Why the most important AI skill might be the oldest one we have

The more time I spend helping association teams work with AI, the more convinced I am that the skill that matters most is not prompt engineering, and it is not knowing which tool to buy. It is asking good questions. That may sound a little old-fashioned for a post about new technology; after all, curious people have been asking good questions for a very long time. But when I look closely at a disappointing AI result, or at a confident answer that turned out to be wrong, I usually find a question that nobody stopped to ask.

Questions show up at two moments in any AI task:

• Before the AI responds. What am I actually trying to accomplish? Who is this for? What does "good" look like, and what should the AI never do?
• After the AI responds. How do I know this is true? What is missing? Whose perspective is not in the room? Would I put my name on this?

So this post looks at both moments, why they matter more as AI gets more capable, and what the research says happens when we stop asking. I have also included two concrete examples of each, so you can try them on your own work this week.

In my white paper, Stop Teaching Your Team to "Use AI", I describe a pattern I kept seeing in association teams, which I call the vending machine stage: put in a request, take out a result, move on to the next task. It is a natural place to start (it is where I started, back in 2023, with a clumsy request to draft an email reply). The trouble is staying there, because the vending machine stage has no questions in it. You ask for a thing, and you get a thing.

The Lada Consulting AI Fluency Model breaks fluency into four domains: Task Orchestration, Instruction Crafting, Output Judgment, and Responsible Application. When I look back at the model now, I notice that every domain is built on a question:

• Task Orchestration: Should AI be part of this task at all, and if so, a chat assistant or an agent? Where does it have to stop?
• Instruction Crafting: What does the AI need to know to do this well, and what should it never do?
• Output Judgment: What would I keep, change, or throw out, and how do I know?
• Responsible Application: Can I show how I checked the facts and protected member data?

The two domains I call the tactical loop — Instruction Crafting and Output Judgment — are the ones most people run dozens of times a day, and they are where the quality of your questions shows up fastest. So those are the two moments this post focuses on.

One more point from my white paper that matters here: AI agents raise the stakes. A vague question to a chat assistant gets you a vague paragraph that you can fix before anyone sees it. A vague instruction to an agent that can touch your membership database, or send a message on its own, can produce a bad action. The better your questions before and after, the less you are trusting luck.

Handing a task to AI is not the problem. I hand AI tasks every day, and I think most association professionals should be doing more of it, with guardrails. The problem is when we hand over the thinking along with the task. I've seen more and more research that suggests that is easier to slip into that trap more than most of us would like to admit.

Here is some of what I have been reading:

• Confidence in AI can crowd out critical thinking. Researchers at Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using generative AI at work. They found that higher confidence in the AI was associated with less critical thinking, while higher confidence in one's own abilities was associated with more. They also found that the thinking did not disappear so much as change shape, moving toward verifying information, integrating responses, and overseeing the task (Lee et al., CHI 2025). In other words, the job becomes asking good questions of the output.
• Offloading attention has a cost. In Harvard Business Review, David Rock of the NeuroLeadership Institute warns that when we consistently offload our attention to AI, "we miss opportunities for deeper thought, and sacrifice our moments of insight" (HBR, December 2025).
• Leaders are already seeing it. A 2026 BCG survey of 70 C-suite executives found that half are already observing de-skilling in their organizations, and nearly 90 percent pointed to uncritical acceptance of AI outputs as a symptom. The skills they named as most at risk include judgment, problem framing, and analysis (BCG, June 2026).
• Polished is not the same as useful. Another HBR piece coined the term "workslop" for AI output that looks polished but lacks substance, which shifts the real thinking onto whoever receives it. In their research, 41 percent of workers had received work slop, and each instance took close to two hours to clean up (HBR, September 2025).
• Ownership fades, too. An MIT Media Lab study, with a small sample of 54 participants, found that people who wrote essays with an AI assistant reported the lowest sense of ownership over their work and had trouble accurately quoting what they had just written (Kosmyna et al., 2025). (Even though it's a small sample, it matches my own and my clients' experiences when using AI in this way.)

The encouraging part is that the same research points toward the fix. A July 2026 HBR article on designing AI that strengthens human reasoning recommends practices like "reverse prompting," where the AI asks the person questions instead of simply handing over an answer, along with building in AI-free time for independent thinking (Tamisier-Fayard, Evgeniou, and Fayard, HBR).

The lens through which I see this is the member on the other side of our work. Associations exist, in large part, because they hold and grow a field's knowledge: they convene experts, capture what subject matter experts know, build bodies of knowledge and competency frameworks, and understand the context their members work in. If staff and volunteers gradually stop questioning what AI hands them, that knowledge does not just sit still; it becomes less robust, one unexamined draft at a time. Members come to us for judgment they can trust, and judgment is exactly what we risk offloading.

A good prompt usually starts with questions you ask yourself before you type anything. Then, once you have answered them, you can invite the AI to ask you questions too, which is a simple way to practice the reverse prompting the HBR authors describe.

The vending machine version:

You will get something, and it will probably read like every other renewal email in your members' inboxes.

The questions to ask yourself first:

• Who exactly is this for? First-year members, long-time members, or people whose membership lapsed last month?
• Why do people in that group tend not to renew? (If you don't know, that is a question worth asking your data or your membership team before you ask AI anything.)
• What is the one action I want them to take, and what would make it easy?
• What should this email never say or promise?

The questioning version:

That last line changes the conversation. The AI's questions often surface gaps in your own thinking before a single word of the email is written.

The second moment is after you have the output. This is where most people stop, especially when the result looks good. It looks great, and that is exactly when to slow down.

Say you asked AI to summarize 200 open-ended member survey responses, and it hands you a clean list of themes with a line like "72% of respondents mentioned professional development." It looks great, and that is exactly when to slow down. Questions to ask, of yourself and of the AI:

• Where did that number come from? Open-ended responses don't usually produce a clean percentage. Ask: "Show me how you calculated 72%, and list ten responses that support it." If it can't, the number may be invented (the self-check in my white paper uses exactly this scenario: catching an invented retention statistic before it reaches the board deck).
• What did you leave out? Ask: "Which themes appeared in fewer responses but might matter to a smaller group, such as students or retired members?"
• Does this match what I know? Compare the themes with what your membership and education staff hear every day. If something surprises you, that is worth a follow-up, not an automatic yes.
• Could I defend this to the board? If a board member asked how you know, would you have an answer beyond "the AI said so"?

Now say you have set up an agent to assemble segmented renewal reminders from approved templates and queue them for staff review. The drafts look fine. With an agent, though, the output is only half the story; you also need to judge what it did.

Questions to ask:

• Which records did it touch? Check the activity log, if it keeps one. Did it stay inside the approved member segment, or pull from a broader list?
• Did it use only approved sources? Are the templates and data it drew on the ones you authorized, or did it pull in an older version of a benefit description?
• When should it have stopped to ask, and did it? If a member record was missing an expiration date, did the agent pause and flag it, or guess?
• What would happen if it got this wrong at scale? One bad email might cost you an apology, but four hundred of them can become a member-trust problem, which is why a person, not an AI tool, still hits send.

In both examples, the questions do the same job: they keep you in the role of the one who decides, rather than the one who approves whatever arrives.

None of this requires a budget or a technical background; it requires reps. If you want to start this week, here is a short list to try on one real task:

• Before you prompt, write down who the output is for, what "good" looks like, and what the AI should never do.
• End at least one prompt with "Before you answer, ask me up to five questions."
• Ask the AI to argue against its own answer, or against yours.
• For every number or fact that will carry your name, ask where it came from, and check it against a real source.
• For any agent, review what it did, not just what it produced.
• Keep one task a week AI-free, so your own thinking stays in shape.

If you lead a team, these questions are also worth building into how you work together: in a shared prompt playbook, in a review step for anything that goes to members or the board, and in the rubric from my white paper, which gives you a way to see where each person is today in Instruction Crafting and Output Judgment.

Here are five you can paste onto the end of almost any prompt to improve the results using the questioning philosophy shared in this post:

1. "Before you start, ask me up to five questions about anything you'd need to know to do this well." This is the reverse prompting idea from the HBR piece. The AI's questions often show gaps in your own thinking, not just in the prompt.
2. "What assumptions am I making in this request that I haven't tested?" This helps most with strategy or planning prompts, where it's easy to ask AI to confirm what you've already decided.
3. "If anything is unclear or missing, tell me what it is instead of guessing." This gives the AI permission to stop and flag a gap, which cuts down on confident filler.
4. "What would a skeptical [member / board member / first-time attendee] ask about this?" This keeps the end user at the center and brings up objections before your audience does.
5. "Give me two or three different approaches, and explain the tradeoffs of each, before recommending one." This keeps you in the role of the one who decides. You're choosing between options you understand, not accepting the one answer you were handed.

AI is getting better at answers every month. I am less sure it is getting better at knowing which questions matter to your members, your field, and your mission; that part is still ours. So I will leave you with a question of my own: what is one question you have stopped asking since you started using AI, and what might it be worth to start asking it again?

If you would like to go deeper, you can download the AI Fluency white paper or reach out about custom AI fluency training for your team.

Cheers, Cathy

In the spirit of Responsible Application: I drafted this with AI as a thinking partner and checked the sourced claims myself. The point of view and the examples are mine.

1. Goel, S., Martin, D., & Kaffe, C. (2026). When Everyone Uses AI, Companies Risk Losing Critical Skills. Boston Consulting Group.
2. Kosmyna, N., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, arXiv preprint.
3. Lada, C. (2026). Stop Teaching Your Team to "Use AI": The Lada Consulting AI Fluency Model. Lada Consulting LLC.
4. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking. Proceedings of CHI 2025. Microsoft Research and Carnegie Mellon University.
5. Niederhoffer, K., Rosen Kellerman, G., Lee, A., Liebscher, A., Rapuano, K., & Hancock, J. T. (2025). AI-Generated "Workslop" Is Destroying Productivity. Harvard Business Review.
6. Rock, D. (2025). What's Lost When We Work with AI, According to Neuroscience. Harvard Business Review.
7. Tamisier-Fayard, M., Evgeniou, T., & Fayard, A.-L. (2026). Design AI Systems That Actually Strengthen Human Reasoning. Harvard Business Review.