Most firms start by buying. Legal research assistants, contract review tools and AI features inside practice management software are quick to roll out and need no engineering team. The trouble starts when a firm's most valuable work does not fit a standard tool, such as reading hundreds of pages of mixed evidence per case and turning them into an opinion in the firm's own format.
Law firms should buy AI for tasks that are the same at every firm and build custom AI only where their own documents, templates and judgment rules create the value, because that is where generic tools produce work that attorneys have to redo.
This article maps the legal AI market, gives a decision matrix for buying, configuring or building, and shows what a custom build looked like at a US personal injury firm, including cost, architecture and ethics requirements. It does not rank legal AI vendors, and it is not legal or ethics advice for a specific jurisdiction.
Key takeaways
- Buy AI for standardized work such as legal research, contract review and drafting common documents, where vendors have already solved the problem.
- Build custom AI when output must follow the firm's own templates, when documents are messy and practice-specific, or when volume makes per-seat or per-case pricing expensive.
- A custom case-preparation tool at a US personal injury firm processed bundles of about 250 pages for roughly $3–4 in model usage per case and improved attorney throughput about 1.5 times.
- Clio's 2025 Legal Trends Report found that 79% of legal professionals use AI, but 53% say their firm has no AI policy or are not aware of one.
- ABA Formal Opinion 512 applies the existing duties of competence, confidentiality, communication and reasonable fees to generative AI, whichever route a firm takes.
- The most common mistake is buying a general chatbot for practice-specific work and expecting attorneys to fix the output.
Where legal AI stands in 2026
Adoption is no longer the question. According to the Clio 2025 Legal Trends Report, 79% of legal professionals use AI in some form, with 87% at large firms and 71% among solo practitioners. Only 40% use legal-specific AI tools, down from 58% a year earlier, which means many lawyers rely on general-purpose assistants for legal work.
Governance has not kept pace. More than half of respondents said their firm has no AI policy or that they do not know of one. Among lawyers who use AI widely, 69% reported a positive effect on revenue, and 45% had adjusted their pricing.
The professional rules are now clear in outline. In July 2024, the American Bar Association published Formal Opinion 512, which applies existing duties to generative AI tools. Lawyers must understand the tools they use, protect client confidentiality, keep clients informed and charge reasonable fees. A lawyer may bill for time spent using an AI tool and reviewing its output but should not bill a client for time spent learning the tool.
The legal AI market in four categories
Before deciding whether to build, it helps to see what can already be bought. Legal AI products fall into four broad groups.
| Category | What it does | Typical pricing model | Strength | Limitation |
|---|---|---|---|---|
| Research and drafting platforms | Answers legal questions from case law and statutes, drafts memos and documents | Per seat, often bundled with a research subscription | Trusted legal sources and citations | Built for general legal work, not a firm's own case files |
| Practice-specific platforms | Handles one practice area end to end, such as demand letters and medical chronologies in personal injury | Per case or per document, sometimes per seat | Deep workflow for one practice | Output follows the vendor's format, data sits on the vendor's platform |
| Practice management with AI built in | Adds summaries, drafting and search inside case management or billing software | Add-on to an existing subscription | No new system to adopt | Limited to the data and features of that platform |
| General-purpose assistants | Enterprise versions of large language model assistants | Per seat | Flexible, fast to start | No legal sources, no firm templates, high risk of unchecked output |
Personal injury is a good example of how specialized the market has become. Platforms such as EvenUp combine demand letter drafting, medical chronologies, settlement benchmarks and integrations with case management systems like Litify, SmartAdvocate and CASEpeer. For a firm whose process matches the platform, buying is the faster route.
Buy, configure or build: a decision matrix
The right choice depends on the task, not on the firm as a whole. Most firms end up with a mix: bought tools for standard work, configured assistants for internal knowledge and a small number of custom systems for their core workflow.
| Question | Points to buying | Points to building |
|---|---|---|
| Is the task performed in a similar way at most firms? | Yes, such as legal research or NDA review | No, it depends on the firm's own method |
| Must the output follow the firm's own templates and style? | Vendor format is acceptable | Output must match firm templates exactly |
| How messy are the inputs? | Clean, standard documents | Scanned, handwritten and mixed evidence bundles |
| How many cases a month? | Low or irregular volume | High, steady volume where per-case fees add up |
| Where may client data be stored? | The vendor's platform is acceptable | Data must stay in the firm's own environment |
| Does the workflow give the firm an edge? | No, it is routine | Yes, it reflects how the firm wins cases |
| Does the firm have, or can it hire, engineering support? | No | Yes, in-house or through a development partner |
Configuring sits between the two. A firm can connect a general assistant to its own document management system through retrieval, restrict it to approved sources and add firm prompts and templates. This works well for internal knowledge questions and first drafts of routine documents, and it costs far less than a full custom build.
Weighing a legal AI tool against a custom build? We review your workflow, documents and volume, and recommend the route that fits, including when buying is the better answer.
See legal tech projectsWhat building looked like at a US personal injury firm
Brandon J. Broderick, a US personal injury law firm, prepares legal opinions for insurance compensation claims. Each case arrives as a bundle of about 250 PDF pages: medical records, police reports and notes, many of them handwritten. Attorneys used to read every page in order before drafting an opinion.
Why a general tool was not enough
The firm's first tests used a general chatbot with a single prompt. It could not handle the volume of a full bundle and could not keep topics such as medical findings and liability facts apart. The opinions also had to follow the firm's own templates, with key conclusions highlighted so that attorneys could verify them quickly. No general tool could meet those requirements without heavy manual rework.
What .wrk built
- Five specialized assistants. Each one handles a single topic of the opinion and has its own examples, decision templates and instructions for edge cases. Splitting the work kept topics apart and made each part easier to test.
- Per-topic queries. Instead of one long prompt, the system sends focused questions to each assistant, which keeps the context relevant and the output predictable.
- Automatic recovery. When an assistant got stuck, for example by refusing a correction, the system reset the session through the API and continued.
- Structured output in the firm's format. Results follow the firm's templates, with key conclusions in bold so attorneys can check and complete them quickly.
- A simple internal tool. A Vue.js interface and a FastAPI backend with LangChain, packaged with Docker Compose, so the firm's staff use it like any other internal application.
The team was small: a frontend engineer, a backend engineer and an AI engineer. Prompts were refined against sample opinions approved by attorneys, and model costs were tracked from the first production run.
Results
Structured opinions are now generated in minutes instead of after hours of linear reading. Model usage costs about $3–4 per case bundle, and attorney throughput improved about 1.5 times, because the AI handles the first-pass synthesis and attorneys focus on judgment. Read the legal document processing case study.
The other half of a law firm's AI stack: the website and content platform
Legal AI is usually discussed in terms of casework, but many firms also compete for clients online, and that side benefits from custom engineering too. The same firm runs several law firm websites on its own content management system, Verba.
- Permissions that match how the firm works. Fixed user roles could not scale across many sites and content types. .wrk replaced them with a policy-based model of resources and actions, with revocable access tokens.
- Faster pages under load. Bots hitting uncached pages pushed server load to its limit. Caching shared page fragments made pages about three times faster under load testing and cut server CPU load.
- One design system for every site. Shared interface components moved into a versioned package, so new firm sites launch from existing building blocks.
- Content generation grounded in real sources. A content service in Go combines keyword and semantic search over indexed material with a trend analyzer that scores topics. The firm's content generation tool draws on both, which keeps drafts tied to real sources instead of invented references.
Read the Verba CMS case study.
What custom legal AI costs
Costs depend on the number of document types, the integrations and the accuracy the firm requires. These ranges reflect typical projects of this kind.
| Item | Typical range | What drives it |
|---|---|---|
| Proof of concept on real case files | $20,000–$40,000 over 4–6 weeks | Number of document types, access to sample cases and attorney time for review |
| Production case-preparation tool | $60,000–$150,000 over 2–4 months | Integrations with case management and document systems, number of output templates, review interface |
| Model usage | A few dollars per case | Bundle size, number of passes and the model chosen |
| Hosting and monitoring | $300–$1,500 a month | Cloud or private hosting, logging and backup requirements |
| Maintenance and improvement | 15–20% of the build cost a year | Template changes, new case types, model upgrades |
Volume decides the economics. A firm that prepares 400 cases a month at about $4 of model usage per case spends roughly $1,600 a month on the model itself. Against a per-case platform fee, a high-volume practice can reach the break-even point within the first year, while low-volume firms are usually better off buying.
Our guide to AI development cost explains what moves a project toward either end of these ranges.
Ethics and risk checklist
Opinion 512 applies to bought and built tools alike. A custom build gives the firm more control over each point, but also makes the firm responsible for designing it in.
- Confidentiality. Use model providers and hosting that do not train on client data, sign a data processing agreement and limit access to case files by role. For the most sensitive matters, consider a model hosted in the firm's own environment.
- Competence and verification. Attorneys must check AI output before relying on it. In Mata v. Avianca, a federal court in New York fined two lawyers and their firm $5,000 in 2023 for filing a brief with case citations invented by ChatGPT. Design the tool so every statement links back to its source page.
- Supervision. Treat AI output like the work of a junior colleague: reviewed, corrected and signed off by a responsible attorney.
- Client communication. Decide when clients should be told that AI was used, and when their consent is needed, particularly if their information is entered into a tool outside the firm's control.
- Fees. Bill for time spent using the tool and reviewing results, not for time spent learning it. Firms on flat or contingency fees keep the efficiency gain, which is one reason personal injury firms were early adopters.
- Records. Keep logs of what the system produced and what attorneys changed, both for quality control and for any later dispute.
A 90-day path for a mid-size firm
Days 1–30: map the work
List the five most time-consuming document tasks in one practice area. For each, record monthly volume, average time per case, the template used and whether a vendor product already covers it. Agree on an AI use policy, since most firms still lack one.
Days 31–60: buy where the market is ready, test where it is not
Pilot a vendor product for the standard tasks. For the one task where the firm's method is unique, run a proof of concept on 30–50 real, anonymized case files and compare the output with attorney-written versions.
Days 61–90: decide and plan
Compare the proof of concept with the vendor option on accuracy, attorney review time, cost per case and data handling. If building wins, plan a production version with the review interface, integrations and logging the firm will need.
Frequently asked questions
What AI software do law firms use?
Law firms use legal research and drafting platforms, practice-specific tools such as personal injury demand and medical chronology platforms, AI features inside practice management software and general-purpose assistants. Some firms also build custom tools for their core workflows.
Should a law firm build or buy AI software?
Buy for tasks that are similar at every firm, such as legal research or standard contract review. Build when output must follow the firm's own templates, when documents are messy and practice-specific, or when high volume makes per-case pricing expensive.
How much does custom AI software for a law firm cost?
A proof of concept on real case files usually costs $20,000–$40,000. A production case-preparation tool typically costs $60,000–$150,000, with model usage of a few dollars per case and ongoing hosting and maintenance.
Is it ethical for lawyers to use AI?
Yes, within professional rules. ABA Formal Opinion 512 requires lawyers to understand the tools, protect client confidentiality, communicate with clients about AI use where relevant, verify output and charge reasonable fees.
Can AI prepare personal injury case files?
Yes. AI can summarize medical records, build chronologies and draft opinions or demand letters from case bundles. Attorneys still need to review and complete the output, but the first-pass reading and structuring can move from hours to minutes.
To sum up
Law firms get the most from AI when they buy for standard work and build only where their own documents, templates and judgment create the value. A custom tool pays off in high-volume, practice-specific workflows, provided confidentiality, verification and supervision are designed in from the start.
.wrk builds AI case-preparation tools and content platforms for US law firms, from document processing pipelines to multi-site CMS and grounded content generation.
Planning AI for your firm's casework or content? Share the workflow and a sample of anonymized documents, and we will propose an approach and a budget range.
See legal tech projects