I built a private LLM agent for a boutique law firm that runs on a $5/month VPS. No subscription. No vendor lock-in. Total cost: $0.02 per run.
SaaS companies charge $500 per seat for a wrapper around the Claude API. I wrote the real plumbing instead.
The Fallacy of Legal SaaS
Generic legal AI platforms can't access your active Clio matters in real-time. They store your confidential client data on third-party servers you don't control. True custom AI agent shells give you complete ownership. For the full service picture, see document processing agents.
True human capability multiplication means owning the infrastructure. You run the code. You hold the data. No middleman takes a cut.
The firm I worked with had 6 attorneys and about 40 active matters. They were paying $500 per seat across 9 users. That is $4,500 per month, $54,000 per year, for drafting help that still needed an associate to check every output.
Privilege is the second problem. Once client notes leave your network, the audit story gets complicated. Corporate counsel asks where the data sits. The honest answer matters.
The Custom Agent Architecture
I deploy a local Python script on a secure VPS that queries the Anthropic API directly. Zero data retention. Full control.
import anthropic
import sqlite3
def invoke_legal_agent(case_notes):
client = anthropic.Anthropic()
prompt = f"""
You are a Senior Litigation Paralegal.
Review the following case notes and draft a formal demand letter.
Do not invent facts.
{case_notes}
"""
response = client.messages.create(
model="claude-3-5-sonnet-20241022",
max_tokens=2000,
messages=[{"role": "user", "content": prompt}]
)
return response.content[0].textOne script. $0.02 per execution. You own every line of code. You own every byte of data.
Key insight: The SaaS tax exists because firms don't realize they can write 30 lines of Python and eliminate $500/month in subscription fees.
What the Firm Runs
The demand letter script is one piece. The full agent does four jobs.
- Matter intake. A form drops new facts into the matter file and flags missing documents.
- First draft generation. Demand letters, engagement letters, and routine correspondence built from the firm's own templates.
- Clause review. Incoming contracts get scanned for indemnity, venue, and auto-renewal language. Findings land in a short brief. That pipeline is covered in depth on custom AI agent shells.
- Deadline watch. Dates pulled from matter notes and pushed into the firm calendar every morning at 06:30.
Model spend for the whole month, all four workflows, came to $87. The VPS is $5. Backup storage is $1. Total running cost: under $100 per month.
Privacy and the Data Path
Client data lives in one place: the firm's encrypted volume on the VPS. The Python worker reads matter notes, sends only the text required for that run to the model, and writes the output back to the same volume.
Anthropic API calls go out with zero data retention enabled. Nothing trains on client files. Nothing sits in a third party document store.
When outside counsel asks about the data path, the firm hands them a one page diagram. That answer has closed the question every time.
What Ownership Means in Practice
Ownership is concrete. The firm holds the repo in its own GitHub organization. The templates live in a shared drive the firm controls. If I disappear tomorrow, the workflows keep running and a competent Python developer can read the code in an afternoon.
The Claude API is a dependency, and a healthy one. Model upgrades become a version bump in one file. Nothing about the firm's matter data is stuck inside a vendor's product roadmap.
When the firm added a fourth practice area last spring, we cloned the drafting workflow, swapped the templates, and shipped it in 9 days. With a SaaS seat model, that change would have been a support ticket.
The Build Timeline
Two weeks for the first workflow. That included the template work with the managing partner, which is the real bottleneck. Software is quick. Agreement on how the firm drafts takes longer. A comparable legal build is documented in the Dalessio Law case study.
Week three covered clause review and the deadline watch. Week four was testing against 60 closed matters. We compared agent drafts to what an associate produced. Differences got folded back into the prompts and templates.
For the deeper stack, see why RAG fails for unstructured legal documents and my notes on clause extraction with the Claude API. If you want the tool permission layer, read tool-calling architecture.
Want the full deployment guide? Download the Blueprint or AI Workflow Repair Intake. For more on agent architecture, see business AI agents.
| Operational layer | Fragile default | Production architecture |
|---|---|---|
| Document ingestion | Manual reading and copy-paste into matter fields | Automated OCR and classification with strict JSON schema validation |
| Conflict & risk check | Ad-hoc search after the consult is booked | Pre-booking contact search and mandatory human gate on low confidence |
| Case file retrieval | Naive chunking that splits clauses mid-sentence | Section-aware chunking with citation anchors back to page and line |
| Compliance & custody | Prompts sent to consumer chat accounts | Zero-retention API endpoints with full per-document audit logs |
FAQ
How long does it take to deploy custom ai agents for boutique law firms in a law firm?
A focused legal workflow covering intake, OCR extraction, or matter sync takes two to four weeks from mapping to production handover, with human approval gates on every client-facing step.
Does client data stay private when using AI for legal workflows?
Yes. Production legal pipelines use zero-retention API endpoints, scoped field permissions, and full audit logs so client documents never train public models.
Can this integrate with Clio Manage, MyCase, or Lawmatics?
Yes. The architecture connects directly to Clio Manage, MyCase, Lawmatics, and HubSpot via webhooks and REST APIs with search-before-create deduplication.