The Switch on the Wall: The Hard Part Is Invisible

You don't have to be an engineer to turn on the light, and you shouldn't have to understand how large learning models work to use an AI-powered college and career planning assistant.

Last week, I was talking to a student about using Soar, the new college and career planning assistant I wrote about in April. She logged in and asked Soar a question related to college research. As she read the answer, under her breath, I heard her mutter, "*Wow*!"

Thinking about it later, I was reminded of a story about Shankaran, a man who worked for my great-uncle all his life. My father grew up in Vadakanchery, a small village in Kerala in the south of India. In a tradition that's thankfully now almost dead, Shankaran was the family's lifelong servant. In that capacity, he accompanied my Dad and great-uncle to the nearby town of Palakkad, where Dad was to meet my mom for the first time: the first step toward an arranged marriage.

Palakkad, unlike Vadakanchery, had electricity. Decades later, I remember Shankaran telling the story: someone touched the wall, and the lights came on. Shankaran lit the lamps in my great-uncle's house every night, and to him, electric light was a miracle.

Who knows if AI will transform the world in the way electricity did? Electricity changed everything about life and work. The families who use Soar are not experiencing the equivalent of electric light for the first time, but the simplicity of turning on a switch reminded me of that student's experience.

I speak to many audiences in the higher education counseling community about AI, and I've noticed there are three distinct approaches to it. While some are against AI (a defensible position, since there are families who want their children to be counseled the old-fashioned way), and some are *builders*, who want to get under the hood, build agents, and use tools to their maximum capacity, most people are *pragmatists*, who want the benefits without learning all the details. They want to turn on the switch and see the light come on, not build generators, transformers, transmission lines, and household electric systems.

What We Were Trying to Solve

That simplicity was my guiding principle as I designed Soar: I wanted something that students would use, and that would make the process of working with a counselor more *effective*, not just more *efficient*. That meant helping solve the challenges that counselors and students face in the process, and seeing where AI can help:

First, students are already using AI to research colleges — ChatGPT, Google, Reddit, TikTok, College Confidential — and they're doing it at ten o'clock on a Sunday night, without their counselor anywhere near the conversation. They arrive at sessions with opinions they can't fully articulate, conclusions drawn from sources they can't name, and sometimes misinformation they've internalized as fact. The counselor has no idea what they've been told. That's a problem that existed before AI and has gotten significantly worse with it.

The second is that we lose momentum between sessions. One great meeting, then a week goes by. The student doesn't always know how to make progress on their own. They come back. The counselor spends the first part of the session finding out where things stand. The work restarts rather than continues.

The third was scope. College counseling is bigger than college. It's majors, careers, scholarships, financial aid, and summer plans. Students and counselors have to navigate a thicket of details. It demands expertise across a range of subjects that no single counselor can always have at their fingertips, and students need answers to questions that don't always wait for the next scheduled session. One of my coaches, Claire Schadler, told me a story about a session with a student who suddenly said at the beginning of a meeting that she wanted to consider becoming an occupational therapist. Instead of spending a week doing research and losing the moment, Claire used Soar to outline the exact path to this kind of career with the student in a few minutes, and spent the rest of the time talking about it and meeting the teen where she was.

And finally, there's AI itself: parents and students want to know that teens are preparing for the life they are headed toward, where familiarity with AI — and the ability to use it ethically and well — are critical life skills.

What Happened When We Launched

Still, when I built Soar, having launched many software products before, I had modest expectations. Software rarely works when you launch it: there are a bunch of issues to be fixed and improvements made before it's really usable.

So the student's reaction was remarkable. I also heard from a Waldorf mom — deeply thoughtful about technology and appropriately skeptical — who told me she "doesn't love AI at all," then sent me an email after using Soar to say it was "truly miraculous."

These aren't people who were primed to be impressed. They were just people who encountered something that worked.

Where AI Helps, and Where It Doesn't

Building Soar has meant deciding not just where AI helps, but where it doesn't. AI introduces judgment into every answer, which is powerful when judgment is what you need, and a liability when you need consistency.

• Hallucination: Most AI tools hallucinate: they invent facts or misinterpret the results of their training. So in building Soar, I integrated databases from many reliable sources: the Common Data Set, the Integrated Post-Secondary Education Dataset (IPEDS), and the Department of Education's College Scorecard. I also integrated scholarship databases from reliable sources and databases about summer and enrichment programs. I also built in anywhere between 17,000 and 75,000 words of context to questions, so Soar has information about how college counselors work and provides data you can have much greater confidence in.
• Sycophancy: AI tools have a known — and not accidental — tendency to tell people what they want to hear. These tools are designed this way because people come back to something that flatters them. This sycophancy can mean overestimating likelihoods and being misleading about programs. To counteract that, I had to build a likelihood model that uses years of experience in working with students, and use our instructions to help students, parents, and counselors get the information they need, not what they want to hear.
• Voice: Many people now use AI as a writing coach, and it isn't a very good one. So as I approached essays, I had to think about the limits of using AI. When it comes to writing, Soar's approach is that human beings coach you to write: AI doesn't. So as I built Soar's writing hub, I integrated a process we've tested over many years: step-by-step approaches to self-reflection, personality and values assessment, brainstorming, drafting, and revision. Coaches assign work. Students write and reflect. Coaches respond with feedback. The back-and-forth between coach and student is the process. Soar is the container that makes it organized, trackable, and accessible — but the heart of it is the same human coaching relationship we have always had.

Building Soar took two and a half months of intense, focused work — and more than a decade of counseling knowledge before that. It required deep software engineering expertise. It involved agentic AI, Claude Code writing actual software, infrastructure that most people will never see and don't need to think about. I'm telling you this not to impress you, but to make a point: the sophistication belongs in the building, not in the use.

The Hard Part Is Invisible

Shankaran didn't need to understand electrical engineering to recognize that something had changed. The student who said "wow" didn't need to understand large language models. They just needed something that worked — something built carefully enough to deserve that reaction.