By Bryce Engelland
Agentic AI is emerging as legal technology's next frontier, but its added complexity and unique demands may make it far harder to operationalize than generative AI (GenAI) — and some efforts in legal are destined to fail
Key takeaways:
- GenAI success is not a proxy for agentic AI readiness. GenAI augments individual tasks and forgives messy systems; agentic AI demands clean data, documented workflows, and explicit processes. Those law firms that mistake one kind of maturity for the other will discover the gap painfully.
- Misaligned incentives kills more pilots than bad technology. When no one is rewarded for adoption, but everyone bears risk for visible failure, the rational response is quiet non-compliance. That’s why successful implementation requires restructuring who owns the outcome, not just who owns the project.
- The work that matters most is unglamorous. The firms that succeed with agentic AI will be the ones that spent the preceding years documenting knowledge, identifying institutional workarounds, and fixing information architecture — even though this preparation pays no dividends until it prevents disaster.
It's 2028, and your law firm's agentic AI pilot has officially been shelved. The budget is gone, the champions have moved on, and the postmortem sits in your inbox, unread. What went wrong?
The truth is, almost nothing surprising. The mistakes were predictable — obvious, even, in hindsight. The warning signs were all there in 2025 and 2026, back when the industry was still figuring out basic GenAI deployment, let alone the next stage that agentic AI would come to represent. The warnings went unheeded, however, and now you're explaining to leadership why 18 months of investment produced nothing but a cautionary tale.
That's the future we’re trying to help you avoid.
Right now, in the early days of 2026, agentic AI exists mostly as demos, developer previews, and a few new features worked into pre-existing infrastructures. The technology can perform impressive feats in controlled environments, such as chaining together complex tasks, navigating multistep workflows, and making decisions without constant human input. In the legal realm, however, most law firms are still wrestling with first-generation GenAI adoption. The agentic future still feels quite distant.
It isn't. Two years can pass quickly, and the law firms that spend that time preparing will have options in 2028 that the unprepared won't. The challenge is that preparing doesn't mean what most people think. It doesn't mean buying early, and it doesn't mean waiting for maturity. Rather, preparing means understanding now why these systems fail, and building the institutional capacity to avoid those failures when the technology arrives in full.
Let’s look at two premortems — a kind of autopsy of failures that haven't happened yet, written so you don't have to write your own. The two firms are fictional, but the failure modes are not. Many law firm professionals will recognize themselves and their firms in one of these stories; some even will recognize themselves in both, and all will find lessons to learn.
Premortem No. 1: The eager overreacher
This law firm had every reason to be confident. Composed of 400 lawyers and a revenue trajectory pointing straight at entering the Am Law 200 by the end of the decade, the firm regularly succeeded at outpacing client expectations. More specifically, the firm had just completed a GenAI rollout that managed to come in on time, on budget, and meet all its key objectives. While competitors were still forming committees, the firm had already trained its people, built internal tooling, and moved on. The firm wasn’t reckless — it was ahead of the pack. So, when agentic AI emerged as the next frontier, the firm’s leadership saw a chance to widen the gap.
The firm launched its agentic AI pilot in early 2027 with real momentum. A small team of early adopters — three associates, key technical personnel, and a senior paralegal who'd driven the GenAI rollout — took the lead. The team built a sandbox environment, mapped target workflows, configured the agents, and ran test cases with synthetic data. The results were promising. The AI agents handled routine intake and document routing faster than expected. The firm’s leadership was impressed and scheduled a board presentation. The team prepared to move from sandbox to production.
That's when it fell apart.
The sandbox had been clean; however, the real environment was not. The firm had never actually fixed its information architecture — it had just trained humans to navigate the mess intuitively. Previously, GenAI had been forgiving enough to work with that arrangement. For example, you could paste text into a tool and get useful output regardless of where that text lived. Unfortunately, agentic AI was not as forgiving. The agents expected governed data, documented processes, and clear system integrations. Instead, the AI agents found 13 years of ad hoc folder structures, a contract repository split between the DMS and senior partners' Outlook folders, and workflows that existed in five different configurations based on who one was talking to at the time. Worse yet, none of this configuration matched the documented set-up the agentic AI had been trained on.
The conflicts check was where the AI pilot finally broke.
On paper, the workflow was simple — four steps, fully documented, ready for automation. In practice, it required seven steps — plus a phone call to Linda in accounting, who had been at the firm for 19 years and simply knew which clients had complicated histories, which matters had been walled off informally, which names triggered exceptions that had never been written down.
The agent, of course, didn't know it needed to call Linda — it couldn't know. So, it processed a new matter intake for a client that, technically, cleared conflicts; but, it was one that any human who'd been at the firm more than a year would have flagged immediately. The client in question had been involved in contentious litigation with an existing client's subsidiary four years prior. The formal conflict had been resolved, but the relationship remained sensitive. Partners on both sides had agreed, informally, that the two clients would never be served by the same practice group.
The agent, finding no documented conflict, routed the intake to exactly that practice group.
By the time anyone caught it, an associate had already billed six hours to the new matter, and the partner on the existing client relationship found out about the situation from the client, not from his own firm. Unsurprisingly, the conversation was brief and unpleasant. Eventually, the firm was able to keep the client — barely — but the agentic AI pilot was frozen within a week.
Still, other problems compounded. For example, the human-in-the-loop review process had become theater, with reviewers approving workflows they couldn't parse. Then, two of the original champions left mid-pilot, taking irreplaceable expertise with them. The conflicts failure was just the kill shot.
The failure revealed something the firm’s leadership hadn't wanted to see. The firm's apparent operational maturity was fiction, maintained by institutional knowledge that had never been documented and couldn't be automated quickly. It was a castle on a sandy foundation that couldn’t be held.
Premortem No. 2: The cautious giant
The firm was built to last, with more than 1,000 lawyers, a century of institutional history, and a client roster that included household names across four continents. The firm saw technology hype cycles come and go, and watched smaller firms chase trends and flame out. Firm leadership’s approach to innovation was deliberate, methodical, and — as they would tell you — appropriately skeptical. After all, firm leaders watched the evolving legal technology sphere closely and built a well-oiled machine for testing, evaluating, and implementing emerging tech. They knew some technologies were just marketing hype, but others were serious opportunities, and they intended to sort the wheat from the chaff.
When agentic AI entered the conversation, leadership formed a committee, commissioned a risk assessment, and approved a carefully scoped pilot with extensive guardrails so it never touched a single live matter until it proved itself under even unexpected conditions. Leadership knew that this was how serious firms handled serious technology.
The pilot launched in mid-2027 with all the right infrastructure — a dedicated budget, executive sponsorship, and a cross-functional team with representatives from IT, practice groups, risk, and operations. The pilot also had governance documentation that ran dozens of pages. The AI agents were confined to low-risk, back-office workflows, such as document classification, deadline tracking, and routine client communications — nothing that touched substantive legal work or that could embarrass the firm. The sandbox performed well, and the agentic AI rollout was approved.
That's when nothing happened.
Not a failure, exactly — just friction so pervasive that forward motion became impossible.
The deadline-tracking pilot was supposed to be the easy win with low stakes, clear value, and minimal risk. The agents would monitor matter deadlines across the docketing system, flag upcoming due dates, and send automated reminders to responsible attorneys. The system worked exactly as designed — and the attorneys ignored it completely.
Not out of spite; not even out of skepticism. They ignored it because they had spent years building their own systems — Outlook reminders, spreadsheet trackers, paralegal check-ins — and those systems worked. What’s more, they knew and trusted those systems.
The new system was one more thing to monitor, one more source of alerts competing for attention, and one more place where something could go wrong. The rational response was to keep doing what they were already doing and let the pilot run in the background, officially adopted, functionally unused.
Project leadership tried incentives, training, executive pressure — nothing moved. The attorneys had done the math, even if they'd never articulated it; no one's bonus depended on the pilot succeeding, but someone would absolutely be blamed if an agent missed a deadline, or sent a reminder to the wrong person, or did anything that resulted in an attorney losing face. The asymmetry was obvious — the upside of championing the pilot was diffuse and long-term; while the downside of a visible failure was immediate and career-limiting.
Meanwhile, the practice group leaders who controlled day-to-day operations were measured on billable hours, budget stability, and headcount retention. For them, a successful pilot meant disruption to workflows they'd spent years optimizing and potential threats to teams they'd built. They didn't sabotage the pilot necessarily, but they raised concerns, emphasized edge cases, escalated risks, and ensured that nothing moved faster than the slowest stakeholder. Of course, each objection was reasonable, and each delay was defensible; yet, the cumulative effect was paralysis.
The firm was also fighting its own history. By 2028, staff had lived through multiple transformative technology initiatives that had been announced with fanfare and abandoned within a year. Thus, the agentic AI pilot arrived pre-discredited. People had learned that enthusiasm was risky and skepticism was free. Indeed, the immune response wasn't cynicism — it was pattern recognition.
The firm’s agentic AI pilot didn't die of a single wound. It died of a thousand reasonable hesitations, none of them sufficient to kill the project alone, but together forming an institutional immune response that no technology, however promising, could survive. Within 18 months, the firm had spent its budget, exhausted its champions, and produced nothing but a case study.
The common thread
These two firms failed in their agentic AI visions for opposite reasons — one moved too fast; the other couldn't move at all. However, they both shared a common point, in that their agentic AI failures were never about the technology.
The eager overreacher firm had capable tools and motivated people. It failed because firm leaders mistook a successful GenAI deployment for institutional readiness. GenAI augments individual tasks — a lawyer could paste text into a tool and get useful output regardless of how chaotic the surrounding systems were. Agentic AI, on the other hand, demands something different — clean data, documented processes, explicit workflows, and distributed expertise. The firm’s operational maturity was a performance, maintained by humans who'd learned to navigate dysfunction so well that the dysfunction had become invisible. When that dysfunction became visible, it was only because the agentic pilot had already collided face-first into it.
The cautious giant firm, by contrast, failed because every layer of the institution was individually incentivized to protect itself, and collective self-protection produced collective paralysis. No one was rewarded for the pilot's success; yet, everyone was at risk if it failed visibly. The firm had seen so many failed technology pilots at this point that staff had become numb. Their rational response was to nod in meetings and quietly ensure the technology never touched anything with their names on it.
Both failures were predictable, and both were preventable. Both will happen, unfortunately and repeatedly, over the next three years to those law firms that assume technology is the hard part.
Of course, legal leaders may not be too far wrong in thinking that. Agentic AI represents a genuine leap in complexity, risk, and implementation difficulty — particularly within legal environments in which the stakes of autonomous decision-making are uniquely high. However, these premortems illustrate that the technology's capabilities per se are not the only concern. A system that performs flawlessly within its design parameters can still fail catastrophically when it encounters an organization that isn't ready to receive it. The eager overreacher had the technology, and the cautious giant had the governance. Neither had what mattered — the institutional coherence to turn capability into outcome.
This is not a problem that solves itself with time. The firms that succeed with agentic AI in 2028 will not be the ones who bought in earliest or waited longest. They will be the ones who spent 2026 and 2027 doing the unglamorous work of documenting the undocumented, cataloging the institutional knowledge that lives in Linda's head, aligning incentives so that adoption isn't someone else's risk and not their reward.
Simply put, the firms that succeed with agentic AI in 2028 will be the ones that recognized that a premortem is not pessimism — it's preparation.
