You will be able to explain why AI can be excellent at one task and poor at a similar-looking one.
Farah, a marketing executive at a Singapore bubble tea chain, has been using an assistant for months. It writes her social media captions, turns rough notes into tidy briefs and suggests twenty campaign names in a minute. So when her manager asks which of the chain's eight outlets should get the next promotion, she gives the assistant the sales spreadsheet and a page of notes from outlet managers and asks it to decide. The answer is confident and well argued. It also misses the one line in the notes explaining that the Bugis outlet's sales dipped because of three weeks of renovation, and recommends the wrong outlet.
To Farah, the two kinds of task felt about equally hard. To the model, they were very different. This lesson is about that gap, and why it is so easy to fall into.
In 2023, researchers from Harvard Business School and other universities ran a field experiment with Boston Consulting Group, published as "Navigating the Jagged Technological Frontier". Hundreds of BCG consultants worked on realistic consulting tasks. Some had access to the most capable AI model available at the time, and some worked without it.
One set of tasks involved creative and analytical work of the kind consultants do every day, such as coming up with ideas for a new product, planning how to bring it to market and writing about it persuasively. On these, the consultants with AI did better. They finished more of the work, finished faster, and their output was judged to be higher in quality.
Another task was chosen because the researchers expected it to sit just beyond what the AI could do well. Consultants had to work out a business problem from spreadsheet data and interview notes, where the right answer depended on noticing details the data alone did not make obvious. On this task, consultants using AI were less likely to reach the correct answer than those working without it. The AI gave plausible, polished, wrong recommendations, and many people accepted them.
So the same tool, used by the same kind of skilled people, made them better at one task and worse at another.
The researchers described the boundary between what AI does well and what it does badly as a jagged frontier. If AI ability rose smoothly with task difficulty, you could say "it handles easy work and struggles with hard work", and plan accordingly. That is not what the experiment found.
Instead, tasks that look equally difficult to a person can fall on opposite sides of the line. Writing a persuasive product pitch and reasoning through a messy business case both feel like skilled consulting work. One sits well inside the frontier, and the other falls just outside it. Farah's captions and her outlet decision are the same pattern on a smaller scale.
Everything in this course explains why the edge is jagged. Language tasks such as drafting, summarising and rewording play to what a model learned from huge amounts of text, as lessons 3.2 and 4.1 showed. Tasks that hinge on one specific detail, a precise figure, fresh facts or careful reasoning over data are where invented answers and quiet misses happen, as module 6 showed.
The catch is that the AI gives you no signal when you cross it. Lesson 6.1 made the point that the confident style is the same for right and wrong answers. In the experiment, the outside-the-frontier answers were well written and convincing. People who had grown used to good results on other tasks had little reason to doubt them.
That is the real risk. Success on one kind of task builds trust, and the trust carries over to a neighbouring task where it has not been earned. The better the tool has served you, the more easily that happens.
General claims will not tell you where the edge sits for you. A vendor's demo, a colleague's success story or a headline about what AI can now do describes someone else's tasks. Even the experiment above tells you that the edge exists and roughly what lies on each side of it, not exactly where your own tasks land.
The only reliable way is to test on your own work, using tasks where you already know the right answer or can check it. Lessons 5.4, 6.4 and 7.5 gave you three such tests. Module 8 builds on them. Lesson 8.2 gives you questions to score any task, and lesson 8.4 turns the scores into a map of your own job.
Start by looking for pairs. Somewhere in your week there are probably two tasks that look much alike, as Farah's did, where you suspect the assistant would do one well and stumble on the other. Spotting those pairs is the first step to seeing your own edge.
Name two tasks from your job that look similar but where you suspect AI would do one well and the other badly, and explain why.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).