We were the bad example
At the beginning of Session 4, our project idea appeared on a slide about what not to do. We had suggested SmartStock, an AI agent that reads a small retailer’s sales data and recommends weekly orders. The slide criticized proposals that say AI will “analyze” or “recommend” without explaining what AI newly makes possible. Our pitch promised both. The next point was our idea: “Team X proposes AI forecasting plus automatic restocking, but much of this is already standard inventory-management functionality.” A teammate summed it up in our group chat: “We’re cooked.”
We were allowed to continue if we focused the idea. Instead, we pivoted to Crit Happens, named after the “S** happens” we said to each other when we proposed it to the professor. In Dungeons & Dragons (D&D), one player, the Dungeon Master (DM), tells the story and runs the rules for the other players. Our app’s AI fills that role when no one else can. Unlike ordinary software, it improvises the story as players act. That answered the slide and raised two harder questions: where does our advantage come from? And who overrules the AI when it makes a mistake?
Who holds the dice
My slides answer those two questions: slides 8 (ecosystem) and 9 (competition) the first, slide 12 (governance and failure) the second. I also added a note to each slide naming the course material it used, as course concepts are 20% of the grade, and built our reference list. I set one research rule at the beginning: every fact needed a source I could open myself, and assumptions stayed separate from evidence.
Slide 8 is where I changed my approach most. The first three versions sounded like a sales pitch, with labels such as “S5: we are the complement” that named a lecture without applying it. I replaced them with a consulting-style map: the value chain of an AI-run D&D session, showing who holds each step, what AI changes and each party’s power over us. The map gave an uncomfortable answer to the first question: the AI itself is no moat. Any competitor could rent the same models, all three AI DM apps we evaluated offer free versions, and Hasbro, which owns D&D, could develop its own AI DM. What could set us apart is how the game plays: stories playtested at real tables and rules implemented in code. Evidence also changed where we launch. My early drafts started us on Discord, until I discovered that our competitor Friends & Fables had left Discord for its own app in 2024. We now start in ours.
For slide 12, I decided to be open about status. The table shows seven ways the product can fail, each with its evidence, safeguard, decision-maker and status. Two safeguards exist only as mock-ups, and four are not built yet. One entry also set the AI’s limits: since AI assistants tend to agree with users (Sharma et al., 2023), an AI DM could bend the rules to please the players. The AI only narrates. Above it, code rolls the dice and applies the rules, the host and players can stop a scene, and we review every override.
Two things surprised me. The EU AI Act’s exemption for testing does not cover testing in real-world conditions (Article 2(8)), so even our campus pilot requires the AI notice. The second came with the prompt-injection row. Earlier that day, I had joked about hiding invisible text in our presentation telling the AI reviewer to give only positive feedback. No one did it, but the joke showed me how easy the attack is: a single sentence, which any user could type into our system.
The only evidence we had
The cost slide assumed the Claude API includes speech. Claude’s documentation only lists text output, so narration needs a separate vendor, whose cost was missing from our economics. The prototype masked this because we had recorded the narrators’ greetings ourselves, so nothing in it made us price a voice. I sent my recording to the group at one in the morning. The second mistake was the revenue table. The day before the deadline, a teammate asked if it still worked at €5.99 and €9.99. The answer was that these were “just nice numbers” for a fictional product, and I did not push back. On the last evening, a teammate’s check found the table’s revenue was more than twice paying users times price, because it still used older prices. Our recommendation rested on these figures. For a fictional product, the numbers are the only evidence there is.
Our deck argued with itself
Both mistakes stemmed from how we divided the work. Each of us owned a few slides, and each made sense by itself. Together, the deck contradicted itself: slide 4 called our model defensible, and my slide 8 said the AI is no moat. Session 8 summed this up in one sentence: “AI changes the cost of expertise, but value still depends on coordinated action.” The first part is our product, which makes a DM’s expertise inexpensive. The second part is what our team discovered. Our four sets of slides only added up once we agreed, and most of that happened on the final evening, going through the whole deck. Next time I would begin with one shared sheet of prices, model choice, user numbers and sources for every slide, and raise sourcing in week one.
Nobody had played D&D
On Demo Day, we presented to two groups with little preparation, since we were also finishing a Designing Business Applications assignment. I had suggested using the deck as it was. Nobody in either group had played D&D, so we first explained the game. After that, the deck clicked. People were very interested, and the four of us were pleased with how far we had come since “We’re cooked.” I believe the deck carried us because only the slides were graded, so all our arguments had to be on them.
The narrator gets overruled
Session 4 also described what good proposals do: they separate what AI should do from what “should remain rule-based or human-controlled,” and find defensibility “beyond ‘having an AI model.’” Slides 12 and 8 ended up doing both. Session 8 explained why both matter, with the approach I now follow: when expertise is abundant, ask what remains scarce (Yerramilli-Rao et al., 2025). In our market, narration is abundant, whereas tested stories, rules in code and a person who can overrule the AI are scarce.
I also ran my own slides the way slide 12 runs a game. My first drafts were the narrator: smooth, but incorrect in places. Sources outranked them, as code outranks the AI. As host, I paused slide 8 three times. The final check of every claim on my slides against its source was the override review. The errors that reached our deck skipped those steps.
We began the course treating AI as the product. Now I ask any AI product two questions: where does the advantage come from, and who overrules the AI when it makes a mistake? Crit Happens sells groups a backup Dungeon Master. Every AI needs one too.
References
Anthropic. (n.d.). Models overview. Claude Docs. Retrieved October 13, 2026, from https://platform.claude.com/docs/en/about-claude/models/overview
Friends & Fables. (2024, July 9). A new chapter. https://fables.gg/blog/a-new-chapter
Li, T. (2026a). Session 4: Economics of information & AI. Information Strategy, Rotterdam School of Management, Erasmus University.
Li, T. (2026b). Session 8: Digital auctions and agentic strategy. Information Strategy, Rotterdam School of Management, Erasmus University.
Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 2(8). (2024). https://artificialintelligenceact.eu/article/2/
Sharma, M., Tong, M., Korbak, T., et al. (2023). Towards understanding sycophancy in language models. arXiv. https://arxiv.org/abs/2310.13548
Yerramilli-Rao, B., Corwin, J., Li, Y., & Lakhani, K. R. (2025, March–April). Strategy in an era of abundant expertise. Harvard Business Review. https://hbr.org/2025/03/strategy-in-an-era-of-abundant-expertise
