The case for cannabis AI guardrails – and how to create an AI tool that works

This is part of a regular series of MJBizDaily interviews with major THC industry players. To be considered for an interview, contact editorial@mjbizdaily.com.

Cannabis-specific artificial intelligence tools are here and multiplying fast.

Software vendors are pitching licensed operators on products that promise to optimize cultivation rooms, flag compliance errors in ad creative, audit inventory, replace budtenders with AI-powered kiosks and deliver instant business insights through conversational analytics.

But according to Nick Afonsky, founder and CEO of Washington, DC-based IT consulting firm Palomar Group, the gap between what the tools promise and what they reliably deliver is wide.

Using AI to reach a useful answer is one thing. Building a governed, repeatable enterprise system that sits on top of company data, returns consistent results and keeps sensitive information out of public application programming interfaces (APIs) is a different challenge.

And that’s a challenge that most AI software vendors have yet to meet, Afonsky said.

“The promise is really big,” Afonsky told MJBizDaily in a recent interview. “Harnessing and making it work is the challenge.”

Afonsky will discuss this topic at MJBizCon Dec. 1-4 at the Las Vegas Convention Center in a panel discussion titled “AI in the Grow: Separating Real ROI from Vendor Hype.”

Why is reliable data the key to success?

The trouble starts when operators try to get the same accurate answer every time. Without training and resources, it’s hard to reach consistency.

Operators also must account for a basic trait of the technology. AI is probabilistic. Ask the same question twice, and the answer may differ.

“They’ve [generative AI juggernauts] done a great job of democratizing the magic of AI,” Afonsky said. “We’ve all become prompt engineers.”

“But AI is a little like a kid in the sense that you need to tell kids what to do – always do this, and never do this.”

“You have to give it guardrails.”

For Afonsky, AI projects are like any other IT initiative: success depends on clean, well-structured data, executive sponsorship, patience and trust.

To get the most out of an AI-powered tool, his teams spend much of their time using AI to clean, normalize and enrich retail point-of-sale data. That’s because the raw output is often poor, with too many product categories, missing strain names, absent products and duplicate entries.

Once the data is reliable, more use cases for AI open up, he said.

Palomar builds automated quality-control checks that validate inbound data before it reaches clients.

From there, operators can forecast demand, evaluate marketing content and plan promotions.

Where does AI pay off for cannabis operators?

In cultivation, AI paired with internet-connected sensors can scan for pests, adjust lighting and temperature and catch anomalies the human eye misses at scale.

For larger operations, the return is likely worth the investment, Afonsky said.

On the compliance side, Afonsky said Palomar built a tool that reviews ad creative assets against state rules.

An operator feeds the tool an image of an advertisement and a target state, and it checks whether the license number appears where it should and flags images prohibited by state law, such as cartoons or other state-specific bans.

The tool then produces a report with recommended fixes.

AI can also help with inventory audits, he said. Duplicate or inaccurate items make audits unreliable, and AI can quickly identify and correct those errors, letting operators run audits faster.

“If you don’t have accurate inventory or you have duplicate items because someone entered the same product in two different ways, there’s a risk because you can’t audit your inventory accurately,” he said.

“AI tools can spot and clean that up quickly.”

Why should operators beware of prototypes?

Operators should watch for gaps between a prototype and a production system. A prototype proves an idea can work for one person, while a production system scales, protects data and delivers validated results for multiple users – repeatedly.

“It’s one thing to say we have an AI budtender, but is it really functioning?” Afonsky said. “Outside a pilot, is it generating results your shopper is going to appreciate?”

Afonsky advises buyers to demand validated results.

Operators should:

• Confirm the tool answers the question they have
• Works on their own data rather than demo data
• Ties back to a business use-case scenario, whether that means saving time, growing revenue, cutting cost or managing risk.

Even well-designed AI can fail when something obscures a grow-room camera or when the developers never learned how a budtender phrases a question.

“It’s easy to do a demo of these things,” Afonsky said. “But putting it into your environment with your data in a dispensary where there’s an end user who’s maybe not thinking like a data scientist is different.”

What should operators discuss with vendors selling AI-powered solutions?

Afonsky urges operators to have an upfront conversation with vendors touting AI solutions and guard against “scope creep.”

“Everyone has big ambitions for AI, but it’s important to figure out what we can do in the short term that will generate the greatest benefit, address a use case and not be too complex,” he said.

He recommends rapid prototyping and iterative development: keep projects simple and achievable, prove the benefit, then expand. Operators who take on too much can head down a path that’s hard to reverse.

Rapid prototyping is changing how consultants and software firms sell, Afonsky said. The demo era is over because consultants can now prototype an idea and show a client the result within hours or days.

“It’s a show-me type environment,” he said.

This topic and others shaping the cannabis economy will be front and center at MJBizCon. Keep up with the latest conference news here.

Margaret Jackson can be reached at margaret.jackson@mjbizdaily.com.