Who Pays for AI? A Cost-Ownership Framework for Cannabis Operators
The most expensive AI mistake a cannabis operator can make in 2026 isn't picking the wrong model. It's not knowing who's paying for it.
If you've talked to any software vendor in the last six months, you've heard about AI. Embedded copilots. Agents. Auto-summaries. Anomaly detection. Natural language reporting. The pitch is always some version of: we've added AI, and it's going to transform how you run your business.
Some of it will. Most of it won't. But the question almost nobody is asking and the one that's going to separate the operators who actually capture value from those who write checks for hype is much more boring:
Who pays for the inference?
Every AI interaction has a cost. Someone is paying it. If you don't know who, you don't understand the deal you're being offered. And as cannabis operators face continued margin compression, regulatory expansion, and capital scarcity, getting AI cost ownership wrong is going to quietly inflate software spend in ways that don't show up cleanly on any line item.
This is the framework we use at Flourish to think about it. It's the same framework we recommend our customers use when evaluating any AI capability ours included.
The State of AI in Cannabis Software in 2026
Some context first, because the framework only makes sense against the backdrop of how fast this market is moving.
According to Gartner's 2026 forecast, worldwide AI spending will hit $2.52 trillion this year, a 44% year-over-year increase. Gartner separately projects that 40% of enterprise applications will embed AI agents in 2026. Cannabis software is no exception. Every ERP, POS, cultivation platform, and compliance tool in the category is shipping AI features, and the pace of release notes has shifted from "we added a new report" to "we added a new agent."
The underlying use cases are real. According to BCC Research's AI Impact on Cannabis Market report, cannabis operators are deploying AI across cultivation (computer vision for yield and disease detection), retail (recommendation engines and demand forecasting), compliance (METRC reconciliation, COA parsing, audit prep), and finance (280E optimization, inventory forecasting). One published case study saw PurpleFarm report a roughly 20% yield increase after deploying an AI-powered crop monitoring system. Industry surveys cited by BCC also note that 86% of cannabis consumers say they'd be loyal to a dispensary offering personalized recommendations.
What is changing and what makes the cost-ownership question urgent is that AI in cannabis software is moving from a feature checkbox to a workflow dependency. When AI is the thing reconciling your METRC submissions, monitoring your COAs, or driving your reorder logic, the question of who pays for it stops being theoretical.
The Three Patterns of AI Deployment
Every AI-powered feature you'll encounter in cannabis software falls into one of three patterns. The patterns differ in who triggers the AI, who runs it, and most importantly, who pays for it.
Pattern 1: Bring-Your-Own-AI
The customer has their own AI subscription, Claude, ChatGPT, Gemini, whatever and connects it to their operational software. They ask their AI questions. Their AI fetches data from the software. The AI does the reasoning. Their subscription pays for the inference.
The software vendor's job here is to be a connector: expose data and actions through an open protocol (the emerging standard is the Model Context Protocol, or MCP), authenticate the user, and let the customer's AI of choice do the work.
Who pays: The customer's existing AI subscription. Vendor's role: Open connectivity layer. Best for: Ad-hoc analysis, ops questions, reporting, "I wonder if…" exploration, workflows the operator wants to drive themselves.
In a cannabis context, Pattern 1 looks like a compliance lead asking their Claude assistant, "Pull every package transferred from cultivation to processing in the last 30 days and flag any with COA results that haven't been received yet." The AI does the reasoning. The ERP exposes the data. Nobody at the software vendor is paying for the inference.
Pattern 2: Embedded In-App AI
The software vendor builds AI directly into their product. A cultivation manager opens the app and the system has already flagged a discrepancy between the lab report and the inventory record. A compliance officer reviews a METRC submission and gets an automatic readiness check. None of this requires the user to "talk to an AI" the intelligence is just baked in.
Behind the scenes, the vendor is making API calls to an AI provider, and paying for every one of them. That cost gets capitalized into the software price, the same way database hosting and uptime monitoring do.
Who pays: The vendor, who prices it into the subscription. You pay indirectly through your software bill. Vendor's role: Operator of the AI capability, on the hook for cost, quality, and reliability. Best for: Workflows the vendor can run consistently across the entire customer base anomaly detection, document understanding, compliance pre-checks, automated alerts.
Pattern 2: Where most cannabis software has spent the last two years. It's also where most of the "AI feature creep" risk lives see below.
Pattern 3: Managed AI Services
This is a hybrid. The vendor runs an AI agent on the customer's behalf a "concierge," an automated monitor, a scheduled task that watches for problems and acts on them and prices it as a separate service.
The vendor is paying for the inference, but unlike Pattern 2, it's not bundled into the base subscription. It's a discrete line item the customer chooses to buy.
Who pays: The customer, through a dedicated managed-service SKU. Vendor's role: Provider of a productized, ongoing AI workflow. Best for: High-touch, high-value workflows where the vendor's expertise plus AI plus continuous operation creates outcomes a customer can't easily run themselves.
For cannabis operators, Pattern 3 often shows up as continuous compliance monitoring (an AI agent watching for METRC discrepancies in near-real-time and opening tickets when one fires) or as a forecasting service (a weekly demand model that produces buy recommendations the retail team executes).
Why This Matters Now
For most of the last two years, cannabis operators have been offered AI almost exclusively under Pattern 2 embedded features baked into platform subscriptions, with no clear pricing signal about what the AI itself costs. That worked when AI was a feature checkbox. It's becoming untenable now that AI is moving from "useful" to "structural."
A real AI-powered workflow one that runs across hundreds of packages, dozens of orders, daily compliance submissions, weekly forecast generation has real cost. If that cost is hidden inside a fixed subscription, one of two things happens. Either the vendor restricts how much AI you can actually use (rate limits, narrow scope, capped features), or the vendor swallows the cost and looks for ways to recoup it elsewhere. Either way, the customer ends up with less AI than they expected.
The macro environment makes this sharper still. The Diligence Stack's analysis of enterprise AI budget architecture found that roughly 28 cents of every incremental AI dollar at enterprises comes from net-new IT budget. The other 72 cents is pulled from existing software budgets, BPO contracts, headcount savings, license consolidation, and business-unit budgets. Translation: most enterprises are funding AI by reallocating spend, not expanding it. In cannabis, where capital is scarcer than in most industries, that reallocation pressure is even more acute. Hidden AI cost inside software bills competes directly with line items operators can see and control.
The operators who are going to win with AI in 2026 and beyond are the ones who think clearly about which workflows belong in which pattern.
How to Apply This as an Operator
Three questions to ask any vendor pitching you AI capabilities:
1. "Is this AI you're running, AI I'm running, or AI we're sharing?"
If the answer is fuzzy, push harder. The pattern dictates the economics, the limits, and the strategic implications. A vendor who can't answer cleanly hasn't thought it through.
2. "What happens when I want to use this more than you expected?"
For Pattern 2, this is the critical question. Embedded AI sounds great until you hit a usage cap. For Pattern 1, it's a non-issue your subscription, your limits. For Pattern 3, the SKU should price scale into it explicitly.
3. "Can I bring my own AI to your platform?"
This is the question that separates platforms with a real AI strategy from platforms with an AI veneer. Open connectivity MCP, API access, structured data egress means your existing AI investments (your Claude Team plan, your ChatGPT Enterprise seats, your internal data science team) can extend onto the platform. Closed AI means whatever the vendor decided to build is all you'll ever get.
A useful supplemental check: ask how the vendor handles data residency, audit logging, and access scoping when AI yours or theirs touches your operational data. In a regulated industry like cannabis, where state tracking systems and patient data both carry compliance weight, the AI strategy question and the data governance question are the same question.
The Flourish View
We think the right answer for a vertical ERP isn't to pick one pattern. It's to be deliberate about all three.
We're investing heavily in Pattern 1 — opening Flourish to any AI assistant our customers want to use, with a secure MCP server that lets your team query, analyze, and act on Flourish data through Claude, or whatever AI tooling you're already paying for. Your existing AI subscriptions become more valuable the moment Flourish data is in scope.
We're building Pattern 2 selectively — for workflows where the vendor running the AI is structurally the right answer. COA discrepancy detection. Audit anomaly review. Compliance readiness pre-checks. These are places where Flourish, running AI consistently across the whole customer base, produces outcomes no individual operator could replicate.
We'll offer Pattern 3 where it earns its keep — productized agent services for operators who want continuous monitoring or expert-augmented workflows without staffing them internally.
This isn't a roadmap announcement. It's a posture. We think open connectivity is the foundation. We think embedded AI should be honest about its scope. And we think operators should always know who's paying for what.
What's Next Then
If you're evaluating cannabis software in 2026, add AI cost ownership to your diligence checklist. Ask the three questions above. Get clear answers in writing. And when you compare platforms, compare them on how well their AI strategy compounds your existing investments not on how impressive the demo looks.
The vendors that earn their seat in the cannabis stack over the next few years won't be the ones with the flashiest AI features. They'll be the ones who let you build a real AI practice on top of their platform at a cost structure you actually understand.
Flourish Software is the vertical ERP for cannabis operators, serving cultivation, manufacturing, wholesale, and retail across the supply chain. Our MCP server is launching this quarter and brings open AI connectivity to Flourish data for any customer with a Claude subscription. Reach out if you'd like to be among the first operators connected.
