Contract review and contract generation are the two highest-volume artificial intelligence use cases in legal practice. Spellbook owns the SMB contract generation category. Harvey ships generation and review as part of its generalist platform. Kira / Litera and Luminance are the long-tenured contract-review incumbents. eBrevia rounds out the specialist tier.

Every one of those platforms processes the firm’s confidential contract corpus inside the vendor’s cloud. For SMB and mid-market firms with matter mixes that touch engagement-letter restrictions, NDAs with AI-use clauses, outside counsel guidelines, or jurisdiction-specific confidentiality duties (US ABA Opinion 512, UK SRA, Canadian Federation of Law Societies, Australian Solicitors’ Conduct Rules), that vendor-cloud trust boundary is the deciding factor.

This guide covers the third path: private, self-hosted artificial intelligence for contract review and generation, inside the firm’s own tenant. Same workflow depth as Spellbook, Harvey, or Kira; confidentiality posture the contractual stack requires; SMB and mid-market economics. If you already know you want a managed private RAG deployment for the firm’s contract corpus, jump to our Private RAG service →

Contract artificial intelligence is two workflows, not one

Most vendor marketing collapses contract AI into a single bucket. On the firm side, it’s two distinct workflows sharing one corpus.

Review (incoming)

The firm receives a draft — NDA, vendor MSA, target’s customer agreement set, financing commitment letter — and answers: where does this deviate from market or the firm’s playbook?

Needs: clause extraction, market-terms comparison, anomaly detection against the firm’s playbook, disclosure-schedule reconciliation, cross-document analysis (“compare assignment provisions across 200 customer contracts”). Kira / Litera, Luminance, and eBrevia were built primarily for this.

Generation (outgoing)

The firm produces a first draft — NDA, employment agreement, vendor agreement, SOW, financing commitment, next-round purchase agreement.

Needs: first-draft generation from a matter brief, clause-library suggestion as the lawyer drafts, term consistency checks, redline generation against the firm’s preferred clauses. Spellbook is the leading vendor here (Word add-in, generation-first UX); Harvey and Luminance support generation inside broader platforms.

Firms running real volume need both. A platform that handles only one half is a partial solution.

Where Spellbook, Harvey, and Kira fall short for SMB and mid-market firms

The pattern: the vendor-cloud trust boundary is the structural problem, not the feature set. None of the three tiers resolve the contractual-stack constraint driving the actual buying decision at SMB and mid-market firms in 2026.

For the deeper M&A diligence workflow specifically, our companion guide on best legal AI tools for lawyers and law firms covers the vendor tier in more depth.

The contractual stack that’s reshaping vendor selection

The contractual environment around confidential contract work in major common-law markets is converging to restrict vendor-cloud AI processing. A stack of overlapping constraints:

• NDAs with AI-use clauses. Sharp rise through 2025–2026 in NDAs prohibiting confidential content from being uploaded to AI systems that may retain, expose, or train on it. Widespread across M&A, JV, commercial-licensing, and supplier negotiations.
• Outside counsel guidelines (OCGs). The Association of Corporate Counsel’s mid-2025 AI guidelines template is widely embedded in US enterprise OCGs and mirrored in UK / Canada / Australia. Standard provisions: AI-use disclosure, no public AI tools, no training / no retention, pre-approval per matter. Vendor-cloud platforms trigger an OCG review per matter; private deployments satisfy these by default.
• Bar duties across common-law markets. ABA Formal Opinion 512 (July 2024) requires US lawyers to evaluate disclosure risk before any confidential content goes into generative AI, with informed consent for self-learning tools. Six US state bars (CA, NY, IL, TX, FL, DC) have followed. UK SRA + Law Society, Canadian Federation of Law Societies + provincial bars, and Law Council of Australia have published parallel guidance under their respective conduct rules.
• Engagement-letter clauses. Firm-to-client letters are starting to include AI-specific clauses. Still emerging rather than standard.

For SMB and mid-market firms without a dedicated AI compliance function, an AI layer that doesn’t create the contractual problem in the first place is the simpler answer.

The architecture — private, self-hosted artificial intelligence for contract review and generation

One managed deployment, inside the firm’s tenant (or a firm-owned VPC in AWS, AWS GovCloud, the firm’s preferred region in London / Toronto / Sydney, or Azure), handles both review and generation on the same calibrated corpus. The contract content never leaves the firm’s perimeter.

Standard timeline: 4–6 weeks. Weeks 1–2 cover tenant deployment and SSO; weeks 2–3 the contract corpus ingestion; weeks 3–4 retrieval tuning and clause-library calibration. From week 4 forward, lawyers run review and generation workflows in production. The deployment runs continuously — the firm’s corpus grows over time and the system gets sharper at recognizing the firm’s market positions, preferred clauses, and matter-type patterns.

Five capabilities private self-hosted artificial intelligence gives the firm

1. 100% contract coverage with citations. Every review flag, clause suggestion, and redline links to a source paragraph — defensible to the partner, client, or bar reviewer.
2. Cross-document analysis on demand. “Compare assignment clauses across 200 customer contracts.” Queries that take associate-weeks run in minutes.
3. Anomaly detection against the firm’s playbook. Deviations surface automatically during review; matches surface automatically during generation.
4. Redline and first-draft generation against the firm’s preferred clauses. First redlines using the firm’s preferred positions; first drafts from a matter brief plus the clause library.
5. Matter-type templates by practice area. M&A, employment, commercial, financial-services, government-contractor, healthcare playbooks live as tunable templates; each matter uses the right one automatically.

The economics for SMB and mid-market firms

Where the private self-hosted model wins decisively against vendor-cloud licensing (Spellbook, Harvey, Kira / Litera, Luminance, eBrevia):

• Per-deployment cost: less than one year of major-tool licensing.
• Managed service materially below the vendor annual license.
• Contractual stack satisfied by default. Documents stay inside the firm’s tenant. No third-party data processor. OCG, NDA, engagement-letter, and bar-rule requirements pass review automatically.
• Cumulative cost diverges every year. The vendor license starts from zero every year; the deployment is already paid. By year two, the firm is operating at a fraction of cumulative vendor cost.

When private self-hosted artificial intelligence isn’t the right answer

• Light confidentiality posture, generation-only need. Spellbook fits.
• Top-tier firm with generalist-platform budget. AmLaw 100 firms without engagement-letter restrictions run Harvey as the default.
• Very-high-volume transactional practice. Past 200–300 matters a year, an annual Kira or Luminance license amortizes well.
• Firms with dedicated AI ops teams. Some larger mid-market firms operate their own platforms (and usually outsource ops to a managed partner later).

For the broad SMB and mid-market middle (30–300 lawyer firms, 50–500 contract matters a year), the private self-hosted model wins on economics, on the contractual stack, and on operational simplicity.

Frequently asked questions

Common questions from SMB and mid-market firms about deploying private, self-hosted artificial intelligence for contract review and generation.