The legal industry has spent a decade trying to reprice routine work. From Atrium's $75M collapse to today's AI-native law firms — the billing model keeps changing, but the structural inefficiency hasn't.
The core dysfunction in enterprise legal spend is not that legal work is expensive — some of it should be. The dysfunction is that inexpensive work is being completed by expensive resources, and every model the market has produced so far has failed to fully close that gap.
An NDA review is not a complex act of legal reasoning. A standard MSA redline against a known playbook is not partner-grade work. Yet the delivery chain for these tasks — whether it runs through Big Law, a mid-size firm, an ALSP, or one of the new AI-native firms — still involves a costly human doing most of the thinking, with technology on the side.
The question isn't whether AI can do this work. It's been clear for over a year that it can. The question is why the market still prices as though it can't.
Atrium launched in 2017 with a thesis that sounds prescient now: merge a technology company with a law firm, use machine learning to eliminate busywork, and charge predictable prices instead of hourly rates. Backed by $75.5M from Andreessen Horowitz, Y Combinator, and General Catalyst, the company grew to nearly 200 employees and 450 clients.
The pricing evolved through three phases — and each revealed a deeper structural problem.
Atrium's failure is often read as a cautionary tale about hybrid models. But the real lesson is narrower: subscription pricing absorbs unlimited demand against fixed revenue. When the technology can't dramatically cut fulfilment costs, you bleed out. Atrium's clients were getting $2,000+ of legal work for $500/month — the subsidy was unsustainable.
What Atrium got right was the diagnosis. What it lacked was the technology to make the economics work.
In the years after Atrium's collapse, the outsourced legal market consolidated around Alternative Legal Service Providers — companies like UnitedLex, Elevate, and Axiom that combine managed teams with process automation for contract work, compliance, and discovery at scale.
By 2025, 57% of corporate law departments engaged ALSPs directly. The model works — it delivers genuine savings over Big Law. But it works by making the same paradigm cheaper, not by changing it. ALSPs reduce cost by substituting lower-cost human labour (often offshore) and wrapping it in process discipline. Technology accelerates workflows but doesn't replace the humans in them.
The ALSP pricing model — blended rates, retainers, fixed-fee engagements — is still fundamentally labour-priced. Economics improve at scale, but they don't transform. And critically, the incentive structure for technology investment is muted: ALSPs capture efficiency gains as margin rather than passing them through as price reductions. If your ALSP adopts a better AI tool and saves 30% of review time, your invoice stays the same.
"If AI reduces a ten-hour drafting or review task to ten minutes, a firm billing hourly loses 98% of the revenue on that task. An ALSP charging a fixed fee keeps the gain. The incentives are fundamentally misaligned." — Razorhorse, ALSP Market Report, January 2026
Starting in late 2024 and accelerating through 2025–2026, a new class of firm has emerged — what Artificial Lawyer's Richard Tromans calls "NewMods" (New Business Model law firms). By March 2026, a dedicated directory tracks 27 of them. They share a common architecture: experienced lawyers, AI in the production workflow, and fixed per-document pricing.
The pricing is genuinely different from Big Law. Crosby charges approximately $400 per document. General Legal charges a flat $500 per contract, all the way to signature. Soxton claims $45,000 in average savings for clients in their first month. These are real numbers, and they represent a meaningful drop from what came before.
But here's the question worth sitting with: is $500 for an AI-assisted contract review actually cheap?
If the AI is doing the heavy lifting — clause extraction, playbook comparison, redline generation — and a lawyer spends 15–30 minutes on oversight, what justifies a $400–$500 price point? At those economics, the effective hourly rate for the human work alone exceeds $1,000/hour. That's Big Law partner territory — for what is, by the NewMods' own description, routine commercial work.
One reading: these firms have built genuinely differentiated technology and are pricing for the value delivered. Another: $500/doc is the price at which you can be profitable without having built differentiated technology at all — it's the price that lets you use off-the-shelf legal AI, wrap a lawyer around it, and still clear 40% margins.
General Legal reports approximately 40% profit margins on $500 contracts. That's a healthy business. But healthy margins at the current price point reduce the pressure to radically improve the technology. If you're already profitable on every document, the incentive to push AI from "accelerant" to "primary producer" is weaker than if you were pricing at $100 and had to make the technology work to survive.
At $500/doc, a NewMod with good lawyers and adequate AI can operate comfortably. At $100/doc, only a firm with genuinely transformative technology could be profitable. The pricing tells you something about the technology.
| Dimension | Atrium (2017) | NewMods (2025–26) | What would need to change |
|---|---|---|---|
| Pricing unit | Monthly subscription (uncapped demand) | Per-document flat fee (bounded) | Per-document with declining price curve as AI improves |
| AI role | Bespoke ML for document ingestion | Foundation models for clause analysis & redlining | AI as primary producer; human as exception supervisor |
| Human cost share | ~90% (lawyers doing the work) | ~50% (lawyers reviewing AI output) | <15% (lawyers supervising at decision points only) |
| Margin driver | None — subsidising labour with VC capital | Comfortable — ~40% on current volume | Technology-driven — margins from AI efficiency, not pricing slack |
| Client economics | Great (unsustainably so) | Better than Big Law, but still $500 for routine work | $50–$150/doc — reflecting actual cost of AI production + supervision |
Every model in this analysis — Atrium, ALSPs, NewMods — makes the same implicit assumption: the unit of delivery is a human doing work, with technology helping them do it faster. The billing model changes (hourly → subscription → per-document), but the production model doesn't. A lawyer is still the primary producer. AI is the assistant.
The structural shift that hasn't happened at scale is inverting that relationship: AI as the primary producer, with human supervision at the points where it matters. Not a lawyer accelerated by AI, but an agent supervised by a lawyer. The difference sounds semantic. Economically, it's the difference between a $500 review and a $50 one.
The reason a contract review costs $500 at a NewMod firm is that a lawyer still spends meaningful time on each document — reviewing, adjusting, signing off. The AI accelerates their work but doesn't replace the cognitive load. If you invert the model — the agent does the review, the lawyer supervises exceptions — you collapse the human cost per unit and the price follows.
Moving the primary producer from human to machine isn't just a technology problem. It's an architecture problem — and it's why the shift hasn't happened inside existing delivery models.
For GCs and CLOs at $1B+ revenue companies, the conversation is no longer "should we use AI for legal work?" It's: why are we still paying for expensive resources to do inexpensive work — and what would it take to stop?
The NewMod wave is a real development. It's pulling prices in the right direction and proving the market will accept non-traditional delivery. But it hasn't closed the structural gap between "AI-accelerated human work" and "human-supervised AI work." Until that inversion happens, enterprise legal teams are still overpaying for routine work — just less dramatically than before.
When evaluating any outsourced legal model — ALSP, NewMod, or agent-based — there's one question that cuts through the positioning: what percentage of the production cost is human labour?
If the answer is above 50%, you're in the old paradigm with a new price tag. The models that will define the next phase of this market are the ones that push that number below 15% while maintaining quality and accountability — and pass the resulting economics to the client, rather than banking it as margin.
The legal market has spent a decade repricing the same work through different wrappers. Atrium proved the appetite for predictable pricing. ALSPs proved the appetite for outsourced volume. NewMods prove that the market will pay for AI-assisted delivery at lower cost.
What none of them have fully proved is that you can move the primary producer from human to machine — and pass the resulting economics to the client, rather than capture them as provider margin.
That's the next phase. And the in-house teams that will benefit most are the ones already thinking about supervision — not just procurement.
This paper draws on publicly available data including: Atrium's TechCrunch and Business Wire disclosures (2017–2020); Crosby's Sacra funding analysis and Sequoia Capital portfolio materials; General Legal's Artificial Lawyer interviews and rate card (March 2026); Soxton's Business Wire announcement (December 2025); the Artificial Lawyer NewMod directory (March 2026); Razorhorse's ALSP market report (January 2026); Brightflag's 2025 Am Law billing rate analysis; Thomson Reuters 2025 ALSP study; and Harvard Law School Center on the Legal Profession's AI impact research (February 2025).
Cost-per-contract estimates are illustrative ranges based on reported pricing, hourly rate data, and typical scope assumptions for commercial MSA/NDA review. The "supervised agents" cost estimate reflects projected economics at sub-15% human labour composition and is not drawn from a single vendor's pricing. Actual costs vary by complexity, jurisdiction, and engagement structure.