Algorithm
Algorithm
What is an algorithm?
- Step-by-step procedures and rules for solving a problem
- The same problem can have very different efficiency depending on the algorithm
- The foundation of all software and AI
What an algorithm actually changes is that it turns work that depended on individuals into a question of 'what to do, in what order.' The moment you write down the reorder decisions a veteran buyer used to make by gut feel as a procedure, 'average sales over the last 4 weeks × lead time + safety stock,' that judgment can be verified, revised, and handed over. Only a written-down procedure can be improved. A standard that lives only in someone's head can't even be questioned when it turns out wrong. That's why the first step of DX is not adopting a system but writing out the business rules you use today in plain sentences.
In practice, you need to tell three things apart to have a useful conversation. A rule-based algorithm is one where people write the conditions by hand ('order when stock falls to 5 or fewer'). A learning-based algorithm finds the weights of those conditions on its own from past data. And an algorithm is not the same as code: the algorithm is the procedure itself, and code is that procedure implemented in a specific language. This is also why a developer saying 'let's change the logic' and 'let's refactor' are completely different conversations: the difference is whether you touch the procedure or only the implementation.
The efficiency gap is far bigger than it feels. Sorting 1 million records takes roughly 1 trillion operations with a simple comparison method and about 20 million with merge sort. A 50,000x difference passes straight through to server costs and screen response time. So a good algorithm is judged on three criteria: does it produce the correct answer, does it always finish on any input, and how cheaply does it produce that same answer.
The most common mistake in practice is believing that 'the algorithm picked it, so it must be objective.' An algorithm faithfully follows the target metric its designer set and the past data it was trained on. If you tune a recommendation engine to click-through rate alone, it grabs attention but pushes ill-fitting products to the top and return rates jump; if you hand a churn-prediction score straight to sales, discount coupons go to customers who would have stayed anyway. The metric you target defines how the algorithm behaves. When results look strange, the question to ask is not 'Is the model wrong?' but 'What did we tell it to maximize?'
Seungtae ParkHands-on AX instructor who teaches AI business development and marketing automation by building real systems
John KimInstructor John Kim, CEO of Nanum Consulting and author of “B2B Marketing Design”The word comes from the name of the 9th-century mathematician Al-Khwarizmi, and the concept became foundational to computer science.
Consider a hypothetical case. Company A sells household goods online, with 210 employees and 40,000 orders a month on its own store. The click-through rate in its product recommendation area is a decent 7%, but items bought via recommendations had a 12% return rate, more than double that of regular orders (5%). When the CEO ordered, 'Get return logistics costs under control within three months,' the data team took a fresh look at the recommendation logic from scratch.
- When Senior Manager Kim from the data team dug into the logic, the training objective was just one thing: 'click probability.' It looked at neither purchase nor return, and was built to push whatever caught the eye to the top.
- At a meeting, Manager Park from the CS team put up a screenshot and said, 'The top return reason for three months running is color or size mismatch, and those products are all sitting in recommendation slot #1.'
- The operations, CS, and data teams got together and changed the success criterion from clicks to 'purchases with no return within 30 days,' then added a rule excluding any product with a return rate of 15% or higher over the last 60 days from the recommendation candidates.
- Instead of rolling it out everywhere, we ran the new logic on just 20% of traffic for 4 weeks and compared it side by side with the old logic. Click-through rate slipped from 7.1% to 6.2%, but the return rate fell from 12% to the 7% range.
- After the full switchover, over three months the return rate for recommendation-driven orders dropped from 12% to 6.4%, and monthly return logistics costs fell from KRW 29 million (≈ US$21,700) to KRW 16 million (≈ US$12,000). That saves KRW 13 million (≈ US$9,730) a month, and recommendation revenue still grew 8%.