Are You in the 99% or the 1%?

The answer may determine your company’s future.

I’ve been looking at Tesla’s Cybercab and wondering why it still looks so much like a car. It has no steering wheel or driving pedals, which is a fairly dramatic departure from more than a hundred years of car development. But it still has two seats facing forward, a hood, a trunk, and an unmistakable front and back. We have removed the driver and kept a surprising amount of the driver’s world.

My next thought was that perhaps it should just be a big box with comfortable seats. Why have a permanent front at all? Once you start pulling at that thread, even four wheels becomes a choice worth examining. Some familiar features will earn their place through physics, road dimensions, or passenger comfort. The interesting exercise is making each feature explain why it is still there. “The previous version had one” is a wonderfully efficient way to end a design discussion before anyone has to do any designing.

Zoox has already pursued some of these questions. Its robotaxi is bidirectional, with a symmetrical cabin and passengers facing one another. It still travels on roads, but the interior can be organized around the people taking the trip. The seating plan no longer has to accommodate someone whose principal social obligation is to stare through the windshield.

The same question makes me look differently at airplane windows. We cut openings in a pressurized fuselage, then add the panes, frames, and reinforcement needed to make them work. Those openings concentrate structural stresses and add weight. And after all that engineering, we pull the shades down. I wonder how much of our attachment to the window comes from a time when seeing outside helped make the extraordinary act of flying feel reassuring. With cameras outside and screens inside, could we keep the view and reconsider the hole?

Emirates introduced virtual windows for the middle First Class suites on its redesigned Boeing 777 in 2017. They show the view outside using live cameras. The passenger gets a window view from a seat nowhere near the side of the aircraft. Giving someone a view no longer requires putting them beside an opening in the fuselage.

Which brings me back to the Cybercab. Why does it need all that glass? Glass can break, and automotive glazing is engineered to manage what happens when it does. The driver who needed to look through the windshield is gone. I’d like to see what happens if we design the glass out and bring the view in through cameras instead. Give the passengers a cabin built around their comfort, with a view that can extend beyond whichever window happens to be next to their seat.

What interests me is how one change opens the way to another. Remove the driver and you can reconsider the seating. Separate the view from the window and you can reconsider the cabin. The first improvement contains possibilities that become visible only when you question the arrangements around it.

I think a great many companies are having their Cybercab moment with AI. They have installed an extraordinary new capability inside a very familiar shape. People have better tools, tasks move faster, and the same departments still negotiate over the same people to get the work done. Those improvements are worth having. The larger question is whether the company can start directing capacity toward its changing priorities without every new assignment requiring a new position, another departmental negotiation, or a hiring cycle.

That is the distinction I mean by the 99% and the 1%. The numbers are shorthand for a widespread approach and a much rarer ambition, rather than a census of the world’s businesses. Most companies are using AI to improve the organization they already have. A smaller number are asking what they could build if the work no longer had to fit that organization. Starting with better tools makes perfect sense. The trouble comes when you assume that buying enough of them will eventually answer the larger question for you.

Table of Contents

• When the old machine keeps designing the new one
• The stories the production system lets you tell
• The idea inside the server-consolidation project
• The CFO’s thousand-percent margin
• An org chart also assigns resources
• Work the staffing plan never covered
• Why we still need horsepower
• What the firm learns as it works
• When the 1% becomes the 99%

When the old machine keeps designing the new one

The horseless carriage gave an earlier generation a familiar way to understand a new machine: you already know the carriage; imagine removing the horse. Early designs carried familiar forms into the new technology, the habit we call skeuomorphism. Familiarity helps people understand what they are buying. It can also quietly establish the boundaries of what they think they have bought.

Once you start looking, skeuomorphism is everywhere. Take the little grids on windows. When large panes were difficult and expensive to make, windows were assembled from smaller pieces of glass. Today, we can make the large sheet and add strips that make it look like several small panes. We have solved the manufacturing problem and retained it as a decorating option. We also buy electric bulbs engineered to flicker like candles. We miss the candle enough to teach the light bulb to impersonate it.

On a phone, the camera icon depicts a separate camera, and the telephone icon is a handset from the age of the desk phone. In Word, the Save icon is a floppy disk. You can know exactly what that symbol does without ever having touched the object it depicts. We put files into folders on a desktop that has no desk. Familiarity is doing useful work here: an old object gives us a way to understand a new action.

Cars carry their own collection of these memories. The hood ornament outlived the exposed radiator cap it once decorated. The radiator moved beneath the bodywork, and the ornament stayed where people could admire it. Other arrangements deserve more questioning. Why should an electric car give me only one place to plug in, behind a flap on one side of the body? The Lotus Esprit is the only car I’ve driven with a fuel cap on both sides, and I thought that was brilliant. Porsche has already offered AC charging from either side of the Taycan. Convenient access is a design choice we can reconsider when the source of power changes. I would like the charging cable to reach the car without having to reenact my last visit to a gas station.

James Watt understood the value of that familiarity. He needed buyers to understand what a steam engine could do, so he expressed its output in horsepower. People understood horses. They knew what they cost and what they could pull. If he could express his engine’s performance in those terms, they could understand what they were buying.

A 1915 account from the U.S. Bureau of Standards describes tests with brewery horses raising weights from a well. A horse pulling a 100-pound load at 220 feet per minute produced 22,000 foot-pounds of work per minute. Watt then added 50%, setting the rating at 33,000 foot-pounds per minute. The extra allowance covered friction and gave the buyer a margin against disappointment. He made the horse a tougher competitor for his engine to beat.

That was a generous estimate of what an average horse could sustain over a working day, built into a comparison intended to help people buy a steam engine. Yet 33,000 foot-pounds per minute became the standard mechanical horsepower, still in use more than two hundred years later. A number chosen to reassure buyers became the way generations understood engine power, long after most of those buyers had stopped working with horses. As a marketer, I find that almost indecently satisfying. Somewhere, surely, an agency is proposing a refresh.

In early electric factories, the old arrangement imposed some very physical limits. A steam engine drove machinery through a network of shafts, pulleys, and belts. The arrangement of the factory followed the arrangement of the power system. When a repair or alteration required stopping the shared shaft, every machine it drove had to wait. A failure in the shaft or engine could halt production across a whole room or even the entire factory. Perfectly serviceable machines sat idle until the repair was finished and the system could start again. A single mechanical problem became everyone’s lost production.

When electric motors arrived, an entirely sensible first move was to connect them to those same shafts. At the Ponemah textile mill in Connecticut in 1895, motors went into the basement near the engines they replaced; one drove 1,200 looms. New power, familiar factory. The investment made sense, and the belts still told everyone where to stand.

The larger possibilities emerged as individual machines acquired their own motors. Power could travel through wiring to wherever the work belonged. Machines could be arranged around the flow of production, factories could spread across a single floor, and one section could be changed without disrupting the whole mechanical power system. Paul David describes this reorganization in “The Dynamo and the Computer.” The steam engine could finally stop managing the floor plan, a position it had held for some time after leaving the company.

The center of the design had moved. The old factory was organized around delivering power to the machines. The new one could be organized around moving work through production. An improvement in the power supply had opened up a much larger improvement in the business.

Auto manufacturing gives us a modern version. Put a robot where a person once stood, have it tighten the same bolt, and you may get a useful improvement. Keep the old sequence, spacing, and material flow, however, and you have bought a faster arm and kept the layout. The more interesting question is what the plant could make if production capacity were easier to redirect.

Toyota’s 2007 annual report described a body-welding system capable of handling up to eight models on one line, with costs for model changeovers and additions 70% lower than under its previous system. It also described measuring precision throughout production and correcting trends before defective parts appeared. The plant could offer a different mix of cars and build them more consistently. Better cars, more choices, and a factory that could respond when buyers changed their minds. That is a considerably bigger prize than tightening the bolt faster.

In each case, the first investment could pay for itself on familiar terms: a better motor or a more consistent operation. The more consequential gains came from discovering new ways to organize production. Robots gave the line the flexibility to change its production mix as customer demand changed.

↑ Back to top

The stories the production system lets you tell

Jeffrey Katzenberg describes what those limits feel like from inside a different business in “The World is Changing: AI for Creativity.” Looking back at Disney animation in the 1980s, he recalls a production system of hand-painted cels photographed frame by frame, with revisions carrying costs measured in months. Then he makes the observation that interests me most: “These degrees of difficulty shaped the kinds of stories we could tell.” The production method had become a limit on creative ambition.

Disney’s Computer Animation Production System, CAPS, allowed artists’ drawings to be scanned, painted digitally, and combined with backgrounds and computer-generated elements. In the ballroom sequence in Beauty and the Beast, Belle and the Beast were still drawn by animator James Baxter. The ballroom was built in a computer, allowing sweeping camera movements around the dancing characters. Disney’s account of the sequence describes how the production team brought those techniques together. New production capabilities gave the filmmakers more freedom to stage the scene and draw the audience into the relationship.

That is a different way to think about the return on a technology investment. It includes what the artists can now put in front of an audience. You could measure how efficiently the production system handled drawings, but the people in the theater experienced the result as a scene that made them believe in the characters. The new method created more room for human choices about what the work should accomplish.

Every business has a version of this. Over time, people learn which ideas will survive the staffing discussion and stop proposing the others. The service is too expensive to deliver. The review would take too many people. The customer is too small to support that level of attention. Those conclusions begin as practical responses to the available capacity. Eventually, they can limit what anyone thinks to propose. A change in execution capacity can reopen those decisions and give the firm a much larger set of possibilities to consider.

↑ Back to top

The idea inside the server-consolidation project

Computing makes this progression especially clear because the freedom kept expanding. IBM’s work on virtual machines goes back to the mid-1960s, with VM/370 announced in 1972. One expensive physical computer could support multiple virtual computers. You could share the capacity without requiring each user to own the machine.

Decades later, VMware brought virtualization to the world of commodity servers. Its early enterprise appeal was easy to understand: get more from the hardware you already own. Instead of reserving a physical server for each application, run several virtual machines on shared hardware. A 2006 VMware customer story described a customer consolidating 502 Windows systems onto 25 blades. The business case was satisfyingly concrete: fewer boxes to house, power, cool, and maintain.

But the idea inside that improvement was much larger than server consolidation. Once computing capacity could be separated from a particular physical machine, pooled, and allocated through software, the company could change how it supplied that capacity. Applications could draw on a shared pool instead of each requiring another box. Virtualization helped make the cloud possible, extending the same underlying freedom beyond the company’s own data center.

In 2006, Amazon launched EC2 with the promise that customers could obtain and start new server instances in minutes, adding or reducing capacity as needed. Owning the machines became optional. A couple of founders with an idea and a credit card could get started without buying the infrastructure first. The same freedom that improved a data center’s economics helped change who could afford to start a company. That is quite a lot of ambition to discover inside a server-consolidation project.

The word connecting these stories is fungibility: units that can substitute for one another for a particular purpose. When usable capacity becomes less dependent on a particular machine or fixed arrangement, you can pool it and direct it where it is needed. Each further move opens more possibilities. Electricity frees the factory layout. Flexible manufacturing frees capacity from a single product. Virtualization frees computing from the box, and the cloud makes that capacity available to people who could never have owned the box farm.

The opportunity grows each time another restriction falls away. Better use of one computer becomes better use of a data center, which becomes access to computing for a business that has no data center at all. The people who made the first improvement did not have to foresee every consequence for those consequences to become possible.

There is a pleasing turn in the story here. Cloud infrastructure now trains and runs AI at scale. We can use that AI to apply a similar idea to work: capacity that a firm can direct toward a problem, with less dependence on the particular organizational box in which the work happened to originate. One expansion of fungibility helps supply the capability for the next.

↑ Back to top

The CFO’s thousand-percent margin

A CFO friend of mine once told me that if he added up every efficiency gain, productivity gain, revenue uplift, and cost reduction pitched to him by SaaS vendors for software he had purchased, he would be running a firm with a thousand percent margin, doubling in size every month, at a tenth of its current cost. At that point, I think you could probably close the business and retire on the savings. Unfortunately, he still had to go to work.

I keep thinking about that conversation as companies announce their AI investments. The tools can be excellent. The improvements can be real. But between making a task faster and making the company better, the gains can get lost in translation. Forty people each saving ten minutes gives you forty people with ten minutes back. It does not automatically give you the capacity to launch a new service, handle a surge in demand, or take on the client you previously had to turn away.

The essay “The Slow Death of the Enterprise” puts a useful test to those savings: “Released capacity can be valuable, but you still need to explain what happened to it and whether spending or output changed.” A spreadsheet can price forty small gaps in forty calendars as though they were one available block of time. The business still has to find a way to use them.

The saved time still belongs to particular people, with particular skills and commitments. It may improve their day, and that has value. Turning it into dependable capacity for a different piece of work requires another step. The comparison with computing concerns the allocation of capacity. People are never hardware, and their expertise is not interchangeable. The opportunity is to give them more execution capacity to direct toward the work they judge valuable.

This matters especially in businesses built on trust. A client trusts the advisor who understands the family, remembers what matters, and can exercise judgment when the situation gets complicated. The client also trusts the firm to open the account correctly, deliver the report, and complete the review when it said it would. A warm relationship will not indefinitely compensate for work that is repeatedly delivered late or incorrectly. And when a client is trying to decide whether they can retire, an immaculate set of forms is only part of the help they need.

At Humanity Labs, we call those two obligations the trust equation: relationships and delivery. Both deserve investment. The headcount model makes them compete for the same hours, even when the team is excellent and fully occupied. To grow, hire more people: That model has a ceiling. Someone needs to speak with a client; someone also needs to finish the work that makes the conversation’s promises true. Often, it is the same someone.

All of that work matters. The client conversation, the compliance review, and the accurate report each contribute to the firm’s ability to keep its promises. More dependable capacity lets the firm deliver across those responsibilities while giving people more time with clients. It also makes it possible to cover work the firm previously lacked the capacity to take on.

In “The Work Before the Advice,” our AI Workforce Journal describes a partner wealth management firm whose advisors prepared prospect analyses by entering holdings from account statements into Morningstar by hand. A complex prospect could have roughly 800 holdings and require up to six hours of entry, followed by preparing the spreadsheet, PDF, and presentation material. Including the wait to begin, the process could take up to four weeks. The firm’s identity is confidential, but the problem will be familiar to anyone who has had to prepare for a client conversation by first assembling everything needed to have it.

Now the team forwards an email with the statements. The AI Workforce identifies and classifies the holdings and returns a standardized Excel workbook and PDF the same day. The advisor spends up to twenty minutes reviewing and validating the analysis and selecting the material needed. Every analysis is reviewed before a recommendation, and the advisor remains responsible for deciding what to recommend. The preparation is ready sooner, and the person responsible for the advice has more time for the conversation.

This answers the question my CFO friend would ask about where the savings went. The firm reports that hours previously spent entering holdings are now spent with prospects and existing clients. The change gives the advisor a completed analysis to work from and lets the advice conversation happen sooner. Those are concrete improvements in delivery and in the time available for relationships. Improving existing work can be enormously valuable. What matters is whether the firm uses that improvement to change what its clients receive.

As execution capacity expands, the people running the firm have more choices to make. Which client needs deserve attention? What outcome should the firm deliver, and what will count as good and done? They set those standards, direct capacity toward the work, and remain responsible for the result. That is a Human Firm in operation: people making deliberate choices about the experience and outcomes their clients should receive, with the capacity to deliver on those choices.

That is a business outcome my CFO friend could recognize. He could see more clients being served, more of their needs being covered, and revenue growing without costs following in the old proportions. “We’re doing AI” tells him very little. So does a count of licenses. Neither tells him whether the firm has changed its ability to deliver.

↑ Back to top

An org chart also assigns resources

Jack Dorsey and Roelof Botha ask a suitably ambitious question in “From Hierarchy to Intelligence.” They describe Block’s ambition to use AI to maintain an understanding of the business and perform coordination that previously depended on information traveling through management layers. I like the scale of the question. But their thesis is incomplete. Hierarchy also assigns resources, and knowing where work belongs does not give a firm the capacity to do it.

An org chart tells you who can commit whose time, who controls the budget, and whose deadline wins. Anyone who has tried to borrow a good person from another department knows this. The presentation says “One Company.” The next forty minutes concern why you cannot have three hours of someone who works in a different part of it. There are rational reasons for the argument: managers have commitments, people have specialized skills, and everyone’s calendar is already full. Better information does not make those constraints disappear.

Dorsey and Botha do give leaders authority to pull resources across teams. The question is what happens when they pull. If each request still depends on negotiating access to scarce, already committed people, better coordination has improved our understanding of the queue. Knowing precisely why everything is late is a management accomplishment, although clients tend to prefer just getting the work done.

Fungibility addresses what happens after the priority is set. AI capacity can sit in one pool and move as demand changes, drawing on shared firm knowledge and operating under the firm’s control. The firm’s priorities become the basis for allocation. Work that used to require a negotiation between departments can draw capacity from the same pool.

Consider a wealth management firm bringing on a large group of new clients. For a few weeks, account-opening work rises sharply. Under the familiar arrangement, the firm might borrow people, pay overtime, postpone something else, or hire for a peak that may be over before the new employee arrives. Give the existing team a faster tool and you may ease the pressure. The capacity still largely follows the people assigned to that team.

Now imagine the firm directing its AI Workforce toward the increase. Account opening and client-review preparation are both work the pool already knows. Its people decide that the onboarding surge takes priority, and capacity moves toward account opening. As demand recedes, that capacity moves toward preparing client reviews. The firm’s knowledge remains available throughout, and its people retain control of priorities and standards. The same pool serves the change in demand.

Four things have to work together for this to be useful. Firm memory supplies one store of knowledge every unit of capacity can use. Pooling makes all the capacity available to the firm’s work. Dynamic allocation moves it as demand changes. Governance keeps execution under the firm’s control, at the levels of automation it chooses. The firm’s outcome owner holds the delivery standard, and its process owner protects the dependencies and controls the work must preserve.

All four are required. A firm could redesign twenty workflows and still have twenty separate systems, each with its own knowledge and capacity. Shared knowledge alone does not move capacity, and a collection of separate departmental tools does not become a pool simply because everyone bought them from the same supplier.

Governance becomes easier to understand when you can see the decision it controls. In Issue 2 of our AI Workforce Journal, “The Wire,” an unnamed partner multi-family office uses its AI Workforce to verify and log roughly 700 wires a week. It matches account and routing details against the custodian’s record and verifies the reference number. Anything it cannot confirm goes to a person at the firm before the log entry is written. The rule specifies both what must be checked and when human judgment is required. The firm’s control is built into the work.

Together, those four parts are what we mean at Humanity Labs by Workforce Virtualization: pooled capacity, dynamically allocated to the firm’s demand, on shared firm context, under firm control. The product a firm buys is the AI Workforce: one pool of AI capacity, on one firm memory, working across the front, middle, and back office, all under the firm’s control. People decide which outcomes matter, set priorities and standards, and own the result. They gain capacity they can point at problems and opportunities throughout the business.

It also explains why an “AI employee” can be a surprisingly conservative idea, and a deeply flawed answer to the problem. Give each AI employee a permanent job description, isolate its knowledge, and keep it inside one department, and you reproduce the old organization in software. The company still allocates work through fixed positions and departmental boundaries. We have invented a new kind of capacity and immediately given it the organizational restrictions of the people it was supposed to help.

This is the robot standing where a person stood, tightening the same bolt on the same line. It is the horseless carriage again: a new capability understood through the system that came before it. That first substitution can improve the work and pay for itself. But treating it as the destination builds the old constraints into the new technology. The company gets better at doing what it already does, while its ability to change what it does remains limited by the same structure.

The real promise of AI in the workplace is to give the business a flexible workforce through pooled capacity. Workforce Virtualization lets the firm move that capacity toward problems and solve them quickly, then direct it toward opportunities that can make the business grow faster. Shared firm memory allows the knowledge to travel with the assignment, and governance keeps the work under the firm’s control. The organizing principle becomes the work the business needs done. As those needs change, the allocation of capacity can change with them.

↑ Back to top

Work the staffing plan never covered

The first assignments will often be work the firm already does. That is sensible. You understand the task, know what good looks like, and can tell whether the new arrangement improves it. But once you can direct more capacity toward a problem, the list of worthwhile problems starts to change. A review that was previously too expensive to perform broadly becomes possible. A service that could only be offered occasionally becomes something a client can rely on continuously.

We are seeing this in our work with Mariner. In our published account of its Institutional business, the team put its AI Workforce to work examining public retirement-plan filings against Mariner’s own criteria. It excludes plans Mariner already advises, ranks potential opportunities, and adds information the team needs to evaluate them. Work previously done one prospect at a time can now cover a much larger market.

Of the first 8,700 plans analyzed, 5,900 cleared Mariner’s screens, representing roughly $20 billion in plan assets it did not advise. Those are prospective opportunities, not assets Mariner has won. The people still have to assess the fit, build the relationships, and earn the business. But they can begin with opportunities the firm has systematically identified and prioritized, including ones the team previously lacked the hours to find.

That also changes the order in which the team works. Previously, qualifying a plan meant lookups across three systems, one plan at a time, and the team worked the opportunities that reached it through relationships and referrals. Now it can start with the best opportunities under Mariner’s own criteria. The people decide what makes a plan worth pursuing and change those criteria as their judgment develops. The AI Workforce applies that judgment across a market the team could not previously cover, giving its people a better basis for deciding where to spend their attention. The firm can choose a broader ambition because it has a way to carry out the work.

Law firms offer another way to see the distinction. Buying AI tools to help with research, review, and drafting can be a good investment. Better tools, however, do not automatically change how matters are staffed, what clients can buy, or how the firm charges for it. In Thomson Reuters’ 2026 survey of 736 law firm professionals, only 2% described rebuilding practice around AI as their daily reality. The report describes firms whose tools have changed while staffing and delivery remain largely familiar.

Imagine a firm offering ongoing monitoring of an agreed portfolio of contracts and obligations, with lawyers handling the consequential issues that emerge. Demand for review might surge when circumstances change, then settle back down. The firm could direct its AI Workforce toward the extra review work as demand rises. The firm would have to design the service, establish its quality controls, and work out the commercial model. But a client who could previously afford an occasional review might now be able to afford continuing attention.

That is where these changes become interesting to someone outside the firm. Clients do not buy an organization’s productivity statistics. They buy help with something that matters to them. A previously unaffordable service, more complete coverage, or a problem caught early enough to do something about it gives the efficiency a purpose.

It also changes the growth discussion. There is work sitting in today’s queues, and there is useful work that never made it into a queue because everyone knew the firm could not staff it. We describe these as bounded and unbounded work: the existing work the firm needs to complete, and the work no staffing plan could cover. The second can be much larger than the first. You will not discover it by asking only how to clear the current backlog faster.

↑ Back to top

Why we still need horsepower

There is a commercial question underneath all of this. How does a firm buy a pool of work capacity? The CFO still needs a unit, a budget, and a way to judge whether the purchase makes sense.

This is where Watt comes back into the story. At Humanity Labs, the Virtual FTE is the unit of provisioned capacity, priced against headcount and funded from the labor budget. The comparison is with what the firm would spend to add capacity through hiring. It gives the CFO a familiar economic reference for something that can work very differently from a new hire.

Workforce Virtualization names the category. The AI Workforce is the product the firm buys. Virtual FTEs measure the capacity in that pool. A Virtual FTE is provisioned capacity, not a virtual employee; buying ten units adds capacity to the pool, without creating ten separate identities, ten stores of knowledge, or ten permanent places on the org chart.

The same distinction explains a role template. Define a role once as a description of work the pool can perform. The pool takes up that work wherever demand appears, then moves when demand moves. The template gives capacity a reusable definition of the work; it does not give an AI employee a permanent job. In the account-opening example, the firm can add capacity for the surge and direct it toward client-review preparation when that becomes the priority.

Using a familiar measure does not require preserving the familiar organization. Watt’s customers could buy horsepower without building a stable for the steam engine. A firm can buy capacity in Virtual FTEs without assigning each unit a permanent seat in a department. The unit makes the economics understandable; pooling makes the operating model more flexible.

Then comes the question every operator will eventually ask: who is responsible when something changes? A custodian alters its account-opening requirements. A system the work depends on is updated. New-client volume doubles. The firm still needs correct work delivered on time. Someone has to keep the procedures current, adjust the capacity, and fix what has stopped working.

When we supply an AI Workforce, keeping that pool working is part of what we are supplying. We agree on the work and the standards, provide the capacity, and remain accountable for delivery. The firm sets the rules and decides what needs doing. It should be able to judge the service by the quality, timeliness, and cost of the work.

This also gives the CFO a better question to ask than how many minutes a demonstration saved. What work will this capacity complete, to what standard, at what cost, and what can the business now afford to attempt? The unit may resemble a line in the staffing budget. The opportunity can be much larger than replacing the line.

↑ Back to top

What the firm learns as it works

For the capacity pool to be useful, it needs to learn how your particular firm operates. That knowledge has to be captured, maintained, and made available under the right permissions. A generic understanding of account opening is helpful. Knowing what counts as a completed account opening at your firm, which exceptions require approval, and what changed in the procedure last month is what makes the capacity usable.

Katzenberg’s essay contains another useful part of the Disney story. He describes turning to the archives, where recordings, notes, and storyboards preserved Walt’s thinking about character, emotion, and what made an audience believe. Those materials gave the team a way to understand the standards behind the work as they reconsidered how to produce it. The studio had retained judgments about what mattered and why. A firm needs that kind of knowledge too: the decisions and standards that help people judge whether a new way of working still delivers the result they intend.

Shared firm memory makes capacity easier to move, because every unit uses the same store of firm knowledge. It also allows experience to accumulate. When a person resolves an exception and the firm approves what should happen next time, that decision can improve subsequent work across the pool. The next assignment can begin with something the previous one taught you.

Work capacity can become interchangeable while a firm’s accumulated knowledge remains particular to that firm. A new unit of capacity can pick up an established assignment using the same approved procedures and relevant history as the unit before it. Adding capacity should extend access to what the firm has learned, wherever in the firm the work originated.

The promise is precise: capacity can be added at provisioning speed for work the pool already knows; a new kind of work is taught once, and the firm memory keeps it. Teaching the pool how your firm opens an account is an investment. Once that knowledge is in the firm memory, adding capacity for more account openings reuses it. Growth does not require repeating the same training for every additional unit of capacity.

The same firm memory also helps management understand how the business is running. When work across the front, middle, and back office shares a common context, the firm can understand how those activities connect to the result a client receives. Management can examine performance across the whole process, including the handoffs between departments. The enterprise essay makes a useful recommendation here: follow the customer’s problem upstream and downstream through the work. A faster task may do little for the client’s waiting time when the delay sits at the next handoff.

At Mariner, this is turning out to be a very big deal. The firm can finally understand how it is running from end to end. Alongside the capacity to do work it could never previously cover, it is gaining a better understanding of its own operation. That changes the basis on which people can make decisions about how the firm should run.

A separate partner story, “The Gate,” in Issue 1 of our AI Workforce Journal, shows why following work through to the client’s result matters. At a wealth management firm whose identity remains confidential, a risk tolerance questionnaire had to be reviewed before a new account could fund. The active review took only minutes. Turnaround from submission to completion was about nine and a half hours. Much of the difference was work waiting in queues and moving through handoffs. Speeding up those few minutes of review could leave most of the client’s wait untouched.

The AI Workforce now reviews and documents the questionnaires against the firm’s own rubric, completing routine cases without a manual handoff. Turnaround fell to about twenty minutes, with quality reported as consistent with manual review. Time to account funding fell from about a week to about a day. That is the electric-factory lesson in a setting with considerably fewer belts: the arrangement through which work travels can matter far more than the speed of an individual task. For the client, the improvement is an account that can be funded sooner.

The same story makes learning concrete. Ambiguous reviews go to the firm’s human overseer. Each ruling resolves the case and informs how the AI Workforce handles subsequent reviews. Human judgment remains responsible for the exception, and the decision becomes useful beyond the one case. This is the kind of accumulated knowledge the firm should expect to reuse as its capacity grows.

This is where understanding the business and being able to direct capacity belong together. A surge in volume may call for more capacity. A recurring exception may call for a change in the process. The firm’s people can decide which response will help, direct the work, and evaluate whether the client receives a better result. Approved changes become part of firm memory, so what the firm learns improves how subsequent work is done. The value of understanding the whole operation grows when the firm also has the capacity to act on that understanding.

That improvement takes ownership and discipline. A pile of old documents is capable of remembering obsolete instructions with tremendous enthusiasm. Someone has to decide which procedures are current, which lessons are valid, and where human judgment remains necessary. The point of firm memory is to make approved knowledge available when the work needs it, and to keep improving that knowledge as the firm learns.

Another firm can buy similar computing capacity and use the same models. What it cannot get from that purchase is your history of resolved exceptions, approved procedures, and decisions about what good looks like for your clients. Those accumulate through work. Starting now matters because the firm begins building knowledge that will make its future capacity more useful, even as the underlying technology changes. An innovation press release is considerably easier to produce. Unfortunately, it knows very little about your customers.

You can begin without reorganizing the entire company. Find three to five people who want to change how work gets done, give them a sponsor who can make decisions, and choose a meaningful problem or opportunity. For every assignment, name an outcome owner and a process owner inside the firm. Those responsibilities make human judgment and accountability explicit as the capacity to execute expands.

The outcome owner knows what good looks like. This person defines the result the work must produce, establishes what good and done mean, and remains accountable for ensuring the delivery standard continues to be met. For an account-opening assignment, that might mean specifying which checks must be complete, which records must be updated, and what confirmation the client should receive. The outcome owner judges whether the work meets those requirements as volume and circumstances change.

The process owner knows how the work is actually done today. This person understands the sequence, the systems, the handoffs, and the exceptions people resolve to get a result. The documented process may show five steps; the people doing the work may know that step three requires a spreadsheet, a phone call, and approval from someone who appears nowhere in the diagram. We may discover a better way to do the work. Starting from an accurate picture of today helps us understand which steps serve a necessary purpose and which can change. As the work changes, the process owner keeps that understanding current.

Both roles belong to people in the firm. Naming the owners establishes responsibility for the work while execution capacity remains available across the shared pool. A new assignment does not require new hires or a reserved portion of AI capacity. Humanity Labs remains responsible for keeping the AI Workforce working.

Product experience tools such as Pendo can help the process owner see the paths people actually take through applications. The clicks can reveal a repeated detour that nobody thought to mention, or a sequence that differs from the official procedure. In “From Clickstream to Control Plane,” I explored how that evidence can help build a more useful account of how a firm operates. A clickstream gives us a record of what people did; the process owner helps explain why they did it.

Together, the two owners can use that end-to-end understanding to improve how the work gets done while holding the result to a clear standard. The process owner identifies what can change and which dependencies or controls must be preserved. The outcome owner judges whether the new approach produces the required result. Their approved decisions become part of the firm memory, available across the AI Workforce. Assign the first capacity, learn from the result, then test whether that capability can serve more work or a different priority without rebuilding everything around it.

Revenue per employee remains a useful headline measure, with revenue growing and service quality maintained or improved. Reducing headcount can improve the ratio without improving the business. The supporting evidence needs to show whether the firm is delivering more of the trust equation and doing work it could not previously undertake:

• Capacity returned to people. Identify the work the AI Workforce now completes and measure the time and attention people can redirect. Show how they use that capacity for client needs, judgment, and decisions about what the firm should deliver. Minutes saved become meaningful when the firm can account for what they made possible.
• Additional work completed. Name the reviews, analyses, or services the firm previously could not cover, and show their new scope, frequency, and completion standard. Distinguish a faster version of an existing task from work that has become possible for the first time.
• What clients receive. Show the additional coverage or service clients now get, with evidence of accuracy, timeliness, and follow-through. Use client feedback to assess whether the extra attention is improving relationships. More available time is a resource; stronger trust is an outcome the firm still has to earn.
• What the firm can now understand and improve. Identify something the firm learned about how work runs from end to end, the decision it made as a result, and what changed in delivery. A more complete view becomes valuable when people use it to improve how the business serves its clients.

The allocation claim needs evidence too: one verified case of capacity moving between different kinds of work as demand changed, with the delivery standard maintained. That demonstrates the flexibility of the pool. A count of Virtual FTEs establishes its size.

Finally, track the routine human attention required for comparable work at the same standard. As approved procedures and resolved exceptions become reusable, that requirement should decline. The outcome owner remains accountable for deciding whether the result is good enough.

These are useful tests because they connect the technology to what the company can choose to deliver. The next growth discussion can begin with the client experience the firm wants to provide and the outcomes it intends to achieve, informed by a clearer understanding of how the business performs today. Under the headcount model, the first question is which people and positions the firm needs to add. With an AI Workforce, the question becomes what work needs to happen and where to direct capacity. The management job is to decide what is worth doing, define what good and done mean, and own the result.

We call the destination the Firm of the Future, with growth decoupled from headcount. People will still be hired for the judgment, relationships, and standards they bring. The firm can provision execution capacity to carry out the work those people decide matters. A service that once disappeared from the plan because it required too many hires can become a real offer to a real client.

↑ Back to top

When the 1% becomes the 99%

Starting in the 99% is a perfectly sensible way to learn. Buy the tools, improve the work, and take the savings. You can do all of that while examining what the new capacity makes possible. The consequential decision is whether you keep going. There is no automatic graduation ceremony, and a collection of successful pilots will not reorganize the company while everyone is at lunch.

Now put yourself on the other side of the desk in that law-firm example. You have been buying occasional reviews because that was what you could afford. Another firm offers to keep watch over the agreed set of contracts, flag issues as they emerge, and bring in a lawyer when judgment is needed. Once that service proves dependable, you have a different expectation of what a law firm can do for you. Your old firm may be drafting documents faster than ever. You would still like to know why it cannot offer the same attention.

“We’d love to do that, but we’d have to hire an entire team” may be a perfectly accurate description of the firm’s economics. It gives the client very little reason to accept less. A staffing constraint that once seemed natural to everyone has become one supplier’s problem. Customers have a remarkable ability to lose interest in your org chart the moment someone offers them a better result.

That is how I expect the 1% to become the 99%. Each new possibility gives customers something else to ask for, until what seemed extraordinary becomes expected. The industry changes around those expectations.

For firms that want to be in that 1%, the opportunity is extraordinary. These are the moments when new businesses emerge and challenge an industry’s old guard. A smaller firm can deliver a service that once required a much larger organization; an established firm can enter markets its old economics could never support. When valuable work becomes affordable to do at a much greater scale, whole new businesses become possible. I think AI could create one of the largest opportunities for building companies and creating wealth that we have ever seen.

If you want to be in the 1%, start with the customer need your old staffing model made too expensive to serve. Use the new capacity to make that service commercially possible. You have a chance to set the standard the rest of your industry will have to meet. Take it.

↑ Back to top