AI Job Impact: Production vs Judgment Tasks

Last week, I wrote about how I built a two-person law firm that handles the workload of a much larger practice, almost entirely because of AI. The piece went viral—millions of people read it. The responses split into two camps. The first: this is incredible, how do I do this? The second, usually unspoken but unmistakable beneath the enthusiasm: if a two-person firm can do the work of twenty or fifty, what happens to the other eighteen or forty eight?

That second reaction is the one worth sitting with, because it leads to a question most people are framing wrong: "can AI do my job?" This is the wrong question. Not because the answer is no. For many tasks, the answer is increasingly, obviously, yes. It is the wrong question because it treats everything you do as a single undifferentiated thing, when in reality your work contains at least two fundamentally different kinds of "knowing." The market is about to price them very differently.

The duration of tasks AI can complete autonomously has been doubling every few months. Software stocks crater on automation announcements. Smart, accomplished professionals are lying awake running the same calculation: how long until the thing I do for a living gets done by a machine that costs $200 a month? The dread is understandable. But it is aimed at the wrong target.

Consider what most professionals actually do each day, and what they bill for. Research. Drafting. Analysis. Formatting. Pattern-matching against known templates. Synthesizing information from multiple sources into a coherent deliverable. This is skilled production. It requires training, intelligence, and discipline. It is genuinely valuable work. And it is being commoditized at extraordinary speed.

But there is another layer. The layer where you decide what to do when the answer is not clear. When the context matters more than the content. When you stake your reputation on a call that could go either way. That layer is not getting cheaper. It is getting more expensive.

The distinction is easier to see in practice than in theory. Here is what it looks like in mine:

A lead investor's counsel sends back a marked set of NVCA financing documents for a Series B round. The redlines run across all five agreements: the amended and restated certificate of incorporation, the stock purchase agreement, the investors' rights agreement, the voting agreement, and the right of first refusal and co-sale agreement. Here is what AI does with those documents today, and does well: it maps every material change across the full document set, flags where the lead investor shifted economics or control, identifies tensions between documents (where the protective provisions in the charter conflict with the drag-along mechanics in the voting agreement), and generates three alternative formulations for a disputed veto right. If you had given this same stack of documents to a second-year associate eighteen months ago, that cross-document analysis would have taken the better part of a day. AI does it in minutes. That is 90% of the billable time a junior lawyer would have spent.

Here is what AI cannot do.

It cannot decide which terms to fight for and which to concede, given the deal dynamics that exist only in this particular financing. It does not know that the founder just received a competing term sheet from a top-tier fund and has leverage she has not yet played. It does not know that this lead investor's push for a unilateral veto right over any exit below a 5x return signals that they likely intend to block any acquisition that is lifechanging money for the founder but not a big enough multiple for the fund. It does not recognize, from the pattern of their markup (for example, insisting on a 1.5x liquidation preference instead of 1x), that this investor is more concerned about downside protection than partnership, potentially because they are already underwater in a prior fund and cannot afford another write-down. And it has not negotiated a hundred of these financings over a decade, developing an instinct for which fights matter and which are posturing—for knowing that this particular firm always caves on the audited financials requirement but will walk away if you touch their pro rata.

Now, the obvious objection. Much of what I just described sounds like an information problem. Give the AI access to the email threads, the CRM, the prior deal history, the Slack messages. Feed it the context of the business relationship. With enough data, couldn't a sufficiently advanced model factor all of this in? Partially, yes. The hallway conversation could be a calendar note. The investor's negotiation pattern could be extracted from a database of prior term sheets. Information asymmetry is a temporary moat, and anyone building a long-term career strategy around "I know things the AI doesn't" is making a mistake. That gap will close.

But there is a layer beneath the information that does not reduce to data points. Reading those redlines and sensing that the investor is positioning for something undisclosed is not a lookup operation. It is a perception shaped by having been in that situation, with those stakes, enough times that the pattern registers before you can name it. The prior deals are not data I retrieve and compare. They are the lens through which I see this deal. Michael Polanyi called this tacit knowledge: we know more than we can tell.¹ You can feed a model the complete record of every deal I have ever worked on. What you cannot feed it is what it was like to be in those deals—to have the phone call at midnight when the transaction was falling apart, to learn from the one where your read of the counterparty was wrong and the client paid the price.

The AI produced excellent analysis. What it could not produce was the three-sentence email to the founder that said: accept the broad-based weighted average anti-dilution rights (they're market, and fighting on this will cost goodwill you'll need for other key terms), push back hard on the vesting reset (your co-founder might not stay if she has to start her four-year clock over, and you should push for double-trigger acceleration so the team is protected in an acquisition), and negotiate the threshold for the lead to keep its special rights down from 75% to 50% of their original position. That email drew on information no model had access to, required a commitment the client would rely on, and reflected a judgment call that could go wrong. It took four minutes to write. It was the most valuable thing I did that week.

The distinction between these two kinds of knowing is about to become the most important fault line in the professional economy.

You have felt this distinction before, even if you have never named it. Every experienced professional has had the moment where a colleague produces work that is technically competent, checks every box, follows the template, applies the right framework, and is somehow wrong. Not factually wrong. Wrong in a way that is hard to articulate. The emphasis is off. The recommendation is defensible but would never survive contact with the actual client, the actual market, the actual counterparty. Something is missing.

What is missing is the thing you cannot put into a sentence. It is the accumulated residue of every prior engagement, every negotiation that went sideways, every client who said one thing and meant another, every deal that cratered for reasons that were invisible in the documents. You do not consciously inventory these experiences when you evaluate new work. They are the lens through which you see.

This kind of knowledge cannot be written down as a set of rules, because it was never a set of rules to begin with. Stripe Press's documentary series "Tacit" explores precisely this: the kind of expertise that can be observed but does not compress into instructions. The ceiling for language models, they observe, appears to be conscious competence—the ability to capture and articulate procedure.² What they cannot capture is the unconscious competence that characterizes genuine mastery.

Watch a novice and an expert approach the same problem. The novice follows rules—frameworks, checklists, decision trees. The expert sizes up the situation and acts. They have internalized so many cases that the rules have dissolved into instinct. AI can encode the explicit rules that novices follow. What it cannot encode is the intuitive, situated responsiveness that experts have transcended those rules into. Ultimately, the latter is a structural feature of what expertise is, rather than a temporary limitation waiting for more training data.

Silicon Valley has spent the last two years arguing that "taste" is the human skill that will survive the AI transition. The argument goes like this: when anyone can generate code, content, and analysis with a prompt, what differentiates you is knowing what to build, what to ship, what matters. Taste, in this framing, is the ability to direct tools rather than just use them. It is product vision, design sense, the instinct for what resonates. "Taste is eating Silicon Valley," as one widely circulated essay put it.³

They are right that something in this territory is the durable human premium. But the word "taste" is doing too much work. It bundles together at least two distinct things that have very different futures.

The first is pattern recognition on quality. The immediate, pre-deliberative sense that something is right or wrong. A senior lawyer reads a draft contract and knows, within seconds, that something is off. A designer looks at a layout and feels the imbalance before analyzing it. This kind of taste is sophisticated and hard-won. It is also, fundamentally, about recognizing what "good" looks like across an enormous corpus of examples. AI is getting increasingly good at this. A model trained on millions of contracts can flag when a clause deviates from market standard. For pure pattern recognition on qual