Hello INNOVATEwest Community,
Adding “AI” to your pitch deck or board minutes might sound good… but how are you actually going to implement it? Or, more to the point, how will AI help you increase productivity, revenues, or profitability? How will it empower your people? These are questions that Rogayeh Tabrizi, CEO and Founder of AI product company Theory+Practice, is begging business leaders to ask before picking an AI tool. Speaking with INNOVATEwest, she shared a practical framework any leader can use to figure that out for themselves.
Key takeaways:
- AI has become a catch-all phrase for multiple different technologies such as large language models, machine learning algorithms, or transformer models.
- AI tools can help you uncover patterns not perceptible by human analysis, but it requires a clear outcome and well-structured test to deliver valuable results.
- The best way to succeed with AI is to start with the outcome you hope to achieve, then think about what tests might help you get there; only then think about how AI can help run those tests.
Dave Tyldesley
Co-Founder/Producer SAAS NORTH & INNOVATEwest
Editor INNOVATEwest PULSE
Don’t let complexity hold you back, ever again. Connect everyone to everything with Boomi.
It’s frustrating for Rogayeh Tabrizi, CEO and Founder of AI product company Theory+Practice, to hear people talk about AI as if it were a singular, monolithic thing.
From her Vancouver HQ, Rogayeh helps clients all over the world realize that AI is a “class of applications” rather than a single technology, and the kind of AI you need depends entirely on the business goals you’re trying to achieve.
Speaking with INNOVATEwest ahead of her talk at the inaugural summit, Rogayeh explained how to build a business-first AI strategy.
Headquartered in Vancouver, BC, VoPay helps businesses navigate the complexities of payments at any scale through their API-first platform. VoPay’s robust infrastructure delivers frictionless payment technology and puts powerful financial tools into the hands of any business. Schedule a Consultation at vopay.com
Understanding the umbrella of AI
The power of AI is that the right tool can help you identify patterns that humans cannot spot on their own—but a single project might require multiple different AI-powered tools to get the job done.
To illustrate, Rogayeh shared the example of a consumer packaged goods (CPG) company that wants to forecast next year’s production with more accuracy.
In most cases, demand forecasting means analyzing the previous years’ sales and forecasts. A rough estimate is then made, with a contingency plan in place if demand or ability to supply changes significantly.
But with a well-defined problem, AI can help significantly. For instance, a machine learning algorithm can simultaneously analyze multiple years’ worth of data while correlating other variables like weather patterns and interest rate predictions that could impact supply chains. At the same time, a large language model can assess consumer sentiment in the news about your product offerings compared to previous years, offering a new variable input that might impact sales.
“You can see how much richness we can actually bring to a table, both in terms of ability to identify hidden patterns that are sitting in joint data sets, as well as benefiting from historical trends,” said Rogayeh.
advantage is a modern consulting, outsourcing and managed services firm hyper focused on great experiences and outcomes for our customers. We help organizations modernize, innovate, and use technology to develop competitive advantage and improve performance. Visit us at avantage.com/innovatewest to start your journey towards innovation.
How to build an AI strategy that gets results
AI can do a lot and add a lot of value, but only when you have the right strategy in place. Here’s how to create it.
Step 1: Describe your outcome with contextual detail
All AI success stories have one thing in common: the people implementing AI had a clearly-defined problem they wanted to solve or an end-state they wanted to attain. In the CPG example, it was about producing the right number of products in the most cost efficient way.
Step 2: Create hypotheses so you can run tests
If you have existing insights from customers, you might already have a strong hypothesis in mind—for example, a bank might have a goal of more multi-product customers, since those correlate to higher revenues and profitability. In that case, a hypothesis might be that existing multi-product customers typically start with a chequing account and then add a credit card or a line of credit.
If you’re in a smaller startup or launching a new offering, you’ll need to start from a new hypothesis. For instance, if your ultimate goal is more revenue through driving B2B revenue rather than your traditional B2C path, you might come up with a hypothesis that B2B buyers prefer a more concierge approach to selling (via calls) rather than your buy-now landing page that had succeeded in the B2C space.
Step 3: Structure tests in a realistic context
Structure tests in the most realistic environment possible, or you risk tainting the study. A famous example of this going wrong was Coke’s reaction to the Pepsi Challenge in the 1980s.
Pepsi ran blind taste tests that found people preferred the taste of Pepsi. Coke repeated the test and, to their horror, had the same findings. In response, Coke changed their formula and launched New Coke—to disastrous results, costing an estimated $30 million by 1985.
But underneath this flop, said Rogayeh, was a failure on Coke’s side to structure tests properly.
Had Coke run the test in a realistic environment, for instance having a full can of Coke over 30 minutes with a snack rather than a sip over 30 seconds, they might have learned that while people like the initial sweetness of Pepsi, Coke fanatics don’t want that sugary taste over longer periods of time.
A digital example of this might be using a large language model to assess consumer sentiment from a survey that only asks what people didn’t like about a platform. The sentiment will obviously be negative—that has nothing to do with the LLM and everything to do with test structure.
Another example might be a productivity-focused SaaS; these are often accessed in the middle of the day in between tasks. A user satisfaction test should mimic that environment, for instance asking people to complete a small mundane task first before logging into the app. Only then can any AI analysis pick up legitimate trends rather than false triggers.
Step 4: Either insight or renewed hypothesis
Once the tests give you results, you can figure out what comes next. Either you repeat steps 2 and 3 again, or you can move to making decisions if you have enough data.
“Whatever kind of data that they are working with, they can actually start asking themselves, ‘what is a scientific methodology here that I can apply to the data that I have access to, in order to advance the strategic metrics that align to my business objectives?’” said Rogayeh.
Don’t just throw AI at it
AI can’t tell you what you want, it can only help you uncover patterns you ask it to look for. Rather than trying to add AI for its own sake, apply a scientific methodology to start with business needs. Then use AI to help you uncover the patterns that humans simply can’t detect.
“It really is about identifying hidden patterns—no human being on their own would have been able to comb through as much data and identify as many of the patterns that are correlated with the thing that you actually want to predict,” said Rogayeh.
STAY IN THE LOOP