AI is fast. Automation is slow. Can they meet?

Recent studies show we can get better results by using generative AI. But to get them at scale, our organizations would have to change, and that's slow and difficult. genAI itself may change all that.

“Come up with an image to kick off a Substack post of mine. Be very creative. Convey optimism, ambiguity, and potential risk. The title of the piece is "AI is fast. Automation is slow. Can they meet?" Do not place text in the image.”

Most voices out there make it sound like the world of AI-change is a runaway train. Just about a year ago we got GPT-4 via a text-based interface. Things have exploded since then. We’ve gotten the ability to interact with these systems via images, audio, and video. In the last week or so we got adaptations that can create short, high-ish quality videos and songs in mere moments. We’ve gotten the ability to ask for code-based solutions to our problems - with these systems testing and debugging their own code. And now we can - just using natural language - program up tailored “GPTs” to solve focused problems (like improving our prompts, tutoring for HBS-style case writeups, and teaching us to identify rhetorical questions) that we can share with each other. Last week a startup (supported by the artist Grimes) announced it has embodied these kinds of agents into “Grok”, a plush toy that will interact verbally with children.

And this isn’t all contained in a few private, closed companies like OpenAI and Google: just in the last week, we’ve had huge new drops of open forms of generative AI that anyone can run on their computers - or even phones - for the cost of the electricity involved. They’re about as good as the GPT-3.5 system that ran under OpenAI’s ChatGPT just under a year ago. Thousands - perhaps tens of thousands - of top coders in the open software community are feverishly at work trying to surpass companies’ closed models. Many say they will achieve all this, and soon. On top of that, the open community is making it easy to get groups of GPTs to work together in digital organizations to solve complex problems - with only initial human guidance. Companies are now offering this functionality in beta trials.

Head spinning with “progress” yet? This advancement isn’t just good for run-of-the-mill administrative, writing, and analytic tasks for the masses: it’s enabling breakthrough science that likely wasn’t possible without generative AI. Also just in the last week we found out that - with a little artful scaffolding - genAI has produced genuinely creative breakthroughs in math and in biology: a new solution to an age-old calculation problem and an entirely new structural class of antibiotics - the last one took 38 years to find. Even Terrence Tao, the world’s top mathematician, has said he’s now better at math with these tools.

And yet…

Automation is slow

Here’s the thing about getting real world results - at scale - from new automating technologies: it’s slow and difficult, even when we want it.

In fact I talked a bit about this two posts ago, to help explain why - for most folks - there’s no need to rush on genAI. The decades-long delay between the arrival of new automating tech and widespread results from automation boils down to how organizations work, but I didn’t explain how that is, exactly. In fact, the most you’ll usually get on this point is some hand-wavy mention that it takes a long time to figure out the right way to put new automating technologies to use, and that lots of other complementary things have to change - in our lives or an organization - for that to happen.

But while it’s accurate to say “automation is hard and complicated, and therefore slow”, it’s also deeply misleading. Stopping there obscures what social scientists call mechanisms - the specific interaction patterns that actually cause outcomes to happen or not to happen. If we don’t try to understand those, we’re ignoring huge swathes of psychological and sociological research on how individuals, groups, and organizations change as new forms of automation emerge. If we do understand mechanisms, we can make better predictions about when automating something will be quick, when it will be slow, when it will be easy, when it will be hard.

One key reason automation is slow is because we don’t have the skill to put it to good use. And this even affects individuals. Let me share one hot-off-the-presses working paper on Kenyan entrepreneurs to illustrate. Productive automation using ChatGPT should be quick and easy for these highly motivated folks, right? They’re certainly not burdened by bureaucracy.

Wrong, it turns out. Or at least only partially right.

In this fantastically well designed study, Nick Otis (a PhD student!) and coauthors gave one group of entrepreneurs access to a ChatGPT-style business mentor to aid in their work. A control group got no such help. I highly recommend this detailed thread explaining the work. But in short, all of these entrepreneurs were super motivated to do well. They were taking significant risk to try to found a business, and rapidly tried to take advantage of the genAI they were given. They could automate instantly.

“Let's go much more creative. Those all have a cleaner, more modern aesthetic and are very straightforward. Remember the wildness associated with my brand and this substack - there should be more strangeness, wonder, optimism, and VERY varied art styles. Go for it, have fun, you can do it!”

Some did race ahead, it turned out: the entrepreneurs who were already performing very well saw a 20% boost in key outcomes. But lower performers saw a 10% decrement in their outcomes. That’s right. They did worse.

And we get some insight into mechanisms here: it looks like it’s because “both groups sought the AI mentor’s advice, but that low performers did worse because they sought help on much more challenging business tasks.” In other words, even if you want to automate something badly and no one is there to try to slow you down, you can fail: either you don’t have enough expertise in the task, it’s highly ambiguous, risky, or all of the above. Trust genAI too much in those circumstances and you're liable to get hurt - not just fail to automate.

But let’s put fast-moving, untethered, hyper-motivated entrepreneurs to the side for the moment. The huge mass of the consequential automation in the world arises through organizations, and…

Organizations are still using dot matrix printers

“Show a dot matrix printer, hyper photo realistic style, lower lighting, with the dot-matrix printed text on its paper that says "incompatible with generative AI". And remember, dot matrix printers have no keyboard. Just show the printer head, the paper, and most of the printer, but not even all of it.”

Over the holidays, my wife and I traveled east to spend time with family (I’m in the air on the way home as I finish this). And I noticed: every gate had a dot matrix printer - a technology commercialized in 1968 (seriously, geek or not, go watch that explainer video). Compared to modern printers, these are more expensive to operate, harder to resupply and maintain, are loud, produce lower quality output, demand additional training, and require weird software and paper that are wildly out of step with modern corporate IT and supply chains.

On the surface, this is profoundly irrational.

But retaining archaic automating technology turns out to be a side-effect of the way organizations automate in general. Every organization, everywhere, has the moral equivalent of dot matrix printers in service - I bet you are already thinking of one or two examples in your context. And the same mechanisms that drive the dot matrix phenomenon also impede automation that relies on the latest technology.

Anyone who’s yelling loudly about the need to put LLMs to use right away doesn’t understand this reality. And if you’re a pragmatist - someone who wants to get great results with this new general-purpose technology without sacrificing your organization or team to do it - you need to understand this reality.

In this section, I’ll explain more why that is. In the last section, I’ll suggest some ways we will use genAI itself to change this game in a way it’s never been changed before.

Organizations are like houses

I’ll illustrate the status quo through two papers of mine. The first is coauthored with Paul Leonardi this year, and it’s summed up in this figure:

The short story? For new tools like ChatGPT to add business value, they need to be integrated with the foundational elements of an organization: its structure (e.g., roles, reporting relationships, norms), infrastructure (e.g., functions like recruiting, procurement), strategy (i.e., what it’s trying to accomplish), and field (e.g., industry). Paul and I show this by reconsidering nearly a hundred years worth of study of organizations, work, and technology, and suggest that it all indicates one thing, summed up by this quote from Stewart Brand in his magnificent “How Buildings Learn”:

“Fast and small instructs slow and big by accrued innovation and by occasional revolution. Slow and big controls small and fast by constraint and constancy. Fast gets all our attention and slow has all the power.”

Some parts of organizations change more slowly, because they are more foundational, and these are interdependent with the parts of the organization that change with the whims of global fashion. From this point of view, ChatGPT is the new Slack - a cool new automation tool that a lot of individuals are jumping on. Sometimes this can become the tail that wags the dog - a lot of individual Slack users within a company can drive the stewards of the company’s IT infrastructure to implement Slack for the firm. But most of the time, new tools like Slack will not get put to use in the organization for a long time - or maybe ever - because no one does the work needed to mesh them with the firm’s infrastructure, structure, and possibly strategy. That’s a huge reason why Microsoft’s teams - a demonstrably crappier product - has wildly outperformed Slack in the market. It’s compatible with the IT backbone of the organization, which was developed, installed and maintained via Microsoft-connected product and knowhow.

On top of this:

Organizations are like investment portfolios

Another key point to consider is that keeping an organization running takes resources - cash, people, space, time, attention, effort, and so on. And while investing these resources produces resources in return - you get things like customers, revenue, employees, skill - no organization has unlimited resources.

That means organizations invest unevenly across their automation portfolios. Some automation gets more attention, cash, and people than other automation. And we’ve learned that organizations overwhelmingly prioritize investments in proven, familiar domains - because it’s the equivalent of printing money.

I showed the implications of this portfolio problem this year in a paper examining robotic surgery (open access). To get my data I spent over two years at one hospital that disrupted its automation portfolio by buying an upgraded surgical robot while retaining the older one.

To maximize short-term ROI and expand into new markets, the hospital shunted the best supplies and infrastructure to their new system - and gave privileged access to inexperienced talent. That safer, easier environment let newer talent get up to speed faster. But that also meant the hospital preferentially allocated its most experienced talent to their older surgical system, while depriving that robot of preferred maintenance and infrastructure. It broke down, often in the middle of surgery. Surgeons had to operate with one eye blinded. Nurses had to doublecheck “failure proof” components. Scrubs had to stash extra supplies and force various components open and closed. Sounds like a recipe for inefficiency and harm, right? It did to me. So I collected years of surgical outcomes data to check. It turned out patients did just as well on what became known as the “shitty” robot.

Read the paper for the (literal) gory detail, but there was no dip in surgical outcomes because that experienced talent invented workarounds to produce the same results as those achieved with their newer system.

Here’s the twist, and the link to the slow automation story above: getting good at those workarounds meant that top surgical talent only “whined” about the older robot to management. They never got data about how difficult it was to use, and how it was breaking down. And management never came by to check things out themselves. Ultimately, that limited execs ability to consider new automation options. From their point of view, things were working just fine. So the hospital was running to stand still because it was carrying aging technology that was difficult to use, put extra burden on top talent, and that was degrading over time, given reduced maintenance.

Thinking dot matrix printers yet? You’re on the right track.

This portfolio-driven blindness to the pain and waste of the automation status quo is part of what will block many organizations from noticing and investing appropriately in generative AI - the most consequential new technology to hit the global market in a couple decades. They’ll have a thicket of automation that seems to do the job just fine, and keeping it all working may mean they may not even see the genAI opportunity in front of them. If they do, it might not look that attractive, because their current ways of automating things are familiar, reliable, and effective. Yet those ways will have accreted slowly, with some (usually older) parts of the portfolio getting short shrift, and therefore demanding more and more hidden effort to keep them on their feet.

In a way, this is all a recipe for what Bob Sutton and Huggy Rao call “friction” in their new book. The longer organizations run, the more they accumulate processes, rules, even culture that make it harder and harder to take on something interesting and new. Automation - essentially a blend of work processes, tools, skills, beliefs and assumptions - is no exception. Organizations accumulate and then master various forms of automation which facilitate good performance at a certain level in certain external conditions. But when a potentially disruptive tool comes around, it can only become part of the way the organization automates things if most of the people involved in related structure, infrastructure, and firm strategy make significant change that has implications for many, many other processes, tools, and situations. When the new automating tool is also a general purpose technology - a tool that helps in tasks throughout the economy - the range and intensity of this change explodes.

So: most organizations won’t try to automate anything with genAI, at least at first.

Many will ban it. Not because of the risk to IP, or for ethical reasons, but because it’s incompatible with the foundational elements of the firm, and its portfolio of prior automation investments.

Experiments will be small.

And all of this will frequently be Just Fine. Those who prefer rapid change, exploration, and risk may dunk on fuddy duddy organizations that are taking it slow with genAI, but betting the farm on a new technology is usually not prudent. The friction and inertia associated with the foundational elements of an organization are part of what allow that organization to adapt to change. If a few small experiments yield solid results, bigger ones will follow, enabled by the very foundational processes that impede rapid, wholesale change. Over the course of a few years to a decade, a firm can reconstitute itself to take wholesale advantage of a new general purpose technology - without tearing itself apart.

And yet…

Change is the new Stasis

“Make a plot of general purpose technologies from the wheel to modern day using code - a simple graphic representation is fine, but I need this to be precise and accurate. So don't do this via your image creation capability.”

Take a look at that timeline.

For most of human history, the average organization experienced *zero* new general purpose technologies. From about 1250 to 1750, we got one about every 150 years. So still, practically zero. Since 1950, we - and our organizations - have experienced five. And three of those have come just since 2000.

Call it exponential if you want. ChatGPT did: it quickly fit the data with a four term quadratic polynomial. It’s clearly not a linear trend. And our most recent arrivals are self-improving in unprecedented ways. So from this point of view there’s a new intensity layered in there on top of a simple increase in the rate of accumulation. (It’s important to note, by the way, that some argue that our best inventions are behind us - I highly recommend that argument if you haven’t heard it before)

Previous ways of running a firm and “ingesting” new automation may not cut it anymore. Top business schools teach techniques for generating “best in class” versions of that house diagram in my paper with Paul. Those techniques were devised between 1950 and 2010 or so. They don’t presume an exponentially increasing rate of change in the arrival of general purpose technologies. And they certainly don’t presume generative AI - most notably its likely accelerating effect on a great deal of innovative activity.

I don’t have a well worked out alternative to offer you. No one does. That’s partly because discovering it will require doing it. The first organizations that operate in ways more amenable to change in general - and adopting “intelligent” automation in particular - will likely race ahead, and others will learn from their example. Andy McAfee’s latest book “The Geek Way” - the definitive, theoretically-grounded account of what makes silicon valley organizations tick - likely offers some clues here.

But even the best of pre-GPT organizational wisdom isn’t liable to explain how organizations will handle automation moving forward. First off, you can see above that putting new automation to work creates unintended side effects for an organization. It can divert resources away from other efforts, which can create strain and inefficiency. But also adaptive behavior that blinds everyone to opportunity. No one pays or can even see these costs clearly. These dynamics are akin to climate change, given rapid growth in our civilization’s energy consumption. For that we’ll probably create…

AutoMGT

“Try again on AutoMGT. Yes, there should be optimism. Cool colors. High energy. But also the unknown, and danger. If you read the material from that section, it says there will be nasty, hard to catch side effects of using AutoMGT. Depict that, too. Try this in three VERY different styles. Really push yourself to be very creative.”

To help our organizations keep pace with exponential change, we share management responsibility with intelligent machines.

Right now, humans are in primary control of each of the elements of an organization - whether lightweight and rapidly changing, like tools, or massive and slow, like infrastructure. When a new form of automation comes along, it’s the human interpretation and action that controls how an organization assimilates it. For a new tool to become part of the automating fabric of the organization, many dispersed humans with diverse interests, skills, roles, and information need to change the way they think and act - even if it’s just a little bit.

That worked pretty well from say 2,900bc to 2022.

If we keep creating general purpose technologies at an exponentially increasing rate, we and our organizations will find ways to adapt at that rate too. Humans can’t do that without help. The only way to keep pace with genAI-enabled automation will be to automate the automation process with genAI.

We’ve been scaffolding ourselves towards a hybrid human-machine form of management at least since the advent of the spreadsheet, the algorithm, and the sensor: managers now get massive, rich datasets to inform decision making, and these are automatically analyzed and presented in ways that help those managers make higher quality decisions far faster. A manager in a top performing organization today can help their organization adapt to change far better than they could in 1980.

What we are just barely starting on is getting LLM-generated analysis and advice on how to structure our organizations in response to change. We can now use generative AI to analyze the vast repository of data that our organizations collect and produce to propose productive changes to tools, structure, infrastructure, symbols, and strategy. Once we build genAI-enabled systems like this, humans will still decide whether and how to buy new technology, change the org chart, rebrand the company, or change its strategy. But excellent managers will rely on this technology for decision support on these topics. For an early example, check out WorkHelix - cofounded by three colleagues - Erik Brynjolfsson, Andy McAfee, Daniel Rock (coauthor of the GPTs are GPTs paper) - and James Millin.

This is AutoMGT. The equivalent of a software-generated design feedback for a semiconductor engineer, predictive analytics for a leader on climate change, a legal and safety compliance alert for an architect designing a neonatal intensive care unit. It’ll just be about how your organization is built and how it runs. If this reality unfolds, we will get used to systems like this. At least in “situation normal” conditions, we will learn to accept their recommendations as routinely as we do the terms and conditions for our phone software updates. We’ll scan, assess, then click “yes”. It’s not the same as changing the organization itself, but this is literally what happens now for front-line hiring and layoffs in the warehousing organizations my team and I have studied for the last few years. A human manager makes the call, but software provides the targets. We’re already in the habit.

We don’t have good precedent for understanding what it will be like to work in an organization reliant on AutoMGT. What organizations like this will be capable of. And what the unintended side effects will be. That last point is key: AutoMGT-style organizations won’t be unambiguously better. If the sociology of work, technology, and organizations teaches us anything, it’s that some of those side effects will be nasty, accumulate slowly, and will be practically invisible to everyone involved - until it’s too late. But the adaptive benefits will be proximate, clear, and significant, so we will take the deal.

Living and working in organizations will start to get a bit… weird as we co-design organizations with genAI. Exciting new opportunities, subtle deep threats. Now is a time to pay fierce attention to both sides of the ledger. Study, test, and share findings on how to handle this new territory, and do our best to make choices that insist on improved productivity and enhanced human welfare.