Eighteen cents of tokens per agreement, then 25 minutes of a senior partner’s time at $160 an hour. Your 25-minute review costs $66.67, pushing the total automation cost to 2.4 times the manual paralegal baseline it replaced. The model was never the costly part. The pay grade doing the checking was.

The big picture:

The token line on an Artificial Intelligence (AI) bill is falling, but the review line is not.

Claude Opus 5.5 bills $4 per million input tokens and $20 per million output. Cache hits cost 20 cents per million [5]. OpenAI’s gpt-6.1-sol sits at $2.00 and $10.00 per million on its standard lane [6].

Those are the numbers quoted in board decks. They are also the only line that keeps shrinking.

Human checking is priced somewhere else: the labor market. O*NET (the US Department of Labor’s occupation database) is the source. It puts the median wage for paralegals and legal assistants at $30.24 an hour, or $62,890 a year [1].

That works out to $0.50 a minute.

The exact same 25 minutes of checking costs $12.60 at that wage and $66.67 at the partner’s rate. What strikes me here: both numbers are true at once.

The automation did not fail. It was priced at the wrong grade.

The other half of this is that nobody can estimate review time from the inside. A controlled trial by METR (a research nonprofit) had 16 experienced developers finish 246 tasks [4]. Some of those tasks allowed early-2025 AI tools, and some did not.

With the tools, they took 19% longer.

Afterwards, those same developers said the tools made them 20% faster. The people doing the checking were the last to know it was slow, and that is the number I would put next to any pilot’s self-reported saving.

By the numbers

  • $30.24 — Paralegal median wage: The 2025 median for paralegals equals $0.504 a minute, making 25 minutes of checking cost $12.60 [1].
  • $38.81 — Compliance officer median: The middle grade for a compliance-gated workflow costs $0.65 a minute [2].
  • $21.53 — Service rep median: The floor grade for output a simple check can clear costs $0.36 a minute [3].
  • 19% — Measured AI slowdown: 16 developers across 246 tasks took longer with AI tools while believing they saved 20% [4].

What I’d watch:

The part I keep circling is a ceiling in minutes, not a rate.

The math starts from the manual cost per unit and takes out the token cost. Divide what is left by the price of a review minute. On our field note, the manual baseline comes to $27.85 an agreement, with 18 cents of tokens on the automation side.

  • $160 an hour: The ceiling is 10.4 minutes. The pilot spent 25, which is why it lost.
  • $38.81 an hour: The ceiling hits 42.8 minutes. A compliance officer can spend well over half an hour on the same document and still beat the manual cost [2].
  • $30.24 an hour: The ceiling reaches 54.9 minutes. A paralegal gets more than five times the checking runway of a partner [1].
  • $21.53 an hour: The ceiling expands to 77.1 minutes. This is the widest margin on the ladder for output a simple check can clear [3].

Between the partner’s desk and the paralegal’s, the allowable time moves by a factor of 5.3 for the exact same checking work.

What I am watching next is whether a vendor publishes review minutes per resolved unit beside the token spend. My read is that the first one to do it will be selling against rivals whose per-task costs only look lower because the customer’s own staff absorb the checking time.

The other thing is the wage tables. Token prices move in quarters. Median wages come out once a year.

A steep token cut against a flat wage pushes the review share of the bill even higher.

The catch

The ceiling is not a license to hand the document to the cheapest person in the building. Our field note’s fix was a rule-based check on the clauses the model kept inventing, not a junior reviewer. A badly checked indemnity clause costs far more than the review ever would.

The wage comparison is not perfectly equal, either. O*NET prints median wages, not loaded costs. Employer taxes, benefits, and overhead sit on top of that $30.24, while a $160 billed rate carries the firm’s margin.

My read: the real gap is narrower than 5.3 times once both sides are costed the same way. But it remains wide enough to change which pay grade sits on the review.

The METR result tracks software developers on early-2025 tools, not a contract desk. I would not move that 19% slowdown directly into a legal workflow. Instead, I would use it for the one thing it does show: self-reported review time runs optimistic.

At a glance

  • The Big Shift: The token line on an AI budget is falling while the review line remains priced by the labor market, meaning the person checking the output dictates the per-unit cost. The same 25 minutes of review costs $12.60 at a paralegal’s median wage and $66.67 at a partner’s billed rate.
  • Why It Matters: Cost per resolved work unit is what a finance team actually approves. An automation pilot can look cheap per token while costing 2.4 times the manual process it replaced because the review minute decides the final bill.
  • What I’d Watch: I am watching for a review-minute figure published next to token spend, and for the annual wage tables that reset the math.
  • Review ceiling: The maximum minutes a check can take before the automation costs more than the manual baseline.
  • Measured review time: The actual time spent checking, because the 16 developers in that trial took 19% longer while they believed they had been 20% faster.
  • Reviewer grade: The pay grade qualified to clear a specific error class, which shifts the ceiling more than the model choice does.
  • The Catch: Cheap reviewers still must be qualified, median wages exclude loaded employer costs, and the current slowdown evidence comes from software development rather than legal review.

Related reading

Sources

[1] O*NET OnLine, “23-2011.00 — Paralegals and Legal Assistants”, median wages (2025). https://www.onetonline.org/link/summary/23-2011.00 [2] O*NET OnLine, “13-1041.00 — Compliance Officers”, median wages (2025). https://www.onetonline.org/link/summary/13-1041.00 [3] O*NET OnLine, “43-4051.00 — Customer Service Representatives”, median wages (2025). https://www.onetonline.org/link/summary/43-4051.00 [4] Becker, Rush, Barnes and Rein, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, arXiv:2507.09089v2. https://arxiv.org/abs/2507.09089 [5] Anthropic, “Pricing”, Claude platform documentation. https://docs.claude.com/en/docs/about-claude/pricing [6] OpenAI, “Pricing”, platform documentation. https://platform.openai.com/docs/pricing