From Fantasy Hockey to Marketing Strategy: Lessons from One of AI's First Reasoning Models

In September 2024, OpenAI introduced o1, one of the first mainstream AI models explicitly built for reasoning rather than generation. At the time, most people were testing it on maths problems, coding tasks and benchmark questions. I decided to try something a little different: using it to draft my fantasy hockey team.

Looking back from when I built my fantasy hockey team with ChatGPT's early reasoning model, the exercise still feels oddly relevant, two years on. Many of the strengths and limitations we now associate with modern reasoning models showed up during that fantasy draft.

‍
RECAP: Why o1 Mattered

While reasoning models are common today, o1 was one of the first AI systems widely available to consumers that intentionally spent more time working through problems before responding.

It represented an early shift away from AI as a content-generation tool and toward AI as a decision-support tool.

That distinction matters because marketers, communicators and business leaders increasingly use AI not only to create content but also to analyse options, identify risks and support strategic decisions.

‍
My Fantasy Hockey League: The Basics

To provide some context, I play in a keeper fantasy hockey league. Here’s how it works:

• Keeper Format: Each team keeps a core group of six players from year to year, making every draft strategic as teams fill the remaining roster spots (6 keepers + 9 drafted players for a total of 15 players per team).
  ‍
  
• Team Composition: Since each team has a different mix of keepers—some might retain more defensemen or goalies, while others favor forwards—the draft is less about picking the best available player overall and more about selecting the best player to complement your keepers.
  ‍
• Balancing Stat Categories: Players accumulate stats in categories like goals, assists, and penalty minutes. Teams earn points based on how they rank in each category. For example:
  ‍
  • The team with the most penalty minutes gets the highest points in that category.
  • The team with the fewest penalty minutes gets the lowest points.
    ‍

The objective is to lead in as many stats categories as you can and ultimately, become league champion.
‍

Why This Was a Good Test of o1's Capabilities

Fantasy sports turned out to be an ideal test environment because success depends on many of the same capabilities organisations now ask of AI:

• analysing large datasets
• weighing trade-offs
• adapting to changing conditions
• identifying weaknesses
• making recommendations under uncertainty

In this case, o1 had to:

• Fill positional gaps: Whether I needed more defensemen, goalies, or forwards, o1 would need to factor in positional needs to ensure a balanced lineup.
  ‍
• Prioritise key categories: o1 would need to analyze my team’s strengths and weaknesses (hits, penalty minutes) and recommend players that could help improve those categories.
  ‍
• Anticipate competitors’ picks: With each team making selections based on their needs, o1 would need to predict what others would do and adjust the strategy accordingly.
  ‍
• Adapt to changes: During the draft, o1 would need to adjust the strategy and pick recommendations when players I wanted were taken by others, dynamically updating its recommendations in real-time.
  ‍

Much like building a marketing strategy and executing a campaign, o1 would need to analyse layers of data and predict competitor behavior. If o1 could handle this level of complexity in real-time, I knew it had potential for more advanced use cases.

Plus, I came in third to last in my pool last year, so it's safe to say my drafting skills have room for improvement!!
‍

The Setup: Feeding the Beast with Data
‍

To prepare for the draft, I wanted to maximize o1's understanding of player performance. Within ChatGPT, I gave the o1 model structured data from multiple sources:

• 2022-2023 player stats for 250 players
  ‍
• 2023-2024 player projections for 250 players
  ‍
• Standard deviation data and team performance metrics for 250 players
  ‍
• Average draft rankings for the platform we use for 300 players
  ‍
• 2024-2025 projected stats for 300 players
  ‍

I also shared information on my league’s setup, how points are awarded, the keepers of each of the other teams in the fantasy pool, and the unique roster positions each team had to fill.
‍

‍

The Strategy: o1 in Action

With all the data loaded, o1 went to work — and it didn’t disappoint! It quickly broke down my team's needs, identifying gaps in areas like right wing, defense, and penalty minutes, and prioritizing players to fill those weaknesses. It identified specific players I should pick for each round of the draft and alternates in case those players were selected by other teams.
‍

The Results – How Did I Do?

This tailored strategy gave me a clear edge throughout the draft. Here’s how it played out:

• Prioritizing Scarce Positions:  o1 emphasized the need to focus on right wingers and defense early due to my team’s gaps. Following its advice, I drafted Timo Meier (NJD - LW/RW) first, adding versatility and strength in goals, shots, and hits. Then, I secured Moritz Seider (DET - D), a defenseman who contributes in multiple categories, boosting my offense.
  ‍
• Balancing Stat Categories: o1 ensured I didn’t overinvest in any one stat category, guiding me to make well-rounded picks. It advised me to draft Alex DeBrincat (DET - LW/RW) to add versatility on the right wing. For defense, o1 recommended Shea Theodore (VGK - D), a player known for his strong assist potential and contributions in power-play points. Later in the draft, I added Cole Caufield (MTL - RW), a young winger with high upside in goals and power-play points.
  ‍
• Targeting High-Upside Players: Late in the draft, o1 recommended high-upside players like Jordan Kyrou (STL - C/RW), a multi-category contributor, and Pyotr Kochetkov (CAR - G), a young goalie with breakout potential.
  ‍

In comparison to past drafts, o1’s strategic insights gave me a clearer path to building a competitive and well-rounded team. I’m feeling much more confident about my lineup heading into the season, and I expect these moves to give me a strong chance at improving my performance this year.
‍

On reflection, what impressed me the most was that the model consistently applied a coherent strategy across multiple rounds of the draft.

‍

What Did I Learn? The AI Draft Debrief

While it’s too early to say if o1's strategy will lead to a championship season, the process revealed five key insights:

1. Reasoning Models Changed the Type of Work AI Could Do – o1-preview impressed me with its ability to create a detailed draft strategy based on a myriad of stats. It didn’t just spit out generic player rankings—it analysed my weaknesses (like PIMs and Hits) and recommended picks to address those gaps. This kind of high-level reasoning can be a game-changer for marketers too, especially when dealing with complex data sets like customer segments or multi-channel campaigns.
   
   ‍
2. Context Remains the Hidden Constraint – This experiment once again highlighted the importance of memory, context windows, and attention for Generative AI models.  
   
   With the large dataset I provided, o1 occasionally struggled to retain all the necessary details, which required me to periodically refresh or re-enter information. I would have been better off considering and managing what the model “knows” (data up until end of 2023) and what it doesn’t (new player data since its cutoff window) to maximise data inputs and management.
   
   o1-preview offers a significant advantage with its larger context window compared to GPT-4o. Back then it could handle 128k tokens (100,000 words), while GPT-4o processes up to 32k tokens (26,000 words). As you use o1 and GPT-4o, the difference in these context windows becomes obvious, allowing for more refined outputs based on your prompt as your chat with the model evolves (e.g. ChatGPT forgets less things you’ve told it with o1 versus 4o).  
   
   When it comes to Generative AI, not only do we need to see improvements in the models, but we need more compute power to handle more memory, larger context windows, attention in addition to inference.  The larger the capacity for these elements, the greater the improvements we will see in productivity.  
   Two years later, this lesson remains true. The quality of an AI recommendation often depends less on the sophistication of the model and more on whether it has access to the right information.
   ‍
3. Real-Time Decision Support – o1 wasn’t just a pre-draft tool; it helped me make picks as the draft unfolded. Its ability to analyze the real-time flow of the draft and keep me on strategy was invaluable—just as real-time data analysis is in marketing, whether you’re tweaking an ad campaign based on live engagement data or adjusting a social strategy mid-launch.
   ‍
   ‍
4. o1 vs. o1 Mini: Both Have Their Place - I alternated between o1-preview and o1-mini during the draft to manage prompt usage and maximize efficiency.  
   
   o1-preview excelled with the overall draft strategy — its deep reasoning abilities were ideal for this more complex, multi-step challenge where deeper chain of thought reasoning was a plus.  
   
   o1-mini was great during the draft for real-time analysis and pick recommendations. o1-mini proved faster and more efficient in the moment when I was on the clock, generating quick insights based on the draft strategy.
   
   For marketing and communications professionals, this experience highlights an essential takeaway when using these reasoning models – the importance of balancing the complexity of the task with the speed of response and cost of using the different models. This is not about comparing the better model but that different tasks require different levels of reasoning.
   ‍
   
5. Overkill for Simple Tasks - After the draft, I asked o1 to help me summarise the experience for this article.  It became clear that o1’s advanced reasoning was overkill for low-cognition tasks like basic content creation. While it excelled in complex problem-solving during the draft, its "chain-of-thought" reasoning made the output unnecessarily detailed, with a lot of over-explaining and rationalization —almost like it was overthinking my request.  Stick to the base models for such things.
   ‍

Post-Draft: Trade Recommendations and Future Strategy
‍

While the draft went really well, I was not able to address all the deficiencies on my team. I may have to do so via a trade with another team in the pool. o1 helped here as well, assessing potential trades and trade partners.

In the context of marketing or comms, Generative AI can assess gaps in a campaign or audience engagement (like gaps in my fantasy roster) and recommend strategies (similar to suggesting trades) to reallocate resources—whether that’s changing the messaging, focusing on a different demographic, or adjusting ad spend to improve overall performance.
‍

Marketing & Communications Tie-In: What Can o1 Teach Us?

So, what does my fantasy hockey draft have to do with marketing and communications? A lot, as it turns out:

• Strategic Insights – Just as o1 identified gaps in my roster and recommended targeted picks to fill them, Generative AI in marketing can do the same by identifying audience pain points or market opportunities. Today, communications and marketing teams routinely use reasoning models to assess campaign performance, evaluate audience segments, stress-test messaging, identify risks and support planning decisions.

• Real-Time Adaptation: One of the most impressive features of o1 was its ability to adapt in real-time. As players I wanted were drafted by competitors, o1 recalibrated and provided alternative picks on the fly. Similarly, in communications, Generative AI-powered real-time feedback can be invaluable when a campaign needs quick adjustments. Whether it’s pivoting key messaging due to public response and media traction, or a social media post based on live data, more sophisticated models can help you react faster.

• Contextual Awareness and Personalisation: o1 wasn’t just looking at isolated stats; it analysed my entire team’s makeup and needs within the context of the league. In marketing, Generative AI can take the same approach by examining individual customer journeys, segment-specific engagement patterns, or even the larger competitive landscape.
  

• Efficiency vs. Complexity: One of my key takeaways from using o1 was learning when to apply the right level of Generative AI power. o1 was fantastic for strategic planning, while o1-Mini was better for quick, real-time recommendations. The same applies to marketing: not every task requires deep cognitive AI. For simpler decisions like A/B testing or routine copy generation, leaner models might be faster and more cost-effective, whereas complex campaign strategy or customer segmentation may benefit from full-powered AI models.
  ‍

In short, just as o1 guided me through a successful draft, Generative AI can offer marketers and communicators the same level of detailed analysis, personalised strategy, and real-time adjustment capabilities.
‍

Final Thoughts

Looking back, the most interesting part of this experiment beyond the fact that AI helped me draft a better fantasy hockey team, was how clearly it previewed where AI was heading.

AI's value is not just about generating content. It's key value is in analysing information, weighing options and supporting decisions.

Two years later, that's become one of the defining shifts in enterprise AI. The organisations seeing the greatest benefit from AI aren't simply creating more content. They're using reasoning models to improve planning, strategy and decision quality.

My fantasy hockey team just happened to be an early test case.