Start with a measurable job
An effective project begins with a specific workflow and a baseline. In a legal contract-review implementation, I helped design, train, test, and deploy agents over roughly three weeks. Analysis and initial markup that had often taken one to two hours moved into a roughly 10-to-20-minute workflow. The result was meaningful because it addressed an actual recurring task, although each matter still required professional judgment and review.
Reliability is the real engineering challenge
Generating a plausible answer is not enough for a professional workflow. The harder objective is consistent behavior across different documents, unusual terms, incomplete information, and edge cases. We spent substantial effort tuning agents to be more deterministic. The right standard depends on the stakes, but variability should be understood and measured, not wished away.
Context determines quality
Generic model capability is only part of the system. Outputs improve when the system can use company-specific playbooks, preferred clauses, example agreements, and relevant matter context. A commercial contract does not exist in a vacuum; a negotiating position may depend on other documents, prior correspondence, and the business objective. Context architecture becomes a governance question when people rely on outputs.
Human review is a design feature
Professional oversight should be designed into the workflow. It is important to know where a reviewer must verify a finding, approve a redline, or resolve conflicting sources. The aim is to improve the quality and speed of human work, not to hide accountability behind an automated output.
Test systematically
Looking back, I would invest in automated testing earlier. Human review is critical, but repeatable regression tests can expose unexpected behavior before it reaches users. A mature program should track representative cases, failure modes, drift, feedback, and changes in the underlying models or prompts.
Scale is an organizational choice
The first successful workflow is evidence of potential, not proof that a business has transformed. Scaling requires process redesign, training, adoption, security controls, workflow ownership, and a clear measure of financial or operational benefit. A board should not confuse the number of licenses purchased with the value created.
For the board
For directors, the lesson is simple: evaluate AI using the same principles applied to other important capital and operating decisions. Set an objective, measure outcomes, understand dependencies, assign accountability, and adjust based on evidence. Technology changes quickly. Those disciplines endure.
Related reading: When an Algorithm Reads Your Proxy and AI Is on the Ballot (Harvard Law School Forum on Corporate Governance, Nov. 5, 2026). The analysis above reflects the author’s perspective.