For Kilpatrick Townsend & Stockton partner Charles Gray, the surge in generative artificial intelligence over the last two years has been as much about the future as it is a return to his roots.
Before law school, Gray worked as a software developer. When enterprise access to advanced AI models opened up, he began writing Python scripts to automate repetitive workflows. Around the same time, he connected with Trajan Unger, a developer who has been with Kilpatrick since 2005.
What started as garage-style tinkering soon evolved into an official internal incubator: Kilpatrick Labs.
“We both work at a large, multinational Am Law 100 firm — an organization that isn’t typically known for being nimble,” Gray said. “We realized we needed something that felt like a lab. We think of ourselves as a well-funded startup where the firm is our investor, and our ‘board of directors’ is a group of attorneys from different practice groups interested in developing AI solutions.”
Moving Beyond ‘Bolting It On’
Many law firms have approached AI by adopting general chatbots or off-the-shelf third-party tools. “You can’t just take a workflow you’ve done the same way for a decade, slap AI on it, and say, ‘Now be faster,’” Gray said. “AI is about taking a workflow and infusing it with data to make you smarter. It’s a melding of the attorney and the technology.”
He pointed to patent continuation filings as an example. Traditionally, deciding whether to keep a patent application open involved a standard, time-constrained file review. With specialized AI agents, the firm can analyze the client’s entire patent portfolio, evaluate related technologies, review competitor filings and scan pending litigation in seconds.
“Now I can make decisions that were previously impossible simply because gathering that data took too much time,” Gray said. “It’s about asking: What am I missing, and how can I enhance the output for the client?”
Specialized Initiatives
Kilpatrick Labs currently oversees dozens of active initiatives spanning internal operational tools and collaborative client features, including:
- Pre-litigation & Deposition Prep: Custom agentic workflows assist litigators during hearing preparation. In one trial run with a patent litigator, the tool predicted five potential questions a judge might ask, three of which were raised verbatim during the hearing.
- Patent Ideation & Prosecution: Multi-agent tools triage backlogs of potentially patentable ideas and track filings through the U.S. Patent and Trademark Office.
- Brand Enforcement: Specialized agents continuously monitor online channels for trademark infringement, unauthorized products and suspicious domain activity.
- Operational Triage: Automated workflows streamline routine administrative tasks, including prebill and pro forma reviews.
- Client Collaboration Portals: The firm is building an external portal allowing clients to test in-house tools or request custom-built workflows tailored to specific data needs.
To power these workflows, Gray builds multi-agent frameworks that operate in adversarial loops. “Think of it like a partner managing five associates and paralegals,” he said. “A leader agent assigns tasks to sub-agents, they cross-examine and check each other’s outputs and they synthesize the best result. It mirrors standard software development architecture.”
Security, Privilege and Measured Deployment
While the lab operates as a sandbox where developers can experiment freely, firmwide deployment follows strict protocols. Every tool undergoes architecture specification reviews with the firm’s information security team and general counsel’s office before entering small-team pilot phases.
Gray emphasized that keeping data safe requires a clear distinction between consumer-grade and enterprise-grade infrastructure. “Consumer-grade tools train on your data with no guarantee of privacy. If you use those for confidential work, that’s dangerous — full stop,” he said. “Enterprise environments isolate your data in a closed, secure cloud where models never train on your inputs.”
The lab also takes a cautious stance toward emerging tools like Model Context Protocol servers, which allow AI models to interact with third-party software. “We only use MCP servers built internally and vetted through security, or those provided directly by verified vendors,” he noted. “When you connect read-and-write permissions into legal data sources, you must be deliberate.”
The Shift Toward Real-Time Judgment
Looking ahead, Gray expects legal work to move toward agent-to-agent communication, where client systems and firm systems handle preliminary research and data exchange directly. This shifts the lawyer’s day-to-day role away from document drafting and toward high-velocity decision-making.
“Historically, lawyers might make a handful of critical judgment calls a week because producing the underlying work took so long,” Gray said. “Now, attorneys might make 20 or 30 strategic calls a day based on curated data.”
For firms, he said, adapting is no longer optional. “If you deliver the exact same contract or patent you provided five years ago, just slightly faster, that won’t sustain you when opposing counsel, courts and patent offices are all leveraging AI,” Gray said. “The value isn’t just speed; it’s delivering an order of magnitude more strategic depth.”