Stories · Legal / IP
From two hours to ten minutes : and 85%+ accuracy on the work that decides USPTO outcomes.
SKGF's trademark classification process was largely manual: attorneys spending one to two hours per application across 45 Nice Classes, against USPTO rejection rates as high as 57%. Mactores designed a production-grade AI-powered trademark classification system on Amazon Bedrock : hybrid retrieval over 76,000 USPTO records, multi-agent orchestration with Claude 3.5 Sonnet, attorney oversight preserved through Bedrock Guardrails. Less than 10 minutes per classification. 85%+ accuracy. Projected $60,000 monthly savings at 100 applications per month, with 5-7x ROI in year one.
Baseline
60-120 minutes per trademark application. Manual classification across 45 Nice Classes. 57% USPTO rejection rate. Client pressure for speed and predictable cost.
Outcome
Less than 10 minutes per classification. 85%+ accuracy. Auditable, attorney-in-the-loop production system.
Projected Savings
$60K/month
5-7x projected ROI in year one
The Challenge
A high-volume practice with no slack and unforgiving USPTO economics.
SKGF's trademark classification process was largely manual : attorneys spent one to two hours per application across the 45 Nice Classes. Volume was growing. USPTO rejection rates were running as high as 57%. Clients were demanding faster service at predictable costs. The classical professional-services squeeze: too much work, too little senior-attorney capacity, and a rejection-rate ceiling that no amount of throughput could break through. SKGF needed a system that could accelerate classifications while preserving confidentiality, compliance, and attorney oversight : and that could scale with their workload without compromising their accuracy standards.
A traditional SI engagement would have proposed a six-to-nine-month build with a parallel compliance workstream consuming the audit-evidence budget. The economics did not work for a law firm running on quarterly billing pressure. The engagement had to fit inside a seven-week assessment that produced a production-ready architecture and an ROI projection a managing partner could sign.
How We Delivered
Seven-week GenAI assessment. Aedeon ran the data and retrieval layers. FDEs designed the attorney experience and the compliance surface.
Mactores conducted a focused seven-week GenAI assessment to understand SKGF's trademark classification workflows, data landscape, and compliance requirements. Historical trademark records, document formats, and attorney review patterns were analyzed to identify where AI could safely augment legal judgment without introducing risk. Based on those findings, the architecture was designed around hybrid retrieval over 76,000 USPTO records using Amazon OpenSearch with vector search and BM25, orchestrated through Amazon Bedrock with Claude 3.5 Sonnet : ensuring high accuracy, traceability, and compliance through built-in guardrails.
Aedeon's Lane
- •Discovery across SKGF's trademark workflows, document formats, and historical classification patterns
- •Hybrid retrieval pipeline: Amazon OpenSearch for vector search + BM25, with cross-encoder re-ranking for precision
- •Multi-agent orchestration on Amazon Bedrock : Claude 3.5 Sonnet for reasoning, Titan Text Embeddings for retrieval
- •Document-ingestion automation: AWS Textract for legal documents stored in Amazon S3, triggering Step Functions and Lambda workflows
- •Caching and serverless data infrastructure: Aurora Serverless v2 and DynamoDB for structured data and low-latency caching
- •Encryption and guardrails: end-to-end AWS KMS encryption and Bedrock Guardrails for AI behavior safety
Forward-Deployed Engineers' Lane
- •Attorney-facing workflow design : where the AI's recommendation sits, where attorney review is required, how confidence is surfaced
- •Compliance and confidentiality architecture : how 76,000 USPTO records and SKGF's own corpus stay separated and auditable
- •Pilot design : accuracy validation, usability testing, adoption playbook
- •Production readiness : security, integration with SKGF's existing systems, attorney training
- •ROI projection signed off by SKGF leadership before pilot commenced
In Production
A production system designed against a real ROI projection : not a science project.
The assessment confirmed that SKGF could reduce classification time from 60-120 minutes to under 10 minutes : a 90%+ time reduction on a high-volume line : while achieving 85%+ accuracy on USPTO-aligned classification. The projected monthly savings at 100 applications per month: $60,000. The projected ROI in year one: 5x to 7x. The architecture is secure, serverless, and event-driven; the attorney-facing workflow preserves judgment where it matters. The system is positioned to scale beyond the assessment volumes as SKGF's classification practice grows.
Compliance & Confidentiality
End-to-end AWS KMS encryption. Bedrock Guardrails enforce AI behavior policies. 76,000-record USPTO corpus stays separated from SKGF's client confidential data. Attorney-in-the-loop preserves judgment where the agent's confidence is low.
“Ninety percent of the classification work is the same pattern repeated across 45 Nice Classes. The agent does that work; the attorneys do the work the pattern can't predict.”
Why This Matters
Two production agentic platforms inside one client : across two practice areas. That is the operating proof.
The two SKGF engagements : patent Office Action automation and trademark classification : are not two pilots. They are two production agentic platforms inside the same firm, across two practice areas, both shipped on agent-native delivery models that fit inside a four-to-seven-week assessment window. Together they are the operating proof that agentic AI in professional services is a deliverable category : not a future-roadmap conversation. Every law firm, accounting firm, or consulting firm running on senior-judgment-hours has the same opportunity SKGF moved on. The constraint is not technology. The constraint is the delivery model. Most firms cannot ship agentic platforms inside the timeline a managing partner will fund. Mactores can : and SKGF is the proof.