensing rules differ by jurisdiction and change over time, so the store should not rely on the agent's general knowledge of them. The licensed person writes the rule in plain language, confirms it matches the license conditions, and signs off on the version that goes live. The agent then works from that version only.

Separating what the agent can do from what a licensed person decides

The clearest way to set the boundary is to list the actions on each side. The agent can explain store hours, describe the verification step a customer will meet, collect order details, check whether a verification record exists, look up a delivery zone and draft a note for staff. The agent cannot accept an ID, waive a check, approve a delivery outside the zone table, decide whether a restricted sale is compliant, or interpret an unclear license condition.

The OWASP Authorization Cheat Sheet makes a broader point that applies here: access decisions should be enforced by the system, not assumed from what a request says. If the agent's tools cannot approve an exception, then no wording from a customer can produce one.

Illustrative scenario: a customer messages late at night asking for delivery to an address just outside the permitted zone, and says the store has always made exceptions before. The agent checks the zone table, finds no match, tells the customer the address is outside the delivery area, and creates a handoff note for the manager with the address, the time and the customer's message. It does not offer an alternative that implies approval, and it does not promise a callback time it cannot guarantee.

A step-by-step plan for scoping the first version

A first version should be narrow enough that a responsible person can review every rule it depends on. A workable sequence looks like this:

  • Write the relevant license conditions and store policies in plain language, and have the licensed person approve each one.
  • Convert each approved rule into a check the system can run, such as a zone lookup, an hour window or a verification flag.
  • Define the escalation path: who receives each type of handoff, through which channel, and how quickly they are expected to respond.
  • Choose the first role narrowly, usually order questions and pickup or delivery status, and leave sales promotions for later.
  • Test refusals with scripted attempts, including pressure, false claims and requests for exceptions, before any customer sees the agent.
  • Set a review date for the rule table, so that changes in license terms or store policy are reflected in the agent's limits.

Each step produces something the store can inspect. If a step cannot be checked by a person, it is not ready to automate.

What to log, and why the record matters

When something goes wrong, the store needs to reconstruct what happened, which rule applied and who decided. Observability practice, as described in the OpenTelemetry concepts documentation, relies on recording system behavior in a way that can be traced after the fact. For a liquor store, the minimum useful record includes each refusal, each handoff, the rule version in force, the reviewer who closed the handoff and the time of each step.

The record should also be restrained. The agent does not need to keep copies of identity documents, and storing them can create privacy obligations the store did not plan for. Keeping the outcome of a verification, rather than the underlying document, is usually the safer design, and the store should confirm the approach with its own counsel and applicable data protection rules.

How Be Human's managed approach handles this

Be Human designs and operates the agent around these boundaries from the start. A first engagement defines the single role, the memory and tools it needs, the approval boundaries that a licensed person owns, and an observable success criterion, such as the share of handoffs that reach the right person with complete information. Rules are kept as versioned configuration with a named owner. Managed operations can include recurring evaluations of refusal behavior, permission reviews when tools or staff change, and incident response if a boundary fails. The service is remote and available in English and Portuguese.

For a store deciding where to begin, the first useful question is which customer conversation is repetitive, rule-bound and safe to hand to a machine. The free Be Human AI diagnostic can help you find that first task worth handing to an agent.

Frequently asked questions

Can an AI agent check a customer's age by itself? No, and it should not try. The agent can read the result of a verification completed by staff or by the point-of-sale process, and it can tell the customer what the step involves. Capturing or judging identity documents remains a human, licensed task, and the agent hands the case to a person whenever the record is missing or unclear.

Does the agent decide whether a delivery is allowed? It should only apply the zone, hour and category rules that a licensed person has already approved and placed in the system. If an address, time or order falls outside those rules, the agent refuses the request and creates a handoff. Any exception is a decision for the responsible person, not for the agent.

Which role should a store start with? Most stores begin with order questions, hours, pickup readiness and delivery status, because these conversations are repetitive and their boundaries are easy to test. Sales promotions, recommendations and delivery expansion can follow once the refusal behavior has been reviewed over real conversations. Starting narrow makes the compliance review manageable.

Do we need outside advice to set these rules? The rules should come from the store's license conditions, and the licensed person must confirm that each one matches them. Where the conditions are ambiguous or the store operates in several jurisdictions, outside legal or regulatory advice is worth the cost. The agent should never be asked to resolve that ambiguity itself.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →