It’s always good to see stories based on real user experiences, and when MIT Technology Review, in partnership with Databricks, published a report called “The great acceleration: CIO perspectives on generative AI” I was naturally curious to see how what they found compared with my own chats with CIOs and others. I want to stress from the start that neither this report nor my own chats are represented as formal statistical research. The former relates conversations from a small group but draws on data from a large sampling, and my own information is drawn from non-structured chats and may involve some interpretation/interpolation. This is an important topic and a long report to analyze, so I’ll organize my comments and findings around key quotes from the report.
The obvious starting point is this, quoting an MIT Associate Professor: “I can’t think of anything that’s been more powerful since the desktop computer.” I think this is an important point both because it represents an involved insider view of generative AI and because the personal computer was the key factor in launching the last wave of empowerment-driven IT investment. If generative AI has that potential, it’s critically important, but for it to present that opportunity, I think everyone would have to agree that its capabilities have to be formidable. Can we say they are that, and can we even say what its capabilities are? It’s not easy.
When I chat with other AI proponents, I’m struck by a seeming dichotomy I’ve not read about before. Is AI, particularly generative AI, valuable because of “intelligence” or “education”? That same question has been asked many times about people. Somebody is “smart” but what specifically does that imply, that they have a high IQ or have learned or read a lot? This is more important than it sounds because it relates to what AI could actually add to our lives and our work. The Internet is a collection of knowledge, so pure knowledge can’t be what generative AI delivers. So what does it deliver, then? That’s what we have to be thinking about, and what we should try to draw from documents like the one I’m referencing. AI has to be able to do something unique, to make a new business case. If it doesn’t, then whatever it does isn’t going to rival PCs, or any of our past major tech drivers.
The executive summary is disappointing in regards to this. It lists a half-dozen key findings, but they’re all more structural or philosophical than evocative of any major new missions. None of the six would provoke any disagreement from the enterprises I’ve chatted with, but none would be a surprise either. We have to look further.
In the second section of the report, “AI everywhere, all at once”, we see a figure that essentially says that generative AI is suitable for just about every application. That shouldn’t be a surprise given that “intelligence” in human form is what’s supporting those applications today, and an artificial form of something that can’t match the natural form isn’t highly credible. The quote that focuses things here is: “This sudden focus on generative AI’s power and potential represents a marked shift in how enterprise thinks about AI…”
What’s different about generative AI versus the “deep learning” that’s its functional superset? Answer: It talks to us, shows us things. Generative AI is approachable AI, something business planners can get their arms around. The third and fourth figures seem to show that for most missions, enterprises expect AI adoption to quadruple between 2022 and 2025. Again, my enterprises would agree with that, but with the qualification that they don’t necessarily mean “generative AI”. There’s a subtle shift in association in the report here, in that it talks about what generative AI has done but doesn’t associate the growth in adoption to generative AI in particular.
My enterprise contacts would put it even stronger. They believe that generative AI in its current form has yet to prove its value. It’s interesting. It can even be compelling, but it makes way too many mistakes and doesn’t seem to provide as much value in solving a company’s specific business problems as it does writing high-school term papers. The quotes from specific companies that are included in the referenced document seem to be talking about the potential of AI as they now see it, and are only now starting to realize that potential.
One thing that both the referenced document’s comments and what I hear from enterprises have in common is that they are really talking about “deep learning” and “large language models” more than about the generative AI that’s offered online. This shows that models trained from the Internet instead of from company data have limited current utility. It’s getting models to work from company data, or at least data proven relevant, that’s holding things back.
We could hope that the “Building for AI” section of the document has the response to that challenge, and it does, but with a bit more indirection or subtlety than we might want. The section opens by talking about the democratization of analytics, based on a hybrid model of the data warehouse and the data lake, that the report calls a “lakehouse.” “…what we’re finding now is that the lakehouse has the best cost performance straight off.” It “ combines the flexibility and scale of data lakes with the management and data quality of warehouses.”
This section of the document is critical, IMHO, because it makes the explicit connection between AI utility and a company’s own data. Yes, and data organization, deduplication, and validity are key to creating that utility. You can’t make AI work without all of that. Figure 5 in that section says that 72% of executives agree that “data problems are the most likely factor to jeopardize our AI/ML goals.” However, the same figure shows that same percentage believes in a multi-cloud approach.
My own user contacts tell me that quality data is critical for AI success, which is why they tend to believe that the widely tested forms of generative AI aren’t as useful. In fact, roughly that same percentage of my own contacts say that models trained on broad data sources, as most generative AI is, have little value. They need to have models trained on either their own data or on the data associated with the vertical market they’re in, meaning data from similar firms. They’ve taken a pretty big step toward the notion that useful AI is user-specific…private.
Which is the subject of the next section of the document, “Buy, build? Open, closed?” The relevant quote here is “If you care deeply about a particular problem or you’re going to build a system that is very core for your business, it’s a question of who owns your IP.” The enterprises I chat with, by a margin over 90%, believe that they would have to “own” their AI model, meaning that they would have a model customized by them, trained on their data, and tuned to their specific business needs. Broad-based public generative AI is seen as being useful to them in limited and almost cosmetic applications, but not business-critical ones.
The report correctly points out that some CIOs are taking steps to limit the use of generative AI for security/compliance reasons, but the biggest limitation that my own contacts cite by far (three quarters) is based on the generality of the models. I’ve found that online generative AI tools like Bard and ChatGPT will fail to provide even accurate “demographic” information about a quarter of the time. Yes, it’s possible to get better results if you structure your queries very carefully, but most companies are reluctant to bet their internal users would do that. This is why most enterprises say you have to treat generative AI like a junior-level worker whose output has to be checked by someone with more skill.
Let’s reinforce all that with another quote from this section: “All the large models that you can get from third-party providers are trained on data from the web. But within your organization, you have a lot of internal concepts and data that these models won’t know about.” Couple this with a quote from one of my contacts: “There is no way I’m going to load up my core business data into a public model. Who sees it there? I don’t know.”
The section “Workforce worries” addresses the risk that AI would steal jobs. The document cites an Accenture analysis that says that 40% of working hours across verticals could be automated using generative AI. In the next paragraph, the document quotes Goldman Sachs saying that two-thirds of US occupations will be “affected” by AI but doesn’t expect widespread job losses. These two quotes seem to me to argue against each other, and in truth they also argue against the validity of an AI business case.
Why do we “automate” or “empower workers with technology?” To improve productivity, which is output per unit time. Where does the benefit of this improved productivity manifest itself? In a reduced need for workers. What else could provide productivity savings, unless you assume that all sales of goods and services are constrained by the output volume. This entire section seems to be trying to say that there’s no real risk of widespread job losses, and that would argue that there is no major AI business case to be made. That, in short, it isn’t the next big thing after PCs.
What my contacts tell me is that they expect that there will be less demand for junior or unskilled people because it’s the tasks assigned to them that are most subject to replacement by AI entities. The quote of this section is: “Do I need someone to be laboriously typing six hours’ worth of code into an engine and trying to debug it for three days? I can see some huge efficiency by not doing that.” Well, that efficiency is going to come from eliminating the worker or you can’t monetize it. If you pay a senior person to sit around while AI does the grunt work, you may have reduced their workload, but you haven’t realized a benefit from having done it.
The problem here, which my contacts are quick to identify, is that you can’t develop senior people if you don’t have any junior people to work with. One CIO said “I agree that AI can resolve a lot of my challenges finding qualified entry-level people, but my pool of candidates for promotion is reduced by the same AI adoption, so I’ve created problems getting senior people instead.”
So what we seem to have here is a transformation of skills at the bottom that, when resolved, creates a void at the top. In fact, it means that the top-level people are going to see greater demands. Doctors who once read x-rays now have to teach AI to do it and monitor the results, skills that demand AI literacy when there is no such demand today. We’d need to rethink skills across the board to support an AI adoption, and deal with the way we develop qualified people toward senior roles, and accommodate a presumably growing population of people who aren’t qualified to play those roles.
Which leads us to “Risks and responsibilities”, the section before the wrap-up in the document. The section notes the security and governance risks I’ve already raised, and adds in the need for reliability, not so much in the sense of “network reliability” but in terms of the accuracy of the responses of AI and our ability to prove out the way they’re derived. This matches what I’ve heard from enterprises; the biggest problem they see with AI answers (cited by over 80% of users through July) is that “we have no idea how it came to this conclusion” and so “auditing AI results is almost as hard as coming up with results the old way.” Without, of course, the cost of AI.
There are, as the document says, a lot of things a user can do to improve AI results, but the problem is that the cost of doing them and the impact on business agility hasn’t been measured. As my enterprise comments of the previous paragraph demonstrate, all the steps needed to manage AI properly add to the cost of AI adoption and demand an every-growing benefit set to generate a reasonable ROI. All of the challenges of any new technology have the effect of making the generation of viable business cases more challenging. They also increase the difficulties in finding people qualified to address the issues the business case will have to overcome.
That leads us to the “Conclusion” and the quote: “If you were one of those people who learned how to work with computers, you had a very good career. This is a similar turning point: as long as you embrace the technology, you will benefit from it.” OK, I was one who learned how to work with computers, at a time when the enterprises trying to adopt them had to build not only applications but also middleware and even operating systems. They trained programmers themselves because formal computer science was still in the future. Yes, AI success will surely generate a demand for people with skills that are almost non-existent today, and yes, those people will be rewarded richly. But the computer revolution of the 1950s and 1960s is still a work in progress. For every step along the path between then and now, major technology innovations were driven along by major business cases.
The document, in many cases, is consistent with what I’m hearing from enterprises. It’s a problem of emphasis. What my enterprise contacts tell me is something the document doesn’t quite address, though it approaches the issue in a couple places, then slides away. It talks about AI as a technology that carries its own proof points, a technology that will be adopted and that only has to be managed. That is not true, according to enterprises I’ve talked with. Their question is way more fundamental, its one of needing proof that there’s any viable AI revolution to manage. Most of them accept that there is, that AI will change how they operate, but almost all who believe that also believe that at this time we cannot provide that essential proof because we’re still trying to address just what an AI adoption would do for us. They’d like more work on AI, but not so much on managing it but on proving it can be valuable enough to commit to. I’d love to see some reports on that, but the enterprises themselves don’t think we’re ready to even do the analysis yet. Wait, they say, a year and they’ll see what’s been done. I hope they’re pessimistic on the time frame, but I can’t be sure.
