The Road to Useful Autonomy

For years, Tesla has seemed on the verge of bringing full self-driving capabilities to its cars. The promise is that one day, without new equipment, every Tesla will gain the ability to drive itself. To Tesla’s credit, it’s come close. But its Full Self-Driving (Supervised) system still requires a human in the driver’s seat with their eyes on the road, ready to intervene at a moment’s notice. As a result, despite the undoubtedly impressive technology, the real-world utility of Tesla’s self-driving system remains limited.

Now consider robot vacuums. They’re quite useful even though they frequently get stuck in corners or miss spots. That's because the reliability required for useful autonomous systems depends on the stakes of the tasks being automated.

For low-stakes tasks, a system can be useful well before it’s highly reliable. But for high-stakes tasks, even a relatively reliable system may offer little practical benefit until it crosses the threshold at which users can delegate a task and trust the result. At that point, even a slight increase in reliability can yield massive gains in utility.

At Walton, we’re building systems that cross that threshold for complex legal transactions.

The Reliability Gap in Legal AI

Generally speaking, legal AI is still most useful when (a) the stakes are relatively low, or (b) a supervising attorney can catch and correct any mistakes.

The challenge is that, for high-stakes legal work, the required oversight can take more time than completing the work manually.

Consider an M&A transaction. A hypothetical AI agent processes comments from opposing counsel and walks the supervising attorney through how each one affects the transaction. The attorney accepts or rejects each proposed change, and the agent drafts a revised agreement reflecting those decisions.

Reliability is essential at every step. The agent must accurately explain the implications of the comments, capture the attorney’s decisions, and implement them in the new draft. If the attorney needs to independently repeat the AI’s analysis and check every edit, the promised efficiency gains shrink. Across an entire transaction, that burden compounds.

Building Useful Transaction Systems

Walton’s autonomous transaction systems are designed to bring useful autonomy to complex corporate transactions. To do so, we’re taking the following approach:

1. Specialization

We build systems for specific transaction types. Debt financings and M&A transactions share common elements, but each has distinct terms, parties, documents, and closing requirements. Specialization lets us address those differences in depth and provides a concrete standard against which we can evaluate our systems’ outputs.

2. Modeling transactions as structured data

We use representative precedent documents to build a comprehensive terms library and a set of master templates. The terms capture the variables that shape a transaction’s documents, while the templates supply the approved language for each combination of terms. This structure makes transactions easier for agents to understand and lets us handle predictable parts of the workflow with deterministic code.

3. End-to-end transaction infrastructure

Once a transaction system is set up, Walton provides the infrastructure agents need to carry out the work. Its five core components—the terms map, document generation engine, data room, closing checklist, and signature manager—were built from the ground up to work together across transaction workflows.

An attorney can ask Walton to execute a venture financing using an agreed-upon term sheet. Walton will populate the terms map, review data room materials, and draft the financing documents, disclosure schedules, ancillaries, and consents. It will review opposing counsel’s comments with the supervising attorney and draft the next set of transaction documents based on which changes the attorney accepts or rejects. It will prepare signature packets, send approved documents for signature via Docusign, and track signing progress through closing. As work is completed, Walton will update the closing checklist.

4. Rigorous evaluation

We’re building dedicated evaluation suites for each supported transaction type to measure Walton’s performance across standard and nonstandard transaction structures. Our venture financing evals cover straightforward equity rounds, as well as more complex scenarios involving note conversions, secondary share sales, or multiple closings. By prioritizing depth within individual transaction types, we’re able to identify even the smallest gaps in our outputs and make targeted improvements.

A Lesson from Waymo

No self-driving car analogy is complete without mentioning Waymo. It took a very different approach than Tesla and now operates thousands of fully autonomous vehicles across multiple U.S. cities. It operates only in defined service areas, and models local road and geographic features in detail before launching. Its vehicles are equipped with a robust set of sensors, including lidar, imaging radar, and external microphone arrays. It continuously evaluates its systems using simulation, closed-course testing, and real-world driving.

At Walton, we’re running a similar playbook. Each autonomous transaction system is scoped by transaction type, built around a structured transaction model, equipped with purpose-built tooling, and rigorously evaluated. Our goal is to build systems so reliable that attorneys can trust Walton with their most consequential work. That’s useful autonomy.

Explore our venture financing transaction system, now available in early access, or learn more about custom transaction systems.