AI Will Pay Off. But Don't Get Complacent About It.
Why the Technology's Benefits Can't Come Soon Enough
Steve Lohr had a great article1 last week in the Times about the apparently slow pace of clear gains from AI. As he writes,
businesses will have to continue to invest billions to avoid falling behind — but it could be years before the technology delivers an economywide payoff, as companies gradually figure out what works best.
Lohr lists several reasons why all the AI goodness might slow to appear: employee and customer resistance; “the hard slog of mastering a technology;” and of course those pesky hallucinations.
Those all sound right to me. But the slow appearance of gains from today's generative AI is still a puzzle, isn't it? After all, this technology isn't just powerful and versatile. It's also easy to use. You just talk to it. Good prompting requires training and practice, but it doesn't require learning an arcane new language or even a new interface. You just type words into a box, which is something we've all been doing for years now. Or even more simply, you do what I'm doing now: you hold down a button and talk into a microphone.
Some Technologies Are Too Good Not to Use
In a sense, today's AI reminds me of the calculator. Before it appeared, professionals who worked with a lot of numbers used slide rules or clunky electromechanical adding machines. By about 1975, ChatGPT tells me, battery-operated 4-function calculators were available for about $300 in today's terms. This made them affordable for people like my dad, who designed and sold HVAC systems. I remember playing with his first one (a Radio Shack with an LED display, IIRC) and being amazed that it could do arithmetic with big numbers so quickly.
My dad was no math whiz, but he quickly learned to use his calculator and used it a lot. It made his job easier and put hours back in his week. I predict the same thing will be true with AI. More and more people are going to use it at work because it offloads drudgery, helps them with their more interesting work, and puts hours back in their week. Does anyone believe that when we check back in, say, five years, the AI fad will have passed? Or stalled out as a tool used in just a few corporate niches, like customer service (chatbots) and IT (coding assistants)?
I don't think either of those scenarios is likely. I think more people in more jobs are going to be using AI more often in the future, and getting benefits from it. So then one approach to bringing AI into your organization is to just let that process unfold. Let experience accumulate and workforce demographics change (new entrants being much more comfortable with AI than the retirees they're replacing) And AI and its bounty will almost certainly spread.
They’re Also Too Good Not to Accelerate
One good reason not to take this approach is that shareholders and boards really dislike it. They want AI's benefits to accumulate pronto, not lento. As they should. If we think of those benefits as a stream of productivity gains, then of course the board wants them tomorrow instead of sometime down the road — they increase the value of the firm more the sooner they show up.
The second big reason not to be casual about AI is that the depth and breadth of its benefits increase as more time, effort, money, and leadership attention are focused on the technology rollout. To drive this point home, let me fall back once again on my favorite historical analogy: factory electrification.
As they became cheaper to operate than steam engines, electric motors found their way into factories. Most factory owners eventually installed several of them, as opposed to relying on one big one in the basement. Great, they electrified and got some cost gains.
But now think about a factory management team that dared to dream, And devoted some serious bandwidth to thinking about the possibilities of this new technology. Overhead cranes, assembly lines, and conveyor belts simply aren't possible in a factory with a few big motors in the basement and a whole lot of drive shafts and power transmission belts (Think of the fan belt in your car, but a lot bigger) everywhere else. So lots of steam-era factory folk never thought of them. It took hallucinogens new thinking combined with lots of tinkering and experimentation to even envision these innovations, let alone make them work at scale.
What are the AI equivalents to overhead cranes, conveyor belts, and assembly lines? I don't know, and I think that's the really exciting part. It's not just that the far future of business use of AI is uncertain. It's that the most impactful uses a few years from now are uncertain. This technology is still a toddler; who knows how it's going to grow up?
Agentic AI is pretty clearly a thing that is going to mature and spread. But anyone who tells you that they know exactly how is either lying to you or lying to themselves. Similarly, the smart division of labor between humans and AI isn't clear at all right now. As the title of a great paper put it, there's a “jagged frontier” between the two kinds of intelligence. Put a task on the wrong side of that frontier and you'll get a suboptimal result - maybe even a lousy or catastrophic one.
For all these reasons, the Stream of AI benefits is not going to flow equally to all companies in the years ahead. Some — the energetic ones — are going to get a torrent. Others are going to get a trickle.
Once that difference becomes clear, so does the third reason not to be patient about AI's benefits: your competition might not be patient. I think there are some industries where incumbents are protected by such a dense thicket of regulations that they don't have to worry much about the arrival of a horde of well-capitalized, AI-native newcomers (looking at you, insurance and US retail banking). It will be interesting to see whether true AI Enthusiasts appear among those incumbents. If so, I bet they'll start hoovering up margin and market share.
In the rest of the economy, I predict a bunch of new AI-native geek companies are going to appear relatively soon. And then things are going to get really interesting in a not-good way for companies that have taken a casual approach to AI adoption and benefit realization.
I don’t say this just because he quoted me. It’s great for other reasons as well.
The great thing about electric motors and calculators is that their behavior is highly reliable and predictable. The sheer narrowness of function enabled this. LLM technology is much less reliable, and its breadth of function makes it impossible to test across that entire breadth. For fail-soft applications, this is not a problem. But in many business applications, failures are expensive. We are in the midst of developing a software engineering methodology for building reliable AI systems out of these unreliable components. Part of the methodology is to focus such AI systems on a narrow set of tasks so that we can carefully test them to ensure good behavior.
Of course, in the research world, we are trying to understand the causes of LLM unreliability and fix them. But it looks like that is going to take a long time. One promising direction is to combine the natural language capabilities of LLMs with the soundness and reliability of traditional AI methods (e.g., theorem provers, search techniques, etc.).
Cyber security is also an immense challenge, because the input (context) buffer mixes control and data. This is well known to be a terrible idea from a security perspective. As far as I can tell, we lack any sound method for fixing this problem.
I'm very interested in hearing how you (and your readers) think the modern corporation will change in order to best employ this new technology!
I was born in 1977, so I don’t really remember my dad’s calculator showing up. By the time I noticed, it was already just part of the desk — invisible, normal. That seems like the first phase of AI: small tools that offload drudgery so quickly they fade into the everyday.
But as with electricity in factories, the bigger leap comes later, when systems are reorganized around the new power source. For architects, that means moving past file drawers into AI-native infrastructure. Some technologies vanish into the background; others rearrange the whole floor. AI will do both.