I am building a coupon tool. The goal is highly practical: to sort a bunch of fast-food and restaurant coupons by their “actual value for money,” saving me the effort of comparing them manually.

The system was originally running quite well on a fried chicken brand. One day, I wondered: could I apply it to a pizza brand too?

I handed this investigation over to the AI. It did a round of research and reported back: this brand was undoable. Then it started talking about choice groups, underdetermined systems, lack of anchor pricing… I did not quite understand.

But I reasoned that there were plenty of brands I wanted to support anyway. If this pizza place did not work, I could just switch to another one. So I said, “Okay, skip it.”

It looked like I had made a clean, decisive call.

But the truth is, that “Okay, skip it” was built entirely on my lack of understanding. I had not truly absorbed a single word of what the AI just explained.

Later, I revisited the issue, and only then did I figure out what the AI had actually been saying.

The core of my tool calculates the “value for money” of each coupon and sorts them accordingly. The difficulty is that the coupon source only provides the total price of the entire bundle. It does not specify how much the fried chicken or the drink inside is worth individually. To rank which coupon is genuinely a good deal, the “implied unit price” of each item must first be reverse-engineered.

The approach is to treat each coupon as an equation—the total price equals the sum of each item’s quantity multiplied by its unit price. More coupons mean more equations, allowing the unit price of each item to be solved.

For the fried chicken brand, there were over nine hundred coupons, which meant over nine hundred equations. There were only thirty-something items to solve for. With equations far outnumbering the unknowns, the calculated unit prices were highly reliable.

The pizza brand was a completely different story. Almost every single one of its coupons was a “choice group”—“choose one of nine: one pizza,” or “choose one of five: pick any side dish.” The problem was that it was impossible to know which item a consumer would ultimately pick. The equations could not be formulated, so they could not be calculated.

After stripping out these choice groups, only six to eight usable coupons remained. Yet, there were fifty to sixty items to solve for (pizza sizes, crusts, flavors, plus a bunch of side dishes). Six equations trying to solve for sixty unknowns—mathematically, this simply cannot yield a meaningful answer. The resulting numbers would be completely random.

To make matters worse, the pizza—the most expensive item carrying the bulk of each coupon’s value—did not even have a baseline reference price to fall back on. Side dishes could still be benchmarked against standard prices, but the official pricing provided for the pizza was not a fixed number. It was a starting price, such as “from $X”—a range, not a single figure, making it impossible to use as a baseline.

So the conclusion was rigid: the pizza brand’s coupons were inherently choice groups. Changing the scraping source or scraping more data would not change this fact. Their value for money simply could not be calculated using this methodology.

Up to this point, the AI’s initial judgment was correct—when it said the task was “undoable,” it really was undoable.

So where was the problem?

Not in the AI claiming it was undoable, nor in my decision to skip it. The problem lay here: when I said, “Okay, skip it,” I had absolutely no idea why the AI said the task was undoable. I skipped it simply because I did not understand, and figured, “Well, it is not absolutely essential to include this brand anyway.”

And the cost of that lack of understanding was much more concrete than I had initially assumed.

In my mind, I treated this pizza place as having only two paths: either calculate the value for money, or skip it entirely. Since the calculation was impossible, the brand had to be skipped. But looking back on it later, I realized there was actually a third path: do not sort by value for money, just list the coupons one by one for me to browse. Even without the value-for-money metric, the tool’s utility to me was not zero—it still saved the effort of hunting down coupons manually and comparing their expiration dates.

I could not see this third path at the time precisely because I did not understand. Not understanding led me to treat a partially useful feature as completely useless, and throw it away.

And later I realized something even more embarrassing: understanding what the AI was actually saying would have required just one simple action—asking it to explain everything again in plain English.

That was it. I did not need to go study terms like regression or underdetermined systems. All I had to do was reply, “I do not understand this; say it again in terms I can understand.” This step was right at my fingertips and cost almost nothing. Yet I skipped it.

The reason I skipped it, frankly, was that I did not care. With so many brands available, skipping this pizza place was fine. Why expend energy trying to understand something I did not strictly need to build anyway? But “asking for plain English” was practically free—because I did not care, I skipped even this trivial step, and my decision ended up being built on a “lack of understanding,” rather than a “trade-off made after understanding.”

After thinking it through, I plan to go back and integrate this pizza brand using that third path: no value-for-money sorting, just listing the coupons for me to browse. That way, it remains a useful piece within the tool.

At the end of the day, I built this tool precisely to offload the massive effort I used to spend evaluating every coupon I saw. I just wanted to look at the results and choose. Whether or not it can calculate the value for money does not negate its usefulness.

And starting from this incident, I have noted down a rule for myself: the AI spouting a bunch of jargon I do not understand is not the problem; failing to ask it to speak human is. Not understanding can be remedied, but by not even asking, I threw away options that were otherwise right in front of me.