If you're evaluating supply chain software with AI features, whether it's cannabis-specific or general distribution software, the same checklist applies. Here's a practical guide to cutting through the pitch and getting to what actually matters.
Start With What the Feature Actually Touches
Before anything else, get specific about scope.
What Data Can It Read?
Ask exactly which records, tables, or systems an AI feature has access to. "Your data" is not an answer. "Your inventory table and order history" is.
What Actions Can It Actually Take?
Separate reading from acting. A feature that can see your inventory to answer a question is a very different risk profile than one that can update records or trigger a shipment on its own.
Who Reviews the Output Before It's Final?
This is the single most important question on this entire list. Anything touching a customer order, a vendor payment, or a compliance record should have a defined person reviewing it before it's final. If the answer is "nobody, it's fully automated," that's not a feature, that's a liability waiting for its first bad day.
Where MCP Fits Into This Evaluation
MCP, the Model Context Protocol, is an open standard that defines AI tool access through specific building blocks: tools it can use, resources it can read, and prompts it can run. You don't need to know if a vendor uses MCP by name. What matters is whether they can answer the questions above with that same level of specificity, because that's exactly the discipline MCP is designed to enforce.
A Quick Litmus Test
Ask a vendor to describe their AI feature using the tools-resources-prompts framework, even if they've never heard those specific words. A vendor who can map their feature onto that structure easily has almost certainly thought about access control carefully. One who can't is likely still figuring it out, possibly in production, with your data.
Questions to Ask About Failure Modes
A demo only ever shows the happy path. The real test is what happens when things don't go smoothly.
What Happens on Messy Input?
Ask to see the feature handle a genuinely messy example from your own operation, not the vendor's clean demo data. How it handles the mess tells you more than how it handles the ideal case.
What Happens When It's Uncertain?
A good feature flags uncertainty and asks for review. A concerning one guesses confidently and moves on. Ask specifically what the system does when it doesn't have a clear answer.
Applying This to a Cannabis-Specific Purchase
Cannabis distributors and brands should run this exact checklist, with extra weight on the review question, since the cost of an unreviewed compliance mistake is higher than in most other CPG categories.
A Working Example That Passes the Checklist
Distru's AI Order Agent answers each of these questions specifically: it reads your inventory and incoming order data, it acts by drafting a clean sales order, and a rep reviews everything before it ships. When an order won't fill, it flags that instead of guessing. Customers using it save 40+ hours a week, a result that held up because the scope was defined this clearly from the start.
A One-Page Version of This Checklist
What data does it read. What can it actually do. Who reviews it before it's final. What happens on messy input. What happens when it's uncertain. Five questions, and if a vendor answers all five specifically, you're evaluating a real feature, not a pitch.
How to Use This Guide in an Actual Vendor Meeting
A checklist only helps if you actually bring it into the room.
Ask the Questions in Order
Start with scope (what it reads, what it does), then move to review (who signs off), then move to failure modes (messy input, uncertainty). Vendors who've thought this through will answer smoothly in that order. Vendors who haven't will start improvising somewhere in the middle.
Write Down the Answers, Not Just the Vibe
A confident-sounding answer and a specific, correct answer feel similar in the moment but aren't the same thing. Write down exactly what they said for each question, and review it afterward without the charisma of the sales call influencing your read.
Want to run your own checklist against a live feature? Schedule a demo with Distru.
