The associate's invisible tax.

Law is almost entirely language work. Research memos, client letters, contract drafts, motion briefs, discovery summaries, engagement communications — a significant majority of what a practicing attorney produces is written language applied to a specific factual and legal context. Which makes it, on paper, one of the best-positioned professions to get real value from AI.

In practice, most firms are barely scratching the surface of what's possible. And the reason isn't the model's capability. It's a setup problem.

What the billing clock is actually running on

Ask any associate who uses Claude regularly what the first five minutes of a session looks like. They'll describe something like this: open a new conversation, type a paragraph explaining who the client is, what kind of matter it is, which jurisdiction they're in, what the firm's conventions are for this type of document, and what output format they're looking for. Then they start the actual work.

That five-minute setup happens every session. On a complex matter with multiple documents, it happens multiple times a day. Multiply that across associates, across matters, across a year — and you have an enormous amount of attorney time going toward re-establishing context that the tool immediately forgets.

That's the invisible tax. It's not dramatic. It doesn't show up on any report. But it accumulates, and it's entirely unnecessary.

The context problem, specifically

Claude doesn't know your client. It doesn't know the parties, the procedural history, the controlling jurisdiction, or the conventions your firm has developed over decades of practice in a particular area of law. Without that context, it produces output that is generically competent — the kind of draft that demonstrates the model understands how legal documents work in the abstract, but doesn't know anything specific about this matter, this client, or this firm.

Senior attorneys can often work around this because they've internalized all that context. They know how to prompt in a way that gets reasonably specific output. Junior attorneys haven't built that intuition yet, so their results are more inconsistent. The firm's AI output quality becomes a function of individual seniority, not institutional knowledge.

The goal isn't for Claude to understand the law. It already does. The goal is for Claude to understand your matter — before you type the first word.

What the language-heavy work actually looks like

The tasks where AI earns its keep in a legal context aren't the ones requiring legal judgment. They're the ones requiring legal language applied to established facts:

• First-draft correspondence. Client update letters, opposing counsel emails, demand letters following established firm tone.
• Research summaries. Synthesizing case law into a structured memo once the relevant authorities have been identified.
• Contract drafting. Standard clauses adapted to the specific deal terms, in the firm's preferred language, flagging non-standard provisions.
• Discovery organization. Summarizing document sets, identifying key facts, drafting deposition outlines from provided materials.
• Motion structure. Drafting argument sections once the legal theory is established, in the format and citation style the jurisdiction requires.

None of these tasks require the AI to exercise legal judgment. All of them require the AI to know the matter — the parties, the facts, the applicable law, the firm's conventions — well enough to produce output that doesn't require an overhaul before it's useful.

The consistency problem across a practice group

The invisible tax compounds at the firm level. When every attorney develops their own Claude habits in isolation, the firm ends up with as many workflows as it has attorneys. One partner has a prompt they've refined over six months. Three associates have different approaches. A paralegal does something else entirely. The outputs don't share a voice. They don't share conventions. They don't reflect a unified standard of practice.

This isn't a hypothetical risk. It's what I see in most firms that have been "using AI" for more than a few months. Adoption without architecture produces fragmentation.

A new associate joining the group should be able to open a matter folder, have Claude immediately understand the client, the context, and the firm's conventions for this type of work — and produce a first draft indistinguishable from one produced by someone who has been on the matter for six months. That's achievable. It just requires building the right layer first.

What good looks like

A firm running this well looks like this: matter folders have persistent context. When an attorney opens a client folder, Claude already knows the engagement. The parties are identified. The procedural posture is established. The firm's preferred language for this type of document is loaded. The citation format for the relevant jurisdiction is in place.

Draft output on a standard motion is close to production-ready on the first pass — not because the attorney typed a longer prompt, but because the context was already there. Time from "need a draft" to "ready for attorney review" drops dramatically. The attorney's role shifts from rewriting to reviewing, which is the appropriate allocation of their time and judgment.

When a new associate joins the group, they don't need to learn anyone's personal prompt techniques. The knowledge is in the files, where anyone can read it, update it, and use it.

Field notes

Legal work is among the highest-stakes language work I encounter. The precision requirements are real, and the cost of a bad first draft is real. Those requirements don't make AI less useful in this context — they make context architecture more important. The firms that get this right will have a significant productivity advantage over the ones still paying the invisible tax.

R.P.