Enterprises are realizing that renting the same general-purpose artificial intelligence model as every competitor is a losing strategy. Prompting the same closed system erases any edge, and every query hands proprietary business data to an outside AI lab. Running that model can also cost five to 10 times more than a specialized alternative, according to Oumi, a Seattle startup founded by engineers who previously worked at Google, Microsoft and Apple.
Oumi launched its Compounding AI Factory on Tuesday (Aug. 11), automating the deployment of a specialized model into production along with its continuous retraining on data from that model’s own real-world use, the company said in its announcement.
“Every company is becoming an AI company, but nearly all of them are running the exact same closed, generalized models trained on the public web, not on their own workflows, policies, or edge cases,” Oumi CEO Manos Koukoumidis said. Enterprises using Oumi to build can export the resulting model’s weights, its training data and the exact recipe used to build it.
Oumi says one customer, a top-five U.S. bank, used the platform to help modernize 100 million lines of legacy code after a general-purpose AI system failed roughly half its code-translation tests, though that figure comes from the company’s own case study rather than an independent audit.
Real Deployments Are Already Show the Payoff
Companies with the resources to build this in-house are already showing what the return looks like and where its limits are.
Morgan Stanley built an in-house tool called DevGen.AI, fine-tuned on its own decades-old codebase, to translate legacy languages like Cobol into plain English specs developers use to rewrite the code. Since its January launch, the tool has reviewed more than 9 million lines of code and saved the bank’s roughly 15,000 developers an estimated 280,000 hours, Mike Pizzi, Morgan Stanley’s global head of technology and operations, told The Wall Street Journal.
JPMorgan has taken a similar approach at greater scale. The bank has about 450 AI proofs of concept in the works, a number it expects to reach 1,000 next year, with tools like its LLM Suite employee platform and EVEE customer service assistant already in production, according to Tearsheet.
Mistral has built a business around the same idea: Its platform lets companies train systems on their own data, with a customer base past 100 companies, including HSBC and Stellantis, PYMNTS reported, and revenue exceeding $400 million.
Financial Firms Are Ahead on AI, but Data Challenges Remain
Financial services firms have moved further on this shift than almost any other sector, reaching high adoption on 27 of 75 AI-supported tasks tracked across eight business functions, more than any other industry surveyed, according to PYMNTS Intelligence’s Enterprise AI Benchmark Report. The report is based on a survey of 60 tech executives at U.S. companies with at least $1 billion in revenue. New AI is broadly deployed or fully embedded in data and technology processes at 81% to 95% of firms surveyed, PYMNTS Intelligence also found.
Even in that leading sector, 30% of financial services leaders named fragmented or poor-quality data as their single biggest barrier to wider deployment, the same report found. A separate PYMNTS Intelligence survey of executives across industries found 85% describe their data as fragmented or only moderately integrated despite 99% expressing confidence their governance supports enterprise AI.
A company cannot train a specialized system on data it has not organized, and that same fragmentation makes owning a specialized AI model harder than renting one, no matter how fast tools like Oumi’s make the technical process.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
