Legal contract drafting: when every word is on the AI invoice

We covered contract review earlier in this series — the workflow of reading inbound contracts and flagging them against a company's standard terms. This post is about the other side of the same function: contract drafting. The generative work of producing first drafts of agreements, redlining inbound documents against templates, generating clause libraries, and increasingly negotiating directly with counterparties' AI systems.

The work is different in shape from review, but it lives in the same legal function, faces the same privilege and confidentiality constraints, and operates at similar volumes. And it is — like everything else in this series — a workflow where the cloud-LLM-default architecture creates costs, privacy issues, and quality problems that local inference would solve.

The work

Modern legal departments do a lot of generative drafting work.

Initial drafts of standard contracts: NDAs, vendor agreements, customer MSAs, partnership terms, statements of work, amendments, addenda. For a busy in-house legal team, the volume is in the hundreds per month.

Redlining inbound contracts against templates: taking a counterparty's contract and producing the negotiated version that aligns with the company's positions. This is generative work that requires understanding both the inbound document and the company's standard positions.

Clause library generation and maintenance: writing the model clauses that the company uses across its agreements. Updating them when laws change, business priorities shift, or counterparty patterns require new positions.

Negotiation drafting: producing the email or document responses to counterparty redlines. The "this clause is fine but we'd like to revise these three items" responses that constitute most of the actual negotiation work.

Specialized drafting: privacy policies, terms of service, employment agreements, IP assignments, settlement agreements. Each has specific structural requirements and specific risks if drafted poorly.

The aggregate volume at a serious in-house legal function is large, and it has been thoroughly AI-augmented in the past three years.

The math

A representative midmarket company with a small in-house legal team — three or four lawyers supporting a fast-moving business — processes somewhere between two hundred and five hundred contract documents per month across all categories. Each document involves multiple AI operations: initial draft generation, clause selection, redline production, response drafting.

A reasonable token budget per contract is fifty thousand input tokens (the template, the counterparty document, the negotiation context, prior history) and three thousand output tokens (the draft or redline). At frontier pricing, about $0.20 per contract.

Four hundred contracts per month at $0.20 each is $80 per month, or about $960 per year. Modest.

But the bill scales aggressively with company size and legal sophistication. A large enterprise legal function processes thousands of contracts per month across many business units. With richer AI workflows — clause library generation, complex negotiation modeling, regulatory drafting — the token consumption per contract goes up substantially. Bills at large legal departments are in the low to mid six figures per year.

For specialized legal operations — large law firms running document automation services for clients — the volume and cost is multiples higher. Top firms are likely at seven figures per year for their AI infrastructure.

Why this is structurally a local-SLM case

The properties for local-SLM suitability are present, and the privilege and confidentiality dimensions make several of them extreme.

The work is narrow. The model needs to know one company's standard terms, one company's risk positions, one company's negotiation patterns. A model fine-tuned on the company's contract corpus outperforms a general model on the company's drafting work.

The work is repetitive in structure. Contracts of each type follow predictable structures. Standard clauses recur across documents. Negotiation patterns recur across counterparties.

The volume is meaningful at any serious legal operation.

The privilege and confidentiality issues are at the extreme.

Attorney-client privilege protects communications between a company and its lawyers. Sending privileged contract drafts and negotiation discussions to a third-party cloud LLM creates legal risk that the privilege has been waived. Many corporate legal teams are explicitly forbidden, by their own general counsel, from sending privileged work product to cloud LLM providers — even ones with strong contracts.

Confidentiality obligations to counterparties protect their information shared during negotiation. Sending a counterparty's draft to a third-party AI provider, even for "internal" purposes like producing a redline, may violate the confidentiality agreement under which the draft was provided.

The brand-voice and standard-positions story matters acutely. A general model produces generic contract language. A model fine-tuned on the company's contracts produces drafts that match the company's specific risk positions, negotiation patterns, and structural preferences.

The audit trail matters for legal work specifically. Decisions about contract terms have material consequences — exposure to risk, allocation of liability, IP ownership, dispute resolution mechanisms. A local model that writes structured logs of what it considered and chose produces an audit trail useful for both internal review and external defense.

What the fine-tuned model knows

A model trained on the company's contract corpus knows what a general model cannot.

It knows the company's standard positions on every clause type. Indemnification, limitation of liability, IP ownership, payment terms, termination — for each of these, the company has specific preferred language, fallback positions, walkaway points. The fine-tuned model knows them.

It knows the company's negotiation history. Which counterparties tend to push hard on which clauses. Which counterparties accept the company's standard terms with minimal revision. Which patterns of pushback have historically predicted deal risk.

It knows the company's specific risk framework. Some companies are aggressive on certain clauses (high-margin businesses on payment terms, IP-heavy businesses on IP ownership), conservative on others. The fine-tuned model produces drafts that align with the company's specific risk posture rather than a generic legal template.

It knows the company's voice. Contracts at law firms read differently than contracts at startups. Contracts at consumer companies read differently than contracts at B2B enterprises. The fine-tuned model captures the company's specific drafting voice.

The output is drafts and redlines that the lawyer can use rather than rewrite. The editorial savings, across thousands of contracts per year at a serious legal function, are substantial.

What changes with local inference

A contract drafting workflow on a local SLM looks like this.

A model is fine-tuned on the company's contract corpus — historical agreements, standard templates, clause libraries, negotiation history. The fine-tuning happens in a controlled environment that preserves privilege and confidentiality.

The model runs on infrastructure legal controls. Contract drafts flow through the inference pipeline; redlines and responses get produced; negotiation drafts get generated. Nothing crosses the legal team's controlled environment.

The privilege story is intact. The cloud-LLM provider does not become a third party touching privileged communications. The waiver risk is eliminated.

The counterparty confidentiality story is intact. Counterparty documents do not leave the company's controlled environment.

The cost flips from per-contract to fixed. Contract volume can grow with the company's business activity without the bill spiking.

When the cloud LLM is still acceptable

A few cases.

For workflows operating on standard public templates without confidential or privileged content. Some early-stage drafting can be done with public templates and depersonalized content.

For very small legal operations — solo lawyers, small startups — where the infrastructure investment doesn't pay back. The cloud LLM works fine at small scale, with appropriate care about what is sent.

For workflows where the company's general counsel has explicitly approved cloud LLM use with appropriate contractual protections in place. Some legal teams accept the risk for the productivity benefit.

For most enterprise legal functions, the local-SLM case is strong on privilege grounds, on confidentiality grounds, and on quality grounds.

The pattern, in the legal function

Avery NXR is not a legal tool. It scaffolds Next.js applications. The architectural pattern repeats, with privilege and confidentiality dominating.

Legal contract drafting is a narrow, repetitive, meaningful-volume, privilege-protected, confidentiality-bound workload. The privilege case alone makes the cloud LLM architecture structurally difficult for most enterprise legal teams. The cost case adds reinforcement. The quality case — that fine-tuned models produce drafts that match the company's specific positions — makes the architectural shift attractive even for teams that aren't blocked on the privilege issue.

The legal tech vendors that build excellent local-inference tools — with appropriate fine-tuning, privilege-protective deployment models, and evidence packages that satisfy general counsels — will own the institutional legal AI market. The cloud-LLM-default products will hold parts of the long tail and the smaller business segment, but enterprise legal will pivot fast as soon as the tooling matures.

The pattern continues. Legal drafting is one of the workflows where the architectural choice is being driven by privilege and confidentiality concerns more than cost — but where all three dimensions support the same conclusion. The companies that move first will have stronger privilege positions, lower costs, and better drafting quality simultaneously.