Two people have the same model, the same subscription, the same blinking cursor. One gets a usable draft in ninety seconds. The other gets bland mush, sighs, and concludes the tool is overhyped. The tool is identical. The skill is not.
Prompt literacy, the ability to frame a problem so a model can actually help with it, is turning into a baseline competence of knowledge work. The comparison that fits is the spreadsheet. When spreadsheets arrived, the machine did the arithmetic, but the payoff went to people who learned to structure a model, name the assumptions, and check the output. Owning Excel meant nothing. Building with it meant everything. AI is repeating the pattern, faster.
The evidence that instruction is a skill, not a setting
The strongest field evidence comes from a randomized trial of 758 consultants at Boston Consulting Group. On tasks that fell inside the model's capability, consultants using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and produced work rated about 40% higher in quality than colleagues with no AI. That is a large effect from a general-purpose tool. But the same study is a warning: the design deliberately included an arm that received prompt training, precisely because how you use the tool, not merely whether you have it, shapes the result. The model is not a vending machine. It is an instrument, and instruments reward technique.
The instrument metaphor also explains why the gains are so uneven across people. In a field study of over 5,000 customer-support agents, generative AI raised resolutions per hour by about 15% on average, but the effect ran up to roughly 34% for novices and close to zero for the most experienced. AI encoded the tacit know-how of the best workers and handed it to everyone else. Prompt literacy is, in part, learning to ask for that encoded expertise on purpose.
What the skill actually is
Prompt literacy is not a bag of magic phrases. It is a small set of transferable habits. State the goal and the audience before the request. Give the model the raw material it needs instead of assuming it can read your mind or your files. Show an example of what good looks like. Ask it to reason, or to critique its own answer, on anything that matters. And, most important, treat the first output as a draft to interrogate, not an oracle to obey.
None of that requires code. It requires the same disposition a good spreadsheet user brings: a habit of making assumptions explicit and checking whether the numbers make sense. The people who struggle with AI are rarely the ones who can't type; they are the ones who never learned to specify.
There is a second layer to the skill, harder to teach and more valuable: knowing when not to trust the answer. The same BCG study that measured the productivity gains also found that on tasks past the model's real competence, AI users performed worse, accepting confident answers that were wrong. Prompt literacy in its mature form includes the judgment to sense when you have wandered off the tool's map, and to slow down rather than ship. Framing the request well and doubting the reply well are two halves of the same competence, and the second half is the one that keeps the first from backfiring.
Confidence does not come with a login
As Ricardo Semler has told it, Semco ran factories led by engineers without degrees and turned away MBAs from excellent schools who could not run a meeting, on the principle that competence does not arrive with a credential. Prompt literacy sits in the same category. It is not conferred by an IT ticket or a job title. It is built by doing, and it is visible in the output long before it appears on anyone's résumé.
That is the uncomfortable news for organizations betting that a site-wide license will lift performance. Access is necessary and nowhere near sufficient. The BCG trial's own structure, testing training as a separate variable, is a reminder that the tool and the skill are two different purchases, and most companies have made only the first.
Teach it like arithmetic, not like magic
If prompt literacy is the new spreadsheet skill, then it deserves the same treatment spreadsheets eventually got: taught as a normal, expected fluency, embedded in real tasks, practiced until it is boring. The alternative is a widening gap between colleagues with identical tools, where a few quietly compound their advantage and the rest decide the whole thing was oversold. The model does not close that gap. Only the teaching does.
Sources
- Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Lakhani et al. (2025). Navigating the Jagged Technological Frontier (Organization Science). RCT with 758 BCG consultants: on tasks inside AI's frontier, GPT-4 users did 12.2% more tasks, 25.1% faster, ~40% higher quality; on tasks outside it they erred more. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Brynjolfsson, Li & Raymond (2025). Generative AI at Work (Quarterly Journal of Economics). Field study of 5,172 support agents: AI raised resolutions per hour ~15% on average, up to ~34% for novices and near zero for the most experienced. https://academic.oup.com/qje/article/140/2/889/7990658
- Peng, Kalliamvakou, Cihon & Demirer (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (arXiv 2302.06590). Developers with Copilot finished a coding task ~55.8% faster than the control group; later RCTs replicate larger gains for less-experienced developers. https://arxiv.org/abs/2302.06590