AI Agents for Dispensaries: Menu Questions at Scale

An AI agent for a dispensary has one obvious job: answer the question a budtender would otherwise stop to answer, is this still in stock, what does it cost, does this particular store carry it. Whether that agent earns its place depends on how often it gets that answer right, and a 2025 benchmark on grounded product question answering found stock and availability questions are the hardest kind for an AI system to answer correctly, with the right context in hand only 51.4% of the time. This post works through what that accuracy gap means for a dispensary deciding whether to put an AI agent on live menu questions, what a wrong answer actually costs at the register, and the test to run before a real customer runs it for you, using the speed-to-lead system as the frame for where an agent like this earns its place.

Can a dispensary use an AI budtender for menu questions?

Yes, for the questions that have exactly one right answer: is it in stock, what does it cost, which store has it. The same research that makes that possible also shows where it breaks: stock and availability questions are the category where a grounded AI system's retrieved context actually contains the right answer least often of any question type measured, at 51.4%. An AI budtender is worth deploying wherever your own measured accuracy on that question clears what a wrong answer costs you, and not before.

That is a narrower claim than most vendor pages selling an AI budtender make. Cova, Blaze's Herbie, Terpli's Zippy, and CannaVeris each pitch a version of this tool, and each is confident in its own demo. None of them publishes a number for how often it is actually right on a live stock question across a real menu, under real traffic, at a real store. The rest of this post gives you a number to hold any of them to, and a short test you can run yourself before you trust one.

What makes a dispensary menu question different from an ordinary retail question?

Three things a generic retail bot was never built to carry: a legal age gate before any product talk, a hard line against medical or dosing claims, and a register that is the only legal source of truth for what is actually on the shelf. None of those is this post's subject. They are the subject of AI customer service for cannabis dispensaries, which walks through the compliant setup in full. This post assumes that setup is already right and asks a narrower question: once the gates are built, how good is the agent at the specific question it answers most, a live menu or stock question, and what should you expect it to get wrong.

Where the compliance line sits

A medical claim, a dosing recommendation, or a disputed order still goes to a person, every time, regardless of how accurate the stock-question numbers below turn out to be. Accuracy on "is this in stock" says nothing about accuracy on "what should I take for sleep," and a system that is reliable on the first question is not thereby safe to trust on the second. Keep the two questions separated, because a dispensary that conflates them is the failure mode the compliance post exists to prevent.

How accurate is an AI agent when it answers a live stock question?

Less accurate than on almost any other question type a retail agent handles, by the one benchmark that actually measured it by category. The 2025 paper behind this post tested a retrieval-grounded product question answering system across twelve query intents and found stock and availability questions had the lowest context coverage of the set, 51.4%, meaning the system's retrieved information contained the correct, current answer barely half the time. Checkout questions, by contrast, had context coverage of 85.9% and the lowest hallucination rate measured, 1.32%. Stock questions sat at a 2.50% hallucination rate on top of that weak coverage, and product-exchange questions hallucinated most often of all, at 6.89%.

Which questions the data says to trust

Checkout mechanics, store hours, loyalty balance, and anything the register answers with one unambiguous field are the strongest category in the data, both in how often the right context is retrieved and in how rarely the model invents an answer instead. Those are also the questions that cost a dispensary the least to get wrong, since a customer who asks your hours and gets a stale answer notices immediately and asks again.

Which questions the data says to route to a person

Live stock and availability sits at the other end, the weakest context coverage of the twelve intents measured, and that is before cannabis-specific wrinkles like a strain that sold out an hour ago or a size variant that only one of your stores still carries. Product-exchange and substitution questions, "do you have something like this instead," hallucinated more than any other category in the same study. Treat both as the categories where an agent needs either a live, second-by-second feed from the register or a standing rule to say "let me check" and hand off, not a confident guess.

Does that accuracy hold up once a group runs more than one store?

No, not without extra work, because every store adds its own live shelf and its own answer to the same question. A single-location dispensary has one inventory feed to ground the agent against. A group running three stores has three, each changing independently through the day, and a stock question is only answerable once the agent knows which store the customer means. Get that step wrong and the agent's context coverage on stock questions does not stay at 51.4%, it gets worse, because the agent is now retrieving against the wrong store's shelf more often than a single-location build ever would. The backend side of keeping every store's menu current is its own problem, covered in dispensary menu sync automation; this post is about what the agent does with that feed once it has it.

The failure compounds with each store you add, not just appears once at the second. A two-store group has one wrong-store risk to manage, between store A and store B. A four-store group has six possible pairs an agent could confuse a customer's intended location with, and the odds of the agent defaulting to the wrong one rise with every location added unless the location-resolution step is tested at each new store, not assumed to still work because it worked at the first two.

The two facts an agent needs before it answers

A location and a timestamp. The agent has to resolve which store a customer means before it touches the stock question at all, the same way a person behind a counter already knows which counter they are standing at, and it has to treat any inventory read older than a few minutes as untrustworthy rather than final. A build that skips the first fact answers confidently from the wrong store's shelf. A build that skips the second answers confidently from a shelf that changed since the last sync.

What does a wrong answer cost at the register?

More than the one sale it misses, and more online than in person. Register-transaction data from Headset puts the average US dispensary basket at $47.03 over the twelve months ending September 2026, across 482.0 million transactions and $24.6 billion in tracked sales. An agent that tells a customer a product is out of stock when it is not, or in stock when it already sold, is not losing an abstract "a sale." It is losing roughly that basket, repeated across however many customers ask the same question that day.

The online order a wrong answer actually loses

Flowhub's 2025 transaction data found online orders averaging $68.01, 35% higher than the $50.56 average for a walk-in sale, while the average walk-in basket itself has been shrinking, 2.7 items per visit, down from 2.9. An online shopper who asks a chat agent whether a product is available and gets a wrong answer does not stand at a counter waiting to be corrected. They close the tab, and the order that leaves with them is the larger one of the two. That asymmetry is the real argument for measuring accuracy on this specific question type before trusting an agent with it at scale, not a smaller argument about convenience.

What does it cost to keep a person answering these questions instead?

A lower number than most owners assume, which is exactly why the accuracy bar for replacing that person has to be high. The US Bureau of Labor Statistics counted 3,897,860 retail salespersons employed nationally in May 2025, the closest labor category to a budtender fielding counter questions, at a median wage of $17.03 an hour. As an illustrative example, not a client result: a store where one staff member spends even an hour of a shift answering "is this in stock" questions is spending roughly that wage on work a well-grounded agent could take at a fraction of the per-question cost, once volume is steady. The gap is not in the hourly math. It is in whether the agent gets the answer right often enough that the time it frees up is not immediately spent fixing what it got wrong.

How do you test an AI budtender's accuracy before a live customer does?

Run it against your own shelf first, with answers you already know, before a stranger runs it against your shelf with an answer only your register knows. Pull five real stock questions from your own counter log or your own menu, the kind a customer actually asks, and check the agent's answer against the register at the same moment, not against the menu site, which is exactly the lagging source the benchmark's weak coverage number is warning you about.

The five-question test

Ask it whether your best-selling product is in stock right now, whether a product you know sold out this morning is still listed as available, what a specific size or weight of one product costs at two different stores if you run more than one, whether it can find a substitute for a product you know is out, and whether it correctly declines a dosing question instead of answering it. Score each answer against the register, not against what sounds plausible. A build that passes the first and third questions but fails the second is telling you its sync lag, not its reasoning, is the weak point, and a build that fails the fourth is telling you it needs the escalation rule, not a bigger model.

Run the same five questions once a week for the first month, not just once at launch. A build that scores well on day one against a quiet menu can slip once real volume hits it, a sold-out strain does not update for twenty minutes, or a new store's feed turns out to be mapped to the wrong location entirely. Weekly scoring for a month gives you a trend line instead of a single snapshot, and a trend line is what tells you whether the accuracy gap is closing as the integration settles in or staying put because something upstream is still wrong.

Run that test before you sign with any vendor pitching a dispensary chat tool, and run it again after your second store goes live, since the multi-location failure mode above does not show up on a single-store test. For the backend work of keeping every store's feed current enough for that test to mean anything, see dispensary menu sync automation, and for the compliance gates that have to sit in front of any of this, see AI customer service for cannabis dispensaries. The same test-before-you-trust discipline applies to a phone-based agent, covered in AI receptionist alternatives, and to the cart an online shopper leaves behind when a question like this goes unanswered, covered in compliant abandoned-cart recovery for dispensaries. When you want this sized against your own stores, get your free Systems Plan.

Methodology

This post draws on four sources. Tangarajan, Rajasekar, Rathi, Dandin, and Ersoy's "Contextually Aware E-Commerce Product Question Answering using RAG" (arXiv:2508.01990, August 2025) supplies the context-coverage and hallucination-rate figures by query intent, including the 51.4% stock-availability coverage figure this post is built around and the 85.9% checkout coverage figure used as the strong-category contrast. Headset's Cannabis Industry Statistics supplies the sales, transaction, and basket figures for the twelve months ending September 2026, computed from register transactions rather than survey data. Flowhub's 2026 retail-trends report supplies the online-versus-walk-in order value comparison from its 2025 transaction data. The US Bureau of Labor Statistics' Occupational Employment and Wage Statistics for retail salespersons (May 2025) supplies the labor-cost baseline. The one-hour staff-time example is a labeled illustrative example, not a client result. Taken together, those four figures are the accuracy and cost bar an AI agent for a dispensary has to clear before it takes a live stock question off a budtender's plate.