Generative AI
Generative AI
What is generative AI?
- Creates new content (text, images, code)
- ChatGPT and Claude are the best-known examples
- Boosting work productivity
What generative AI really changes is less the quality of the output than where work starts. It used to be that half the working time went to staring at a blank screen, wondering what to write first. Now you start with a draft that appeared in minutes and decide what to fix and what to throw away. The person who used to write becomes the editor and reviewer, which paradoxically makes subject-matter knowledge and judgment in that field more important. If you can’t tell that a draft is wrong, you just get faster while quality gets worse.
There is also more than one technical approach. Text and code are handled by language models that predict the most likely next word, while images and video are handled by diffusion models that build a picture by removing noise step by step. Two constraints matter most in practice. First, a model can only read a limited amount at once (the context window). Recent models can handle hundreds of A4 pages, but anything longer must be split up. Second, models know nothing about events after their training cutoff, or about your internal documents. To have one work with internal policies or product materials, you need a separate setup such as RAG, which retrieves supporting documents and supplies them alongside the question. ‘Training the model on our company data’ and ‘having it find and read our company documents’ are completely different jobs in both cost and difficulty.
The most common mistake in practice is treating this tool like a search engine. Models are built to keep producing plausible-sounding text rather than say they don’t know, so they will confidently cite statute numbers that don’t exist or invent precedents and statistics. If that ends up unchecked in an executive report or training material, pulling it back costs more than writing the draft from scratch. The second issue is security. More and more organizations are flagged in audits after employees paste unreleased results, client lists, or contracts into a personal account instead of a company-approved one. The third is the assumption that ‘adopting it will raise productivity.’ Without usage guidelines and a review process, you just get more plausible-looking documents, and more work for the people who have to read and fix them.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayThe Transformer, the technology underlying today’s generative AI, was introduced in a 2017 Google research paper, “Attention Is All You Need.” Most models since, including GPT, Claude, and Gemini, are built on this architecture. ChatGPT’s release in 2022 brought it into the mainstream.
Consider a hypothetical case: Company A, an electronic components manufacturer with 640 employees. Its training team has 3 people, including the team lead, and runs 42 courses a year. Building one new course took 3 weeks on average, and when a business unit asked for a course, it often didn’t open for two months. In the second half of 2024, the team ran an 8-week pilot applying generative AI to course development.
- Manager Kim on the training team used a company-approved account to request ‘a draft 8-session curriculum on client audit response for second-year production engineers.’ In 40 minutes, they had the outline and learning objectives for each session, work that normally takes two days.
- Senior Manager Park from the production engineering team read the draft and said, ‘This is a service-industry manual, not our shop floor.’ So the team summarized 27 client audit findings from the past three years, fed them back in, and rewrote the practice cases to match the company’s own process standards.
- After two rounds of discussion with the information security team, they set a 12-item rule: mask client names, unit prices, and unreleased yield figures before entering them, and personal accounts are banned outright.
- During review, Manager Kim caught that 2 of the 3 Occupational Safety and Health Act provisions the AI cited had the wrong article numbers. After that, “verify laws, figures, and names against the source before marking them” was added to the training material approval checklist.
- By the end of the pilot, lead time for developing a new course fell from an average of 3 weeks to 9 days. Applied to 18 of 42 courses, it cut the training team’s average monthly overtime from 28 hours to 11. Satisfaction with those courses rose from 4.1 to 4.4 (out of 5).