The most important question in financial markets can be put thus: will companies like JPMorgan Chase and Walmart find artificial-intelligence models useful enough to buy them at a price such that OpenAI and Anthropic can stay in business? If they do: rock ’n’ roll. If not, investors will tire of funnelling capital into the AI model-makers, thereby destroying confidence in the vast construction of data centres, and the semiconductors and gas turbines that power them, and so on into economic oblivion.
Whether or not Silicon Valley’s AI labs become white-collar advisers, they have a great deal to learn from studying them. (PEXEL)
Last year the answer was obvious, at least in Silicon Valley. The back-bending labour of office workers in the 21st century would soon go the way of the back-breaking tasks of the 20th. Which is to say it would disappear, but this time into computers rather than offshore: firms would buy tokens, not workers. The messianic confidence of tech bosses on this point has made them unpopular but not yet profitable. Partly it reflects their faith in the technology. But it also reflects a view of office work formed in the early 2020s, when tech firms hired vast numbers of paper-pushers whose emails, dashboards and coffee-chats “technical” staff despised. Since then the “bullshit jobs” view of the world has simply given way to one about “bullshit tasks”.
Yet the war on white-collar work has progressed far more slowly than techies anticipated. The information that constitutes modern companies is not easily organised and fed into AI models. The answers the models spit out are not easily acted on, either. How, for example, should a firm decide which tasks to automate? When should it use cheap open-source models rather than pricey cutting-edge ones? How should it calculate the returns on its investments? Can it hedge the cost of tokens? Answering these questions is arduous and possibly very lucrative work. Hence the idea—simple, elegant and brilliantly ironic—that model-makers should become much more involved in how their inventions are used. The bullish case works like this: companies get more out of AI; model-makers solve their existential cashflow problems; the world economy marches happily on.
To destroy the consultant, then, become the consultant. Each of the AI labs has announced partnerships with big professional-services firms. OpenAI recently bought a small British tech consultancy. In May Anthropic and OpenAI both set up joint ventures with big private-equity houses, presumably hoping that AI models could turn around many software firms languishing in their funds. That private equity cannot sell what it owns is an existential problem for the industry; if things are as bleak even after the attention of the AI labs, it will be an existential problem for everyone.
Bodies must be thrown at the problem, but how many and by whom? When big tech firms built their cloud-computing businesses in the 2010s, much of the grunt work was outsourced to Accenture, a multinational consultancy which grew enormous implementing the innovation of a handful of giants like Microsoft and Google in its client companies the world over. It has warm bodies in abundance, which it describes in the impenetrable language of the corporate central planner: 30,000 workers to be trained on Claude in “a major investment in talent, solutions, and go-to-market muscle”, it declared in December (just a small detachment of Accenture’s total corps of nearly 800,000 souls).
The market does not believe Accenture will play the role of a cash-generating middleman this time. Its shares have lost more than half of their value this year. Perhaps investors are right. That said, snobbery, as well as transaction costs, is an underappreciated aspect of what firms choose to bring in house and what they leave in the marketplace: some at the top of the AI labs think a corporate culture that prizes individual genius is incompatible with hiring many thousands of unremarkable workers to plug their products into the economy. (Banks manage to combine star traders with huge back offices, though AI types are likely to find this analogy insulting, too.)
Palantir is closer to what OpenAI and Anthropic might eventually build—and that firm’s share price has also collapsed recently at the thought. The crux of Palantir’s business is sending consultants into firms, sorting their data out and displaying it in a way that makes a middle manager feel as if he works at GCHQ. It is an impressive company, but also an impressive linguist. In its hands a firm’s data become its “ontology”. Consultants are elevated to the status of “forward-deployed engineers”. What distinguishes them from Accenture’s new “reinvention-deployed engineers” or OpenAI’s “forward-deployed experts”? Who knows, but pity any consultant who remains backward and undeployed.
Which firms strike it rich telling companies how to use AI models depends to a large extent on whether model-makers develop a better understanding of the mind-numbing nuances of modern business than incumbent consultancies do of AI. Yet so far nothing the AI firms have done indicates the slightest understanding of how ordinary firms operate. Perhaps that is not surprising for an industry whose staff pride themselves on being incompatible with the bureaucratic corporate life.
Model students
Whether or not Silicon Valley’s AI labs become white-collar advisers, they have a great deal to learn from studying them. In investment banks, they will find case studies of being hated and dealing with systemic risks: worryingly, popular mistrust is hard to dislodge and regulation generally only follows disaster. In elite law firms, they will see the extraordinary longevity of small and focused companies. And in consultancies, there are lessons on the pains of growing too fast, as the elite firms have felt in recent years, but also on the dangers of making bold predictions about the future of capitalism. That is not a warning they are likely to heed.
