Automation vs Upskilling (in AI)

A common worry I hear about AI is that people using it too often will forget how to do their job. “Delegate your thinking to AI, and you will stop thinking,” they say.

I believe otherwise.

This belief stems from my experience, over ten years ago, with “delegating” spelling and grammar to Microsoft autocorrect. It was 2013, I was working for a multinational, and my boss asked me to follow some projects in Spain, knowing I spoke Spanish. The problem was that I had learned the language by living in Spain, so I could speak it fluently, but I had never learned how to write it well. My orthography was poor. Thankfully, Microsoft Outlook’s autocorrect kept flagging my mistakes and correcting them. Over time, that taught me how to spell Spanish correctly. Autocorrect didn’t make my Spanish worse by “automating spelling.” It made me better, because it corrected me every day, at the exact moment I made the mistake, until I internalized the right form.

Of course, this doesn’t always work (more on this later), but it led me to believe two things: one useful for individuals, and one useful for companies.

First, I believe that AI has enormous potential to lead to better thinking, especially when the user has skin in the game. A student who wants to get homework done in a subject they consider useless will resist any learning that the AI can provide. But when someone asks an AI a question whose answer genuinely matters to them, and the AI replies not just with a one-line conclusion but also with reasoning and context, the person will often learn a lot. I know because in the last 12 months, AI has taught me a lot by answering my questions, not on topics I’m already a top expert about, but on everything else.

Of course, this doesn’t happen automatically. The AI tool must be set to provide feedback in addition to just doing the job (or at least in addition to doing it).

How to do this? I could speak for hours about this, but some quick tips include:

• Use the AI not (just) to get things done, but to check their quality. What standards should your output meet, and does it meet them? And if not, why, and what can be done about it?
• If you or your employees use Claude Code, enforce “Explanatory output style.” If you use other tools, add to the “custom instructions” setting an instruction to “provide educational insights in between helping me complete tasks about why you’re answering or doing things a certain way, or on things I may have missed.”
• Add to the custom instructions for the AI to also provide feedback on your questions, when it notices they reveal some blind spot or imperfect framing of the root problem. And to always provide, together with the output itself, an assessment of its output and what it could have done better.

Second, I believe that, in a business context, AIs have outstanding potential to bring everyone in your organization closer to the skill level of your best employee (or of the best consultant you can afford), by learning from your best people and then applying their standards and best practices to everyone else.

Let me give you two examples of how this might work.

1. Modern AI models, such as Claude, can be taught “skills” (procedures, checklists, quality standards that they can then use when performing tasks, dramatically improving the quality of their outputs). Teaching an AI these “skills” requires no coding knowledge. Get your best employee at a given task to create these “skills,” and everyone else in your company using AI to perform that task will benefit from that. The result is an upskilling of most employees to the skill level of the best employee in the company.

2. Imagine a customer support team: a few agents consistently de-escalate tense conversations, ask the right diagnostic questions, and close tickets cleanly; others miss key details, escalate too late, or create extra back-and-forth. Traditionally, you spread skills with shadowing, call or ticket reviews, and manager coaching, which is slow, expensive, and hard to personalize. With AI, you can distill what your best agents do into explicit standards and playbooks, then give the rest of the team real-time drafts and AI coaching (questions to ask next, calmer phrasing, which option is policy-compliant, when to escalate), with clear guardrails so the human stays accountable.

Crucially, this doesn’t have to turn support into mindless “copy and paste.” Done well, it’s quality control and on-the-job training. AI can review a subset of outbound replies (or only high-risk ones like refunds, cancellations, or compliance-sensitive topics), score them against your rubric, flag what’s missing, and suggest improvements. The support rep remains accountable for what gets sent, but gets immediate feedback until the standards become habit. The work isn’t fully automated, but the floor rises, and outcomes become more consistent.

So, we have seen three possibilities among many:

1. Asking the AI to, with each major response, provide not only an output or answer, but also some insight and an evaluation of its own output and answer. This provides learning material for the user.

2. Getting the best person at a task to codify their skill in the AI so it’s reusable by others.

3. Using AI not just as a tool to automate work, but to provide Quality Assurance and coaching.

Of course, I don’t mean that automation won’t happen. Some tasks, and parts of some roles, will become partially or even fully automated (including parts of customer support, for example). My point is that beyond automation, there’s a major opportunity to use AI as a training tool: to raise the floor and help people learn faster in the flow of work.

That’s why AI’s biggest impact won’t be replacing your best people; it will be making “good enough” performance common. The winners will be the organizations that treat AI as a training and quality system, not (just) a shortcut.

If you are a leader looking to implement AI as a training and quality system in your organization, this is one of my advisory services.