Kill Two, Fund One: A Case Study in AI Leadership Decision-Making
When Nikkhil Gupta stepped in as CPTO at Printful, he inherited three AI initiatives already in development. His first move wasn't going to win him any popularity contests: he killed two of the initiatives.
“Instead of funding three, they can properly fund one,” he explains. It’s about focus for the teams and reducing commercial pressure on the business.
The situation
Printful is a print-on-demand company. Merchants design custom merchandise (think t-shirts, hoodies, bags) then push those designs to their Shopify or Etsy storefronts. Orders come in, Printful prints and ships.
Two of the three AI initiatives were aimed at design creation:
- Prompt-to-design tools. Merchants type a prompt, get a design, put it on a product.
- Fixing fonts in AI-generated images. Back then, large language models failed at typography. “They would do abstract normal designs, but if you ask them to write a bit of text in the design, they would fail.”
The third initiative tackled a real bottleneck: embroidery automation. Embroidery is complicated. You can’t just print a digital file. Someone has to manually convert it into a vectorized file that tells the needle where to move, how dense the stitching should be, what works on denim versus silk. Manual work. Slow. Expensive.
The decision
Nikkhil didn’t hand down a verdict himself. He reframed the question and gave it back to the teams.
“If you could only do one thing in the next six months, would this be it? And can you help me understand what value does this bring to our customers?” He positioned himself as “thought partner… advisor… a challenger.”
For the prompt-to-design tool, the team came back with the key data themselves. “About 70, 75% of the customers actually did not use Printful for designing. They would have their own brands and graphic and creative teams… [This was] a data point that the team who was actually in this initiative came back with. So they invalidated their own direction.”
For the typography problem, Nikkhil made a bet. He looked at the pace of AI development and decided: “Companies whose core business model is producing foundational models” would solve this faster than Printful could. Better to pause, watch the market, and license the tech later if needed.
But he didn’t just pause it and forget it. He installed a feedback loop: “We put that feedback loop of the next six to eight months to constantly and continuously monitor the market… If no one else solves it in six months from now, we will come back and solve it again.”
The bet paid off. “In six to eight months… someone actually came, uh, one of the core foundational model companies… and conclusively said, we can now put typography and text in images.”
The outcome
With the input from the teams, Nikkhil paused the two design creation initiatives. This allowed him to shift resources and funding to embroidery automation.
As a result, the team ultimately deployed deployed new technology into production and even filed a patent. “We were able to deploy a part of this technology in our production workflows,” Nikkhil says. Clear operational value, not just cool tech.
Killing projects without burning bridges
Sounds great for the company, but what about how the individuals involved took it? “Quite a lot of emotions get involved,” he admits. Teams had invested time, effort, passion. Nikkhil's approach was to create a space for an honest emotional response. Let them vent acknowledge the work. Then he introduced a different frame.
“If you are looking as the result to be a finished product, which you don’t have, yes you have sunk cost. But if you are looking at gaining something out of the experience, then let’s just zoom back and focus on what we learned, what we gained, and how this allows us to do something different and better in the future.”
He doesn’t use the phrase “sunk cost.” He focuses on learning. “What did we get out of it, if not a finished product?” Teams usually come back with good insights that transfer to the next build.
The three decision frameworks Nikkhil used
1) Two-way-door decisions (reversible vs. irreversible)
Use when: You can test, observe, and reverse. Move fast with a lightweight process.
How Nikkhil applied it: He treated the prompt-to-design tool as reversible. Instead of mandating the decision, he asked teams to validate their own work: would this move the needle for customers? The team came back with data showing 70-75% of customers didn't need the tool. But if the data had shown that customers were actually asking for it, they could have reversed course and funded it. "This was a two-way door decision. We would not have lost leadership by not investing in that technology."
Learn more: Jeff Bezos on one-way vs. two-way doors in Amazon’s 2016 Shareholder Letter (official site)
2) Explicit feedback loops (“mechanisms”)
Use when: You need the decision itself to include how you’ll learn. Define the signal, the interval, and who revisits the call.
How Nikkhil applied it: For the typography problem, he didn't just pause, he installed a 6–8 month monitoring mechanism. "We put that feedback loop of the next six to eight months to constantly and continuously monitor the market." The loop forced a checkpoint: did foundation model companies solve this? If yes, license it. If no, go back in. "This is a concept of mechanisms that I learned in Amazon—every process should have a feedback loop."
Learn more: Lean Startup’s build-measure-learn loop (official methodology page); AWS/Amazon material on Day-1 mechanisms and decision velocity
3) ROI framing (decide what to stop, and what to fund instead)
Use when: Portfolio choices compete for the same people and dollars. Compare options on expected value, time-to-impact, and confidence.
How Nikkhil applied it: He required teams to return with “what we should be doing instead.” Not just “we shouldn’t do this”, but “we should do that.” In his case, Nikkhil proposed a 15 to 20 day sprint to redefine the investment with better ROI clarity.
The questions he asked: “Is this the right thing to be doing right now? And if not, what is the right thing to be doing right now?” Data-driven when possible. Anecdotal when necessary. But always anchored to customer value and business viability.
Learn more: - Cost of Delay — quantify the economics of waiting (Scaled Agile glossary) - RICE — Reach, Impact, Confidence, Effort (Intercom’s original post)