In Yesterday’s Legal Innovation Club session, Julia Klingberg, co-founder of Kindred, was joined by Aagam Bakliwal, Cornell Tech graduate and co-founder of Kindred, for a discussion on trustworthy AI, explainability, verification systems, and the future of AI in regulated industries.

Klingberg opened the conversation by introducing Bakliwal’s background in AI systems and startup building, highlighting his previous startup exit and Kindred’s recent startup award win. Bakliwal reflected on his journey between India and the US during the rise of the AI boom, explaining how this shaped his belief that “trust is the new compute”. While AI adoption is accelerating rapidly, he argued that organisations increasingly need defensible and explainable AI systems rather than simply fast outputs.

Defining Trustworthy AI

Bakliwal explained that trustworthy AI means systems that are explainable, auditable, and defensible. In regulated industries such as law, finance, and healthcare, hallucinations and inaccuracies can create significant liability. Unlike low-risk AI use cases such as email drafting, legal and regulatory workflows require systems where every decision can be traced and verified.

The discussion then turned to the differences between building in legal tech versus other sectors. Drawing comparisons with his previous recruitment startup, Bakliwal noted that speed and experimentation matter less in legal environments than accuracy and accountability. In legal tech, trust matters more than growth velocity, because every decision must be explainable and defensible from the outset.

From Black Box to Glass Box

Klingberg asked how lawyers and firms can determine whether an AI product is genuinely trustworthy, given that most vendors market themselves as such. Bakliwal described this as both a technical and organisational challenge. He explained that trustworthy AI starts with reducing uncertainty and moving systems away from “black box” behaviour towards “glass box” transparency. Organisations should not blindly trust AI outputs; instead, systems should provide citations, explain reasoning processes, and make decisions auditable.

Bakliwal elaborated on the “black box” problem in AI, where models produce answers without clearly explaining how they arrived at them. This creates serious challenges in legal and regulated industries where accountability is essential. To address this, he stressed the importance of retrieval and verification systems that connect AI models to trusted external sources such as regulations, case law, and internal policies.

The Three Core AI System Approaches

The session explored three common AI system approaches currently used in enterprise environments:

• Pure LLM systems, which are flexible and fast but prone to hallucinations.
• Retrieval-Augmented Generation (RAG) systems, which connect AI to external documents and sources to improve grounding and citations.
• Deterministic or reasoning-based systems, which introduce structured workflows and verification layers to improve explainability and auditability.

Bakliwal argued that the future of legal AI lies not in replacing LLMs, but in combining them intelligently with RAG and deterministic reasoning systems. He used Kindred’s own workflow as an example, explaining how the platform retrieves relevant case law using RAG before applying deterministic verification systems to validate compliance and documentation requirements.

The Rise of Agentic AI

The conversation also explored the rise of agentic AI systems. Bakliwal described agentic systems as AI capable of taking actions autonomously rather than simply answering questions. While these systems could automate large portions of legal workflows such as research, drafting, and verification, he stressed that human oversight would remain essential. In his view, purely autonomous legal AI systems are unlikely to be acceptable because accountability ultimately requires human responsibility.

Ethical AI and Accountability

Klingberg then shifted the discussion towards ethical AI. Bakliwal argued that ethical AI begins with accountable AI. If nobody can understand or verify how a decision was reached, responsibility cannot be properly assigned. He suggested that accountability in AI systems should be shared across the stack: vendors are responsible for reliable systems, firms are responsible for governance and oversight, and lawyers remain responsible for final decisions.

As the discussion broadened to the legal tech industry more generally, Bakliwal emphasised that the sector must move from persuasive AI towards verifiable AI. He argued that explainability, source verification, traceability, auditability, and human oversight must be embedded into products from day one. In his view, “the technology is not the product — trust is the product”.

Industry Adoption and Verification

Audience members then joined the conversation to discuss adoption challenges, verification workflows, and emerging business models. Ron questioned how distribution differs in legal tech compared to other sectors. Bakliwal explained that adoption in legal and regulated industries is heavily trust-driven rather than growth-driven. Credibility, partnerships, and successful pilots play a much greater role than rapid scaling tactics.

The discussion also explored whether a new AI verification industry could emerge. Bakliwal predicted that verification and reasoning layers will become part of every AI vertical, representing the next major evolution beyond content generation. He suggested that independent AI verification companies may emerge in a similar way to external auditors today, though AI-assisted verification will likely remain integrated into broader workflows.

Building Verification Infrastructure

Klingberg and Bakliwal also shared examples of Kindred’s verification infrastructure, including a reasoning engine capable of identifying relevant regulations, validating compliance requirements, and providing source traceability with cited reasoning. They described how legal experts were used to annotate and break down complex medical device regulations, creating proprietary datasets capable of improving auditability and compliance verification.

Conclusion

The session concluded with reflections on how trust, rather than efficiency alone, will determine success in legal AI adoption. Both speakers agreed that while AI can significantly improve legal workflows, verification, accountability, and human oversight will remain central to the future of trustworthy AI.