Why I Ignored AI for Two Years and What Changed

I ignored AI for two years because, for serious professional work, the early tools weren't worth the risk. They were sloppy with facts, vague on privacy, and wildly overhyped. What changed wasn't my appetite for risk. Claude and a few peers crossed a line: accurate enough, controllable enough, and useful enough that not using them now is its own kind of liability.

For a long stretch, almost every AI headline described something that dazzled in a demo and fell apart in the hands of a working professional. Legal briefs citing cases that didn't exist. Support bots that escalated anything more complex than a password reset. Image tools quietly adding extra fingers and teeth. Email assistants that managed to sound both robotic and oddly overeager. Every few weeks someone announced that "this changes everything," followed a month later by a quieter note about limits, guardrails, or an unforced error.

If you watched that cycle and chose to sit out the first round, that wasn't fear. It was judgment. I did the same, eyes open, and with the information on the table at the time, I'd make the same call again.

What were the legitimate reasons to wait?

Two years ago, most general-purpose AI tools weren't safe to trust with work you'd put your name on. They were entertaining, occasionally useful, but not something you'd anchor client-facing work to without triple-checking every line.

The accuracy problem had a particular shape. Early models answered with a smooth confidence that bore almost no relationship to whether they were right. A consultant asking about market structure could get a fluent overview with the wrong revenue numbers and a "trend" that had already turned. A litigator asking about precedent might see a polished citation to a case that never existed. Everyone makes mistakes. The danger here was that the system wrapped its mistakes in the kind of certainty we've been trained, over decades, to treat as a proxy for competence.

Privacy was muddy too. In 2023 you couldn't get a straight, well-documented answer to the questions that matter in real practice: Who can see this data? Is any of it training the next model? What happens if I paste in a client term sheet or a draft merger agreement? For anyone working under ABA rules, GDPR, HIPAA, or plain fiduciary duty, "trust us" is not a policy.

And the hype didn't help. When every product launch is "transformative," the word stops meaning anything. If you've spent 20 or 30 years filtering sales pitches and grand claims, you treated AI like any other shiny tool: watch, wait, see what survives contact with reality. That wasn't dragging your feet. That was doing the job.

What actually changed between 2023 and now?

Three things shifted enough to move AI from toy to tool in my day-to-day work: model quality, control, and the evidence coming from people with real responsibility.

First, the core models got materially better, especially on the tasks experienced professionals actually care about: digesting long documents, following layered instructions, and sticking to the facts where facts exist. Claude crossed a line for me when it became comfortable with 100-plus-page inputs without dissolving into vague summaries. The hallucination problem didn't vanish, but it became manageable. You learn where to trust, where to spot-check, and where to say "show your work," the same way you learned years ago which associate to trust on what.

Second, the control surface improved. You can now tell Claude, "Use only what's in this policy document and these three PDFs," and it largely behaves. You can disable training on your data. You can get enterprise terms your GC can read without laughing. Still not point-and-forget, but you're no longer tossing sensitive material into a black box with a nice website.

Third, the source of the stories changed. In 2023 most of what you heard came from founders, vendors, and tech press, whose job is to make things sound exciting. By late 2024 and into 2025 the stories started coming from a different group: a 51-year-old FP&A lead who cut month-end variance analysis from two days to four hours; a partner at a regional firm who now reviews first drafts instead of writing them from scratch; a wealth advisor who walks into every review with a Claude-built one-page scenario sheet instead of a scramble of notes. Those people have no incentive to oversell. They care about billable hours, error rates, and client retention. When they give you specific numbers, that's evidence, not aspiration.

How does this look for someone with 20 to 30 years of experience?

If you're in your late 40s, 50s, or early 60s, your advantage isn't out-typing a junior. It's pattern recognition and judgment: what actually matters in a contract, which metric is a red flag in a set of financials, how a client will react to a change in terms. That judgment is exactly what AI doesn't have. Claude can read a thousand pages faster than you can. It can't tell, on its own, which clause is politically sensitive in your client's industry or which board member will dig in on a covenant. That's your moat. What's new is pairing that moat with a tool that does in minutes what used to eat unglamorous hours.

Here's a typical pattern I see. A 54-year-old GC swore off AI after the first wave of hallucination stories. She bills at a high rate, carries heavy board work, and isn't shopping for toys. Her turning point wasn't a grand "AI strategy." It was a week where she used Claude to draft three first-pass markups on vendor agreements. She still did the real thinking: risk allocation, fallback positions, the political read of the counterparty. Claude handled the mechanical grind of redlines and cross-references. Net result: roughly three hours saved that week, which she spent not on more admin but on a long-delayed training for her business-unit leads.

What was the first thing that actually moved the needle?

My own turning point wasn't a slick launch event. It was a 40-page partnership agreement that landed in my inbox at 9:47pm, with a client meeting on the calendar for 8:30 the next morning.

Reading 40 pages that late is possible. Reading them well enough to walk into a meeting with clear eyes is something else. So I did what I'd do with a bright junior: I gave Claude a tight brief. I dropped in the PDF and wrote: "Minority partner, first-time investor. Flag any provisions that give the majority partner unusual control, any notice requirements that could bite my client on exit, and any economic terms structured to favor the majority in ways that won't jump out on first read. Don't summarize generically; point me to specific sections and quote the language that worries you."

What came back wasn't a substitute for reading the agreement, and I didn't treat it as one. It was a map. Three control provisions surfaced with section numbers. An exit clause where a 180-day notice requirement sat quietly mid-paragraph. A waterfall that looked like a clean 50/50 but, unpacked, leaned hard toward the existing partners if things went well. I still read the document. I still made the calls. But I did it with the hot spots already circled. Total prep: about 12 minutes with Claude and another 25 with the document, instead of the two hours I'd normally block. That didn't transform my life. It changed a Tuesday in a way I could feel.

	2023: reasons to wait	2026: what's true now
Accuracy	Confidently wrong; invented citations	Still errs, but predictably; you know where to verify
Control	Black box; unclear what it used	Scope to your documents; "show your work" on demand
Privacy	Murky data and training terms	Enterprise terms and no-train options; still your job to check
Evidence	Demos, founders, tech press	Working partners, CFOs, advisors with specific numbers
Your position	Sensibly skeptical, losing nothing	Skeptical and ready; waiting longer now costs you

What did I get wrong about "learning AI"?

My biggest mistake was treating AI as a subject to be studied before it could be used. A vague conviction that I owed it a course, an understanding of the architecture, some new skill set before I'd earned the right to touch the thing. That's backwards, and admitting it stung a little. You start with a task you already have. You describe what you need in the same language you'd use to brief a capable assistant. You read what comes back, notice what's off, and say so. That's the whole workflow. The first use doesn't need to be your most important task; it needs to be real, so you can judge the output against something you actually know. And you will correct it. That's the process, not a defect. Claude gives you a strong first pass, not a verdict. The correction is fast. The starting-from-a-blank-page is what used to eat the hours.

So is it actually too late to start?

No, and the "you're being left behind" framing is almost always wrong, usually sold by someone with something to sell. Two years of watching wasn't wasted. You arrive with skepticism intact and real questions about accuracy, privacy, and usefulness, which beats wide-eyed enthusiasm as a starting point. But waiting from here would be a different mistake, not out of fear, but because the tools are now good enough to do real work and the professionals using them well are compounding an advantage every week. You're not late. You're under-leveraged. Pick one real task on tomorrow's list, hand it to Claude with a brief as specific as the one above, and judge the result against what you already know. That single session will tell you more than two more years of headlines.