A connector is a checkbox. So is a model. Here’s what the moat in cannabis retail ops actually is.

Summary. Treez's Winston makes a fair point: reaching a POS API isn't a moat. We agree, and we'd add that a frontier model isn't one either. The durable parts are boring — a ledger that agrees with Metrc, rule packs that stop a draft before it exists, an approval queue with no “approve all”, and data rules you can read. This is where we agree, where we differ, and what to ask any vendor.

In June, Treez launched Team Winston and published an essay arguing that “a connector is a checkbox, not a moat.” The argument is that any AI can reach your data through an API; the value is in knowing what the data means. It's aimed at products like ours. It's also correct.

Where we part ways is what comes next. If connectors aren't the moat and understanding is, the question becomes: understanding of what, and proven how? In a regulated retail business the answer can't be “the model learned it.” It has to be something an inspector, a CPA, or a GM can open and check.

1. A ledger that agrees with Metrc

Every POS report and every Metrc report describe the same packages and disagree about them. The first durable thing is a single, append-only, package-level ledger — Metrc plus the money columns — where every sale, transfer, invoice line and credit lands against a tag. Once that exists, “the invoice says 30 and the manifest says 28” is a held line with a credit memo drafted, not a Slack thread. No model does this. It's plumbing, and it takes a year.

2. Rule packs that run before the draft

A vendor-funded promo is allowed in Michigan on uniform terms, blocked in New York as an inducement, and unresolved in California. An assistant that drafts the vendor credit report and then asks you to check is faster than a spreadsheet and still dangerous. The check has to run first, with a rule ID and a citation, and “UNKNOWN” has to route to a person rather than a best guess. Rule packs are versioned; each proposal records the version it was checked against.

// pack NY v9, excerpt
RI-NY-06   vendor-funded promos, rebates, credits    → BLOCKED   · OCM inducement rule
RI-CA-07   vendor-funded discount reimbursement      → UNKNOWN → compliance lead

3. An approval queue with no “approve all”

Winston and BudAlly both put a human in front of every action, and that's now table stakes. The details differ. We think a proposal should carry its evidence, a confidence score, and a precondition hash — so that if your POS changed a price between the draft and your approval, the item is superseded instead of executed against stale facts. And the queue should not have an “approve all” button, because the day it exists is the day it gets used at 6 a.m.

4. Learned rules you can read

“Tell it once that a vendor bills in cases and it remembers” is a good feature. Where does the memory live? If it lives inside a model's context, nobody can audit it, version it, or turn it off. We store learned rules as objects — kind, scope, expression, version, who approved it — in a governance tab. Editing one creates a new version; old proposals keep the old one.

5. Data rules written down

Where did the “market median” come from? If the answer is scraped menus, your license is now next to a terms-of-service breach. If it's pooled customer data, whose, and did they agree? We use licensed data, your own price checks, and a network median with at least eight contributors — and every figure carries a source chip. That's slower to build than a scraper. It's also the only version we'd put our name on.

Where Winston is right, and ahead

Domain post-training on real retail patterns matters, and Treez has more of it than we do. First-party write access into Treez is real. Named customers are real. If you're a Treez store and you want insights and chat, it's a strong choice. Our bet is on the operators who aren't on Treez, who want the layer independent of any POS, and who want the menu, loyalty, manifests and close in the same system the drafts land in.

Five questions for any AI ops vendor

1. Can I export the audit trail — proposal, evidence, approver, credential — as a file?
2. Which rule, by ID and citation, blocked or allowed this draft?
3. Where does a learned rule live, and can I version it?
4. Where does market data come from, and is any of it scraped?
5. Is there an “approve all” button anywhere?

---

Sources: winston.team/what-winston-knows (Jun 2026); Treez press release, Jun 10, 2026; NY OCM inducement guidance [cite]; Dutchie Developer Terms [cite].