Two of the most clear-eyed observers of AI adoption, one inside legal and one outside it, have arrived at the same conclusion from different directions. Casey Flaherty, surveying actual generative AI deployments across the legal industry, and Ethan Mollick, studying AI adoption across knowledge work broadly, both find that the binding constraint is no longer what the models can do. It is whether organizations can execute.
This is the sixth piece in our series engaging the leading thinkers on AI and legal practice, and it is the most operational one, because execution is where thought leadership either becomes real or becomes a conference slide.
Flaherty: Past the Hype, Into the Hard Part
Casey Flaherty’s LexFusion report on generative AI in legal, built from closed-door sessions with teams that have moved past experimentation into actual deployment, is the most grounded account of legal AI adoption we collected all year. His consistent finding: the differentiator between firms getting value from AI and firms collecting subscriptions is execution discipline. Workflow redesign, measurement, data readiness, and governance matter more than which tool you pick.
Flaherty also punctures the industry’s self-assessment. His commentary on a Bloomberg Law survey highlights a striking gap: nearly three quarters of surveyed firms believe they are ahead of most clients on AI. They cannot all be right, and sophisticated clients, who can increasingly use AI to keep work in-house, are watching which firms demonstrate real productivity gains rather than announcing initiatives.
There is a practical reason execution matters for pricing, too. A firm that genuinely controls its AI-enabled workflows can predict cost, cycle time, and quality, which is exactly what makes fixed fees and subscriptions offerable with confidence. Execution discipline is the bridge between AI adoption and the pricing models clients want.
Mollick: Interfaces, Agents, and the Management Problem
Ethan Mollick, writing at One Useful Thing, adds a dimension most legal commentary misses. Much of what looks like an AI capability gap, he argues, is actually an interface failure: chat interfaces impose cognitive load that erases productivity gains, while tools built into real workflows unlock them. The near-term leap is agents that work inside your existing systems and files rather than in a separate chat window.
His work on agentic AI and the future of knowledge work pushes further: as AI systems complete multi-step work autonomously, the human role shifts toward delegation, supervision, and evaluation, and AI adoption becomes a management challenge rather than an IT project. His most concrete advice pairs experimentation with governance: give agents real delegated work, but treat permissions, what an agent may access, change, send, or spend, as a first-class design decision.
Put Flaherty and Mollick together and a clear picture emerges. Winning with AI looks like management: redesigned workflows, defined supervision, measured results, and deliberate limits. Losing looks like enthusiasm: licenses purchased, demos admired, pilots that never touch production work.
What Execution Looks Like at a 40-Lawyer Firm
Reading this literature, I kept noticing how much of the execution problem is really a governance-of-partnerships problem. Large firm AI initiatives stall in committees because every workflow change touches someone’s compensation, and no one owns the decision. Flaherty’s execution discipline requires something many firms structurally lack: management that can decide.
This is where a firm of our design has a quiet advantage. AEGIS runs on professional management, so AI decisions, which tools we approve, how output gets verified, what workflows change, are operational decisions made by people whose job is operations. Attorneys adopt and give feedback. Nobody has to negotiate a workflow change against a compensation committee.
The result is not that we deploy everything. It is that we can deploy deliberately: pick real use cases, measure whether they work, expand what does, and kill what does not. That sounds unremarkable until you learn, from Flaherty’s reporting, how rare it is.
Questions Clients Should Ask About Execution
Flaherty’s survey gap suggests clients should stop asking firms whether they use AI, because everyone says yes, and start asking how. Which workflows actually changed? How is AI-assisted work verified before it reaches me? What efficiency gains can you demonstrate, and where do they show up in my fees? Who in your firm owns AI decisions, and how fast can they make one?
Firms with real execution have crisp answers. Firms with initiatives have adjectives.
The Bottom Line
The model-capability race will be won by technology companies. The execution race is the one law firms actually get to run, and it is won with management, measurement, and governance rather than enthusiasm. Flaherty and Mollick have described the playbook in public, for free. The only remaining question for any firm is whether its structure lets it follow one.
By Scott Levine, Founder and Managing Partner, AEGIS Law
Strategic Engagement
Consult with our Managing Partner.
Ready to review your enterprise risk or legacy strategy? Schedule a direct consultation with Scott Levine using the link below.




