Pricing, licensing, and the partner who must sign off

A vendor demo closes with a slide that says "$50,000 per year, unlimited users." The managing partner nods. Six weeks later, the pilot dies, not because the tool was bad, but because the partner whose name is on the matter never signed off on using it for client work. The invoice was never the problem. Trust was.

This lesson looks at how legal AI is actually priced and licensed, and why that commercial structure matters less than the informal risk calculation happening inside every partner's head before they'll let a tool touch a live matter.

The three pricing models, side by side

Per-seat licensing. A fixed fee per named user per month, common for research tools and general drafting assistants. Thomson Reuters CoCounsel and Harvey both use variants of seat-based pricing for enterprise deals, though exact figures are negotiated per firm and not publicly listed (estimate, as of 2025-2026 vendor practice). Per-seat is predictable for budgeting but wasteful if associates barely touch the tool, and it does not scale down when matter volume drops.

Per-matter or per-document pricing. Common in e-discovery and contract review, where cost tracks the actual unit of legal work. A document review platform might charge per gigabyte processed or per contract analyzed. This aligns cost with client billing (many discovery costs are already passed through to clients as disbursements), but it makes budgeting harder because volume is lumpy and unpredictable across a year.

Consumption-based (API and token pricing). Charged per unit of model usage, typically tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → (the chunks of text a language model processes; roughly three-quarters of a word each in English). This is how firms building custom tools on top of OpenAI, Anthropic, or Azure OpenAI Service infrastructure actually pay. It is the cheapest model at low volume and the most dangerous at high volume, because nobody in the finance department knows what a "normal" month looks like until the first invoice arrives.

A worked example

Say a firm's knowledge management team builds an internal research assistant on a large language modellarge language modelEin Large Language Model ist ein KI-System, das auf riesigen Textmengen trainiert wurde, um Sprache vorherzusagen und zu erzeugen. Damit sind Aufgaben wie Schreiben, Zusammenfassen und Beantworten von Fragen möglich.Vollständige Definition ansehen → APIAPIApplication Programming Interface: eine standardisierte Schnittstelle, über die Anwendungen kommunizieren und Daten austauschen, ohne die interne Funktionsweise der jeweils anderen zu kennen.Vollständige Definition ansehen →. Assume, purely as an illustration:

• Average query: 2,000 input tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → (the question plus retrieved documents) and 500 output tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → (the answer)
• Illustrative rate: $3 per million input tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen →, $15 per million output tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → (rates vary by model and change frequently, treat this as an example only, not a current quote)
• 200 fee earners each run 15 queries per working day, roughly 20 working days per month

Monthly queries: 200 × 15 × 20 = 60,000

Input cost: 60,000 × 2,000 tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → = 120,000,000 tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → → $360

Output cost: 60,000 × 500 tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → = 30,000,000 tokenstokensEin Token ist die grundlegende Texteinheit, die Sprachmodelle verarbeiten, oft ein Wortfragment, ein ganzes Wort oder ein Satzzeichen und nicht ein einzelnes Zeichen.Vollständige Definition ansehen → → $450

Monthly total: roughly $810

That looks trivial next to a $50,000 per-seat license. But it scales linearly and silently. Double usage, double the bill, with no natural ceiling unless the firm sets one. This is why many firms wrap consumption pricing in a usage cap or convert to a seat license once usage stabilizes; the seat license becomes an insurance product against volume risk, not just a convenience.

Why the partner's calculation is different from the CFO's

The chief financial officer wants predictable cost per lawyer. The partner wants to know one thing: if this tool is wrong, whose name takes the hit.

A partner's professional exposure comes from several real sources:

• Duty of competence and supervision. Bar rules (in the US, ABA Model Rule 1.1 and its comment on technology competence) require lawyers to understand the tools they use well enough to supervise the output. A partner who signs a filing built on an AI-drafted clause they never checked is not protected by "the software did it."
• Client confidentiality. Uploading a merger agreement to a public chatbot is a potential breach of confidentiality obligations, independent of accuracy. This is the single fastest way a general counsel client relationship ends.
• Precedent for sanctions. US courts have sanctioned lawyers for filing briefs with fabricated case citations generated by AI (the *Mata v. Avianca* case in the Southern District of New York, 2023, is the most cited example). Every partner has heard of it by now.

So the partner's mental price of the tool is not the license fee. It is: probability of an undetected error, multiplied by the size of the disaster if that error reaches a client or a court. A $200,000 enterprise license with a verifiable audit trail and document-level citations can look cheap next to a "free" browser plugin with none.

The change management sequence that earns trust before the invoice does

Vendors sell tools. Firms adopt behavior. The sequence that survives contact with partners generally runs in this order, not the reverse:

1. Narrow, named pilot. One practice group, one tool, one task (for example, first-pass review of NDAs). Not "AI across the firm."
2. Partner-level champion, not just IT sign-off. A senior partner in that practice group needs to use the tool personally and be willing to say, in a partner meeting, "I checked this myself."
3. Visible error-checking, not blind trust. Early pilots should show fee earners the tool's mistakes, not hide them. Paradoxically, a tool that visibly gets something wrong and is caught by the review process builds more partner confidence than a tool that seems flawless, because it proves the safety net works.
4. Measured time saved, documented. Partners bill in hours (or increasingly, are asked by clients to justify fixed and value-based fees). "This cut first-draft NDA review from 90 minutes to 20 minutes across 40 matters last quarter" is a sentence a partner can repeat to another partner.
5. Formal firm policy before wide rollout. Written guidance on what can and cannot be uploaded, which tools are approved, and who is accountable, usually issued by the general counsel's office or a firm innovation committee.

Only after that sequence does the pricing conversation become real, because only then does the firm know actual usage volume well enough to negotiate seat versus consumption pricing intelligently.

Licensing details that quietly matter

Two clauses in the contract decide more than the headline price:

• Data usage and training rights. Does the vendor use the firm's prompts and documents to train its general model? Enterprise contracts from providers like Microsoft (Copilot), OpenAI's enterprise tier, and Anthropic's enterprise APIAPIApplication Programming Interface: eine standardisierte Schnittstelle, über die Anwendungen kommunizieren und Daten austauschen, ohne die interne Funktionsweise der jeweils anderen zu kennen.Vollständige Definition ansehen → typically contractually exclude customer data from model training, but firms must confirm this in writing, not assume it. This is the single most common point of confusion in vendor negotiations.
• Data residency and DMS integration terms. If the tool connects to the firm's document management system (DMS, the repository holding all client files, commonly iManage or NetDocuments), the license needs to specify where processed data sits, how long it is retained, and whether the vendor's subprocessors (their own cloud infrastructure providers) are disclosed. European firms face an added layer here under GDPR (General Data Protection Regulation), which restricts transferring personal data outside the EU without proper safeguards.

🎬 [VIDEO: "How Law Firms Are Actually Using AI in 2025" — youtube.com — search for recent legal AI adoption panels featuring practicing partners discussing real deployment friction, useful for seeing the partner's perspective in their own words]

Key Takeaways

• Per-seat pricing is predictable but wasteful at low usage; per-matter pricing aligns with billing but is volatile; consumption pricing is cheap at low volume and risky at scale, model your actual expected usage before choosing.
• Partners price risk, not cost: the real variable is probability of an undetected error multiplied by the damage if a client or court catches it, not the invoice line.
• Bar competence rules and real sanctions cases (Mata v. Avianca) mean a partner's name and license are on the line for AI-assisted work they did not verify.
• Trust is built through a narrow pilot with a partner champion who publicly checks the tool's work, not through a firmwide rollout announced by IT.
• Read the data-training and DMS-integration clauses before the price clause; they determine confidentiality exposure, which matters more to partners than the subscription cost.