OpenAI preps models to replace banking, consulting jobs

OpenAI’s decision to train its models on the everyday outputs of consultants and bankers is more than just a technical experiment, it’s a signal that AI is being repositioned from a generic tool to a domain-capable resource, Sanchit Vir Gogia, the chief analyst, founder and CEO of Greyhound Research said Thursday.

He was responding to reports of a project from OpenAI that is being developed for the management consulting sector. According to Bloomberg, upwards of 150 former consultants from McKinsey & Co., Bain & Co., and Boston Consulting have been contracted by a third-party to train the models on “how to do the industry’s entry-level tasks,” in a project code-named Argentum.

Last week, it also reported that OpenAI has more than 100 ex-investment bankers from organizations such as JP Morgan Chase, Morgan Stanley, and Goldman Sachs “helping train its artificial intelligence on how to build financial models as it looks to replace the hours of grunt work performed by junior bankers across the industry,” in a project code-named Mercury.

PoC purgatory common

Enterprises, said Gogia, “are no longer content with wide but shallow capability. They want depth. Real outputs. Structured thinking. Work that they can plug into core processes. Our CIO Pulse 2025 shows this shift in action, with 68% of global decision-makers seeing AI as a co-worker rather than a cost-cutter.”

“What OpenAI is doing is meeting that expectation head-on by learning from those who’ve done the work at scale,” he pointed out. “The aim isn’t to mimic language, but to internalize practice. And for enterprise buyers, that matters. What’s being offered now is not an AI that chats, but one that contributes.”

Forrester VP Principal Analyst Craig Le Clair, whose coverage areas include AI, automation, and the future of work, said, “too many AI agent implementations today are struggling to return investment. The problem stems from poor integration of agent models with business workflows. Forrester estimates that 60% of enterprises fall into a PoC [proof of concept] purgatory, with only 15% of firms generating positive and material impact.”

For example, he said, “a company could launch 30 pilots and fail to generate any ROI because all the agents just produce dense reports or insights that few act on. Forrester calls this the ‘Action Gap’ [and it] is the measurable distance between an AI-driven insight (a prediction, recommendation, or analysis) and the resulting value-driving business action.”

A large gap, he said, “means your agents are generating insights but not a return on investment. OpenAI’s announcement is a clear attempt to reduce this.”

According to Le Clair, author of the book Random Acts of Automation, the workplace is set to be disrupted by model-driven AI agents. “They are intellectual peers that are (at least in their defined domains) just as smart as us, with intelligence evolving at a faster clip. This transition to a hybrid workforce is not well understood and expensive,” he said.

He added, “if you think the biggest hurdle in deploying an AI agent is development, the API calls, or cloud consumption costs, you’d be wrong. Service companies tell us that for every dollar spent on AI agent licensing, organizations are spending nearly five dollars on services to get it running at scale, and those dollars are going to organizational changes [to enable a] shift in skills.”

Forrester, said Le Clair, “has defined an Agent Experience (AX) program that focuses on five skill categories reflecting these changes: knowledge curation; change management; critical thinking; interaction skills; and agent oversight.”

Structure is everything

Greyhound’s Gogia pointed out that even apparently smoothly running AI projects could be problematic. He said that one of his firm’s banking clients recently ran a generative AI trial to streamline its credit memo process. “The results were promising: draft documents that once took days now appeared in hours,” he noted. “But when compliance teams reviewed the outputs, they found gaps, statements made without clear source validation.”

Consequently, the project “didn’t fail, but it didn’t launch either. Instead, the institution added a review framework and put legal, risk, and tech leaders in the same room to manage AI output holistically,” he said.

He cited another case in which a consulting firm integrated AI into its analyst workflow for market sizing. “The early gains in productivity were tangible, but so were the cultural growing pains,” he noted. “[Junior staff] weren’t sure where their role ended and the machine’s began.” The firm’s response was to adapt its training model to focus not on how to use AI, but on how to supervise it.

“Across these cases, the lesson was clear: the success of AI has less to do with capability and everything to do with structure,”Gogia said.

He added that Greyhound views OpenAI’s latest moves “as the clearest signal yet that the company is anchoring itself inside the workflows that underpin real business delivery. By sourcing training inputs directly from former consultants and investment bankers, it’s doing more than chasing accuracy. It’s absorbing judgment.”