This keynote presentation grew out of my ten-part Prudent AI series, in which I explore how leaders can make consequential decisions about AI while preserving their organizations’ capacity to learn, adapt, and choose.
The keynote approaches those ideas differently, through three unexpected historical stories and the leadership questions they raise.
In keeping with my longstanding commitment to sharing ideas openly, I’m publishing the complete script. It’s useful on its own, but, of course, it isn’t the whole presentation. The images, the interaction with a live audience, the questions and answers, and the adjustments made for a particular audience and organization are all important parts of the experience.
I hope the script gives readers something to think about, question, adapt, and build upon.
Prudent AI: Accelerate Reversible Learning, Pace Irreversible Commitment
I. Faster Toward What?
Everyone in this room is being told to move faster on AI. Faster adoption. Faster deployment. Faster results. You’ve approved budgets, signed licensing agreements, launched pilots, and listened to a remarkable number of presentations about what AI will make possible. And much of that activity makes sense. I’ve spent more than thirty years working at the intersection of law and technology, and I’ve never seen a technology with quite this combination of promise, speed, and uncertainty.
But let me ask you something. What’s arriving on your desk?
AI can produce a hundred-page analysis in minutes. Polished legal prose. Persuasive arguments. Impressive-looking citations. Something that appears immediately ready to send to a client. And someone still must decide whether it’s right. Someone must notice the missing exception, the changed definition, the argument that sounds perfectly reasonable until you understand the client’s actual situation.
What happens when a firm’s ability to generate plausible work increases dramatically, but its ability to evaluate that work doesn’t keep pace?
You accumulate something I call review debt. It’s the obligation to verify work that has been generated faster than the institution can reliably judge it. Someone must eventually do that work of verification, or someone bears the consequences of not doing it.
Now, I am not suggesting that we should slow down AI adoption. Quite the opposite. But I think we may be asking the wrong question. Everyone wants more speed.
Faster toward what?
That’s where prudence comes in. “Prudence” has a public relations problem. We hear the word and think caution. Slow down. Wait. Don’t take risks. But prudence is not caution. Prudence is the discipline of making consequential decisions under uncertainty. It’s about moving with purpose while maintaining the ability to change course when the evidence changes.
Here’s my six-word rule for prudent AI:
Accelerate reversible learning. Pace irreversible commitment.
That’s not an argument for doing less with AI. It’s an argument for accomplishing more without losing the judgment and flexibility that make progress possible. And to explore what that means for leadership, I want to begin with a woman you certainly know.
Florence Nightingale. The Lady with the Lamp. Except there’s something surprising about what happened after she arrived at the British military hospital at Scutari.
She made things better.
And the death rate went up.
II. Florence Nightingale and the Work We Don’t See
We all know the Florence Nightingale story. Or at least we think we do. The Crimean War. A terrible military hospital. A courageous woman arrives with a group of nurses, brings order to the chaos, and saves lives. It’s an inspiring story, and there are good reasons Nightingale became famous. But the actual history is more complicated. And I think it’s more interesting.
Nightingale and her nurses arrived at Scutari, near Constantinople, in November 1854. The hospital was overwhelmed by sick and wounded British soldiers. Supplies were inadequate. The wards were crowded. Nursing care was inconsistent. There were serious problems with sanitation and ventilation. Nightingale went to work. She organized supplies, improved basic care, attended to patients, and established routines. She and her nurses made a genuine difference to the comfort and care of the men.
But the mortality crisis continued. By February 1855, the death rate associated with the hospital had reached roughly 40 percent in one commonly cited historical measure.
Now, that doesn’t mean Nightingale’s improvements caused the rising mortality. They didn’t. The hospital was dealing with a complicated crisis involving disease, overcrowding, the condition of patients arriving from the Crimea, and serious sanitary defects. In fact, Nightingale herself initially misunderstood important parts of the problem. She emphasized matters such as nutrition while the hospital building had defective drainage and poor ventilation. A sanitary commission later addressed problems that nursing improvements alone could not solve, and mortality subsequently fell.
The point isn’t that Nightingale failed. Nor is it that someone else arrived with one perfect answer. It’s that improving one part of a system is not necessarily the same as improving the system as a whole. Different people possessed different pieces of the knowledge required. Nursing, medicine, sanitation, administration, statistics. And no one initially had a complete understanding of how the pieces fit together.
Think about a firm introducing AI into document production. Lawyers generate first drafts much faster. Adoption rates rise. People are pleased with the results. Those improvements may be real. But have we reduced the total time required to deliver reliable work? Have we made mistakes easier or harder to catch? Are senior lawyers receiving more work to review than they can reasonably assess? Are clients receiving something better, or simply receiving it faster? A firm can become faster at producing work while quietly accumulating obligations that its traditional measurements don’t capture.
But there’s something else about Nightingale that I think matters for leadership. She rolled up her sleeves. She didn’t just read reports about the hospital. She spent time where the work was happening, and she learned from the people doing it.
I’m not suggesting that managing partners should personally review every AI-generated contract. But when was the last time you sat with the people actually using these systems and watched what happens from beginning to end? Not a vendor demonstration. Not a dashboard. The real work, including the exceptions, the corrections, and the things people have learned to work around. You may discover that the people you thought were simply performing tasks are the people who understand your delivery system best.
Nightingale certainly discovered something along those lines. In January 1855, she wrote to Lord Raglan, the British commander, about a problem with the hospital orderlies. These were soldiers assigned to help care for patients and maintain the wards. They were part of the everyday operation of the hospital. And just as the good ones became experienced, the army kept sending them away to other duties.
Nightingale objected. She argued that the comfort of the sick depended not only on the skill of the surgeon but on the care of the orderly. Constantly changing the orderlies undermined the surgeon’s instructions. She was particularly concerned about the sergeants supervising the wards. They were being recalled just as they began to learn their duties. Nightingale warned Raglan that hundreds of lives might depend on correcting the problem.
I find this remarkable. The army could fill the position.
It couldn’t instantly replace the experience.
Experienced orderlies knew the ward, the patients, the routines, the recurring problems, and the practical details of getting things done. They weren’t simply performing a collection of tasks. They were helping maintain a complicated operating system, and learning more about that system as they worked.
Now think about a law firm. The experienced associate who knows how the different parts of a transaction fit together. The paralegal who knows the filing requirements that don’t appear in the standard checklist. The professional staff member who knows where the data came from and which exceptions require special handling.
Partners may possess broader legal judgment and responsibility for the client relationship. But they don’t necessarily know every operational detail of how the work gets done.
In many firms, the people doing the work every day understand the delivery system better than anyone else. They know the handoffs, the exceptions, the recurring problems, and the small adjustments that keep things moving. That knowledge isn’t incidental to the firm’s capability. It’s part of what makes the firm capable of delivering the work at all.
And those may be the very people whose work you’re now considering replacing with AI.
If the people who understand and maintain your delivery system are the people whose work you automate, what will maintain that understanding afterward?
Suppose an AI system can perform a task in five minutes that previously occupied a junior associate for five hours. That may be an excellent opportunity. Some of those five hours may have been spent doing tedious work that nobody needs to do manually anymore. But before we eliminate the task, we ought to understand what else was happening during those hours. Was someone learning to recognize inconsistencies? Discovering the relationships among different provisions? Developing an understanding of what clients actually care about?
That work may have been doing three things at once. It produced something the client needed. It helped maintain the firm’s ability to deliver the work. And it gave someone experience that might become professional judgment. If AI makes the first of those much cheaper, that’s valuable. But it doesn’t tell us what happens to the other two.
Consider an AI system revising a credit agreement. The language becomes clearer, more consistent, and easier to read. Nothing looks obviously wrong. But one negotiated provision has become slightly more generic. That provision affects a definition in another section, a cross-reference somewhere else, and ultimately the allocation of risk between the parties.
Who’s going to notice? Someone who understands the agreement as a system. Someone who knows why the provision was negotiated that way in the first place.
I don’t think the answer is to preserve five hours of tedious work just because that’s how lawyers have traditionally learned. AI may give us an opportunity to develop young professionals much better than we did before. We may be able to expose them to more sophisticated problems sooner, provide better feedback, and allow them to explore alternatives that would once have been too expensive to investigate. But that’s something we need to design deliberately. If we simply eliminate the work without understanding the capabilities it supports, we may lose something essential in the name of efficiency.
Nightingale had an interesting way of thinking about what good leadership required. In her book Notes on Nursing, she devoted a section to something she called petty management. I like that phrase. Today we might instinctively call something like that micromanagement, but Nightingale meant something rather different. Her point was that being in charge didn’t mean doing everything yourself. It meant arranging matters so that what needed to be done would still be done when you weren’t there. You must understand the work well enough to make that possible.
There’s a tension here that matters for leaders. You need to roll up your sleeves and see what’s actually happening. But you don’t want to become the one person who must understand, approve, or perform everything. The goal is an institution in which people develop expertise, knowledge is shared, and essential work gets done reliably without depending on one indispensable individual.
Nightingale leaves us with a question:
Do you understand where your institution’s capability actually resides, and what happens to that capability when you change the way the work gets done?
But I don’t want to spend the whole morning talking about what institutions might lose. Because there is another side to this story. What can people accomplish when they connect the capabilities they already possess? And that brings us to a part of history I suspect most people in this room know very little about.
III. Who Actually Ran the Silk Road?
Who actually ran the Silk Road?
China? Persia? Rome? The great empires certainly mattered. They controlled territory, collected taxes, fought wars, and influenced which trading routes were open. But there’s another answer.
The Sogdians.
The Sogdians were an Iranian-speaking people from Central Asia, with important centers in places such as Samarkand and Bukhara. They didn’t rule a vast empire stretching across the Silk Road. They didn’t control every route, and the Silk Road itself wasn’t one highway connecting China to Europe. But for centuries, especially from the sixth through the eighth centuries, Sogdian merchants played an extraordinary role in commerce across Central Asia and into China.
Think about what that involved. Merchants conducted business across enormous distances, among people who spoke different languages, followed different customs, used different currencies, and lived under different rulers. Communications were slow. Political conditions could change suddenly. A merchant might depend on someone thousands of miles away whom he couldn’t supervise and might not hear from for months. And yet a substantial commercial network developed.
It wasn’t run by one headquarters. There was no Sogdian chief executive who knew everything happening across Central Asia. Commerce depended on merchants, agents, families, intermediaries, and local communities. People brought knowledge of particular places and relationships that others could use.
And some of the evidence that survives is wonderfully specific. In the early twentieth century, the explorer Aurel Stein discovered a group of Sogdian letters in the remains of an ancient watchtower near Dunhuang, in western China. The letters date to around the year 313.
Some described business difficulties and political upheaval. Others were personal. One woman wrote to a husband who had left her far from home. She described the difficulty of finding help and the precarious position in which she and her daughter found themselves.
The letters were never delivered. They remained in that watchtower for roughly sixteen centuries.
I find those letters extraordinary. A trading network wasn’t an abstract diagram. It consisted of actual people, living in particular places, depending on one another, sometimes succeeding and sometimes suffering terribly when the relationships failed. But across generations, those relationships helped sustain a remarkable commercial world.
And Sogdian merchants weren’t merely passive participants in whatever system the great powers happened to provide. In the sixth century, for example, Sogdian merchants wanted better access to western markets for silk. The Persian Empire stood in the way of some of their commercial ambitions. A Sogdian merchant and diplomat named Maniakh helped persuade a Turkic ruler to pursue a diplomatic opening to the Byzantine Empire. An embassy reached Constantinople in 568. The results were complicated, as diplomatic results usually are. But the initiative itself is striking. Merchants without an empire of their own used commercial knowledge, political relationships, and initiative to help open another possibility.
That’s a rather different image of power from the one we usually get when we study the Silk Road. It’s also a useful way to think about professional institutions.
What have you built?
If you’re a managing partner, you’ve spent years helping build something much more interesting than a collection of individual lawyers. Your firm is a network of people who know things, people who know other people, people who have earned trust, people who can bring together forms of expertise that no one individual possesses. A client calls one lawyer and gains access to knowledge and relationships distributed across the institution. That ability took years to create. In some firms, generations. And it almost certainly isn’t finished.
There are areas of expertise that remain difficult to find. Knowledge that exists somewhere but isn’t readily available to the people who need it. Relationships that could be stronger. Services you’ve considered offering but couldn’t previously deliver at a cost that made sense. Now think about AI in that context. What if people throughout the firm could find and use relevant expertise more easily, or explore a client’s problem from several perspectives before deciding which approach made sense? What kinds of services could become possible if small teams could bring together sophisticated capabilities that once required assembling much larger groups?
That’s the part of AI that excites me. Not simply doing today’s work faster, although that can be valuable.
Making the institution more capable.
And this is where I want to say something to leaders who may feel they’re constantly being told that everything they’ve built is about to become obsolete.
You aren’t starting from zero.
You’ve already built relationships, expertise, trust, and ways of working together. Those are substantial assets. AI may give you new ways to connect and extend them. The question isn’t whether the old institution survives unchanged. It almost certainly won’t. The more interesting question is what the institution you’ve built might now become capable of doing.
That requires initiative throughout the firm. I doubt the executive committee will discover every valuable use of AI. Some of the best ideas may come from the people closest to client problems and everyday work. Nightingale reminds us that leaders need to understand what’s happening on the ground. But understanding the work doesn’t mean directing every experiment or approving every useful idea. One of the great opportunities of AI is to make more people capable of contributing ideas, solving problems, and extending what the institution can do.
That doesn’t mean surrendering control of the institution, either. And I think this is particularly important in our relationships with AI vendors. I have no objection to relying on external technology. Professional firms have always depended on outside capabilities, and AI providers can create enormous value. There’s no reason to assume a law firm must build or own every technological component it uses.
But your firm has spent years building its relationships, knowledge, professional standards, and reputation. You should be very careful about allowing an outside provider to become the only practical means of using those assets.
You can delegate execution without surrendering the authority to define and judge the work.
An AI system may demonstrate remarkable capability. We should recognize that capability and use it appropriately. But never confuse demonstrated capability with warranted authority. A system that produces an excellent legal analysis doesn’t thereby acquire the authority to decide what risks a client should accept or what professional obligations the firm should undertake. Those judgments remain with people who are accountable for the consequences. The challenge is to encourage initiative and make good use of outside capabilities while retaining the institution’s ability to set direction.
There’s an idea from finance that helps explain how to do that. It’s called Real Options. A real option is the ability to make a consequential choice later, after you’ve learned more, because you were careful not to give that choice away today. You make a manageable commitment now. You learn something. You retain the opportunity to do more if the evidence supports it or change direction if it doesn’t.
That’s why I keep returning to those six words:
Accelerate reversible learning. Pace irreversible commitment.
You want to encourage experimentation without converting every experiment into an obligation to continue forever. And you want to be ready to make a bigger commitment when the evidence points toward a genuine opportunity. This is how I think about stewardship.
Leadership is the stewardship of collective capability, not the performance of individual brilliance.
It’s about helping people do more together than they could do separately, creating room for initiative, and keeping the institution capable of deciding where it wants to go. You have a remarkable network already.
What could it become?
Now, there’s another side to preserving choices about the future. It’s one thing to create new possibilities. It’s another to discover, too late, that you’ve lost the ability to choose among them. And that brings me to St. Louis.
IV. St. Louis and the Time It Takes to Recover
I lived in St. Louis for thirty-five years. And there’s a story from that city that has a particular meaning for my family. On July 12, 1973, shortly after midnight, a fire broke out at the National Personnel Records Center in the St. Louis area.
The building held millions of military personnel records. These weren’t just old government files. They documented people’s military service, assignments, medical histories, and other facts that might become important many years later. A veteran might need to establish eligibility for medical care or benefits. A family might need proof of a relative’s service.
The fire burned out of control for twenty-two hours. It wasn’t officially declared extinguished until July 16, four days later. An estimated sixteen to eighteen million military personnel files were destroyed or damaged.
My father-in-law’s records were among those lost.
That makes the story personal for me. But think about what the loss meant for the people who needed those records. Someone could have served the country decades earlier and then, when the evidence of that service mattered, discovered that the original documentation was gone.
And here’s an important detail. The government had not maintained complete duplicate sets of those personnel files. There were no microfilm copies waiting in another location. Millions of files had accumulated in one institution, with a catastrophic exposure they largely shared.
Having many assets isn’t the same as having independent assets. A thousand separate records can all be lost to the same fire. And several different AI products can share underlying models, infrastructure, data sources, or other dependencies that aren’t obvious from their different names. Diversity matters only to the extent that the alternatives are genuinely independent in the ways that count.
But I think the most remarkable part of the St. Louis story comes after the fire. The records were gone, but the obligation to serve veterans remained. People still needed to establish their service. So, the National Personnel Records Center began an enormous recovery and reconstruction effort.
Some records had survived in terrible condition. Approximately 6.5 million burned or water-damaged personnel files were recovered. Imagine the work involved. Documents charred by fire, soaked by water, sometimes covered in mold, handled and preserved by people trying to rescue whatever information remained.
But the recovery effort went further. Records specialists began searching for evidence elsewhere: in payroll records, military unit records, medical information, and documents maintained by other government offices. Much of this material had been created for purposes unrelated to backing up the personnel files. Yet it became invaluable when the original evidence disappeared.
Suppose someone needed to establish military service, and the original file no longer existed. A specialist might have to locate different sources, determine which facts they supported, compare them, and assemble enough reliable evidence to establish what was needed. That’s not simply retrieving a document. That’s reconstruction.
In 2023, the National Archives reported that its reconstruction specialists had successfully reconstructed details of military service approximately 5.5 million times since the fire. But it wasn’t a complete restoration. Some information was lost permanently. Some records could be reconstructed only in part. And fifty years after the fire, the work was still continuing.
The fire lasted days.
The consequences lasted generations.
I have a name for the problem this illustrates. Reconstitution Time.
When organizations talk about technology recovery, we tend to measure hours. How quickly can the servers be restored? How long until the data is available? What’s the recovery time in the contract? Those are important questions. But here’s a different one.
How long would it take to restore your institution’s ability to perform essential work reliably?
Suppose you’ve worked with an AI provider for several years. They’ve configured your workflows, programmed your AI agents, connected your systems, and worked through the exceptions and special cases that make everything reliable. It works. Your people rely on it every day. Then the provider disappears.
The technology is still there. You still own your data. You may even have the workflows and agents they built.
But has the knowledge needed to make it all work disappeared with the vendor?
Recovering information is not necessarily the same as recovering the ability to use it.
The St. Louis story contains a hopeful lesson as well as a warning. The reconstruction succeeded as often as it did because some independent evidence survived and because knowledgeable people could work with it. The original system had suffered a catastrophic loss. But not every source of knowledge had disappeared with it.
That matters for how we think about our AI decisions. It’s tempting to believe that we have preserved flexibility if a contract allows us to terminate a service or if another technology is available on the market. But a contractual right to leave doesn’t necessarily mean you can actually leave. If the expertise required to describe the work, verify the output, and operate the process has disappeared from your institution, you may discover that you don’t have a practical alternative. That’s why the choices we preserve need to be real. We need enough independent knowledge, human expertise, and authority to exercise them.
And that’s where all three histories begin to speak to one another. Nightingale saw the consequences of losing experienced people who understood how the work was done. The Sogdians show us the extraordinary value that people can create when expertise and relationships are connected across distances and institutions. St. Louis shows us what it can take to recover reliable knowledge after an essential source has disappeared.
These weren’t people carrying out a common management strategy. Their worlds were very different. But together, they force us to ask an important leadership question:
What capabilities must an institution retain if it is to remain free to choose its own future?
V. Leadership Choice and Exercising the Option
I want to come back to where we started. Everyone is being told to move faster on AI. I think we should move prudently. The possibilities are too important to ignore. But leadership requires deciding what we’re moving toward. It’s about what your institution is becoming capable of doing.
You aren’t starting from zero. AI may let you make the capabilities you’ve already built more powerful than they have ever been. It may allow you to offer services that were previously impractical, develop professionals in better ways, and connect knowledge that has been difficult to access. It may also show you that some of your existing practices no longer make sense. That’s not necessarily a loss. Stewardship doesn’t mean preserving everything exactly as it is.
But stewardship does mean knowing what you’re responsible for preserving. That includes the knowledge of the people who keep the work moving, whether or not their names appear on the organizational chart.
Some things should change. Some things should be replaced. And some things should be stopped. Which brings me to a responsibility of leadership that I don’t think we talk about enough.
You’re the one who has to kill things.
Not the AI vendor. Not the innovation committee. Not the team that put months of work into making the project succeed. You.
Maybe you’ve already spent a substantial amount of money. Maybe the project has a prominent sponsor. Maybe you’ve publicly committed to its success. Maybe it was your idea.
It’s happened to me.
The money already spent doesn’t make the next dollar worth spending. You still must judge what comes next. You must consider what the initiative can realistically accomplish, what continuing will cost, and what stopping will require. And if the evidence says the initiative no longer makes sense, someone must make that decision. That’s what leadership is for.
The option to stop has value only if someone is willing to use it.
And when you do stop something, you shouldn’t have to throw away everything you’ve learned. Preserve the knowledge. Preserve the expertise. Keep the relationships and capabilities that can help you do something better.
And sometimes, you’re the one who must commit more deeply.
Then move.
The point is not to avoid consequential decisions. It’s to make them with an understanding of what you’re building, what you’re risking, and what options you’re preserving.
Accelerate reversible learning. Pace irreversible commitment.
You’ve spent years developing judgment about people, clients, relationships, and difficult decisions. You don’t need to borrow that wisdom from AI. You need to bring it to AI. That’s how the promise of professional judgment is renewed: not by refusing to change, and not by surrendering responsibility to a machine, but through choices your institution can still make. And those choices are yours.
VI. The Landing
Let me leave you with an image. I’m not here to fly your plane. That’s your job. You know your institution. You know your people, your clients, your responsibilities, and where you’re trying to go.
Think of my role today as air traffic control. From the tower, I have a different perspective. I may be able to see conditions, traffic, or an approach that isn’t quite as obvious from the cockpit. I can help you think about the direction you’re heading and what it will take to get there. But I don’t fly the aircraft.
And the pilot isn’t acting alone, either. A safe landing depends on the work of many people—pilots, controllers, ground crews, engineers—whose different capabilities have to come together. That’s the kind of leadership I’ve been talking about today.
Once the wheels hit the runway, my job is done.
Then the responsibility is yours again.
The framework stays.
I leave.
The learning loop stays.
A keynote is never quite the same presentation twice. The audience, the organization, the discussion, and especially the questions shape what happens in the room. That’s an important part of the learning process, and something a published script cannot fully capture.
I’m sharing this script under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are welcome to share and adapt it, including for commercial purposes, with appropriate attribution and an indication of any changes made.
Copyright © 2026 Dennis Kennedy.
Licensed under CC BY 4.0. The license applies to the keynote text, not to third-party images or other materials used in presentations.
[Originally posted on DennisKennedy.Blog (https://www.denniskennedy.com/blog/)]
DennisKennedy.com is the home of the Kennedy Idea Propulsion Laboratory
DennisKennedy.Blog is part of the LexBlog network.
Dennis Kennedy – CC BY 4.0 license