JPMorgan Chase controls AI spending: Banking industry shifts from "cash-burning race" to "actuarial era"?
Recently, according to relevant reports, J.P. Morgan will control AI spending and hire more technology talents. JPMorgan Chase expects the banking industry to adopt tighter internal cost controls on generative AI. Currently, the bank is also controlling its own costs, moving from an unrestricted experimental phase to strict internal budget management. Interviewees pointed out that when banks apply AI, they should consider cost indicators, such as the cost-to-income ratio of large models (the ratio of total AI investment costs to the business benefits created), the proportion of cost savings brought by large models, etc.
Recently, according to relevant reports, J.P. Morgan will control AI spending and hire more technology talents. JPMorgan Chase expects the banking industry to adopt tighter internal cost controls on generative AI. Currently, the bank is also controlling its own costs, moving from an unrestricted experimental phase to strict internal budget management. Interviewees pointed out that when banks apply AI, they should consider cost indicators, such as the cost-to-income ratio of large models (the ratio of total AI investment costs to the business benefits created), the proportion of cost savings brought by large models, etc. In June this year, the State Administration of Financial Supervision and Administration issued the "Guiding Opinions on the Safe Development and Application of Artificial Intelligence in the Banking and Insurance Industry" which mentioned that it is necessary to adhere to pragmatism and efficiency, be oriented towards improving business value, scientifically plan investment in the development and application of artificial intelligence, effectively balance costs and benefits, and promote artificial intelligence to effectively serve the high-quality development of the economy and the efficient operation of financial services. From "burning casually" to "counting the flowers" Regarding JPMorgan Chase's adjustment, Wang Runshi, a distinguished researcher at the Shanghai Finance and Development Laboratory, analyzed in an interview with reporters that the key to JPMorgan Chase's control of AI spending is to switch from "burning it casually" to "spending it as planned." One is model classification. Use lightweight models for daily tasks, use expensive models only for complex reasoning and high-value scenarios, and do not engage in "shooting mosquitoes with cannons". The second is hard budget constraints. Each department has a limit on token consumption, and approval will be required if the quota is exceeded. AI calls are no longer free lunches. The third is task diversion. For tasks such as summarization, classification, and simple questions and answers, small local models are enough, and expensive models are only used on the cutting edge. At the same time, Wang Runshi pointed out that JPMorgan Chase is expanding its recruitment of AI engineers while cutting its token budget. Low-value jobs are being replaced by AI, but people focusing on and managing AI are in short supply. This is not a contradiction, but a structural adjustment. "It can be seen that JPMorgan Chase is pushing generative AI from free experimentation to the era of strict budget management. The core change is to regard token consumption as a production cost that must be carefully calculated, rather than unlimited R&D expenditures." Fu Yifu, a special researcher at Sushang Bank, analyzed that the specific measures may be as follows: First, the bank implements a model hierarchical calling strategy. For daily low-value tasks such as document summarization and information retrieval, only economical lightweight models are used; for high-value scenarios such as transaction decision-making and risk modeling, the most expensive cutting-edge large models are allocated to control token waste from the source. Secondly, an evaluation mechanism linked to input and output should be established. Each AI project must clarify expected returns at the beginning of the project, continue to track actual returns during the operation process, decisively reduce the token quota for applications that are not effective, and optimize the call path through model orchestration to avoid having high-cost models handle simple requests. According to Cai Guiya, Deloitte China's financial services industry strategy, transaction and enterprise risk consulting partner, with the rapid iteration of model capabilities and agent technology, various banks are accelerating the development of AI applications in the direction of specialization, verticalization and scale. Faced with the subsequent increase in computing power and call costs, the banking industry is expected to show the following three trends: First, at the architectural level, hybrid collaboration of large and small models will become mainstream. By building a unified AI infrastructure, banks introduce model routing and task distribution mechanisms, and match models of different sizes according to task complexity, timeliness and risk levels, thereby reducing Token consumption and reasoning costs while ensuring effectiveness. The second is the management level. AI resources will gradually shift from “trial for all employees” to “department budget + project quota”. Token call volume will become an important indicator to measure AI resource investment and application benefits, and will gradually be incorporated into the internal cost measurement and allocation mechanism. In the future, the financial department may generate clear and quantifiable "AI bills" for each business department in the same way as it calculates cloud resource costs. The third is the application level. The prompt word project and intelligent body construction will enter the stage of refined operation. Banks not only need to promote standardization and templates of AI applications, but also set cost and operation guardrails for intelligent agents, including parameter limits on execution steps, context length, number of tool calls, etc., to achieve full-process control over agent costs while ensuring task quality and operational safety. Embed AI into business processes at the lowest cost From heavy investment to careful calculation, how should banks calculate their AI economic accounts?
Wang Runshi pointed out that the core competitiveness of bank AI is not who has the strongest model, but who can embed AI into business processes at the lowest cost. In this regard, he put forward five suggestions. First, establish a model classification system. Small models are used for 80% of daily tasks, and expensive models are only reserved for high-value scenarios such as credit approval, compliance review, and transaction analysis. This is resource allocation, not saving money. Second, set a budget limit for Token. Let each department know how much money it has burned, and any excess will be subject to approval. Without cost perception, there is no efficiency awareness. Third, don’t pursue precision in ROI assessment, start with a run. On the cost side, the Token consumption, API fees, and computing power consumption are calculated; on the revenue side, hard indicators are used first—how many man-years are saved, how many working hours are shortened, and how much the defective rate is reduced. The attribution problem can be put aside until the A/B test generates data. Fourth, small and medium-sized banks should not think about investing heavily in self-construction. Buying a vertical scenario agent worth millions can produce tens of millions of outputs - speeding up approvals, improving risk control accuracy, and increasing delivery volume. These are all proven paths. Large banks output computing power and models, while small and medium-sized banks focus on their own business scenarios and draw on their own strengths. Fifth, and the most easily overlooked - don't let AI adapt to old processes, let the processes adapt to AI. At present, most banks' AI applications are still at the shallow level of "auxiliary substitution". The fundamental reason is that the process has not changed. What really needs to be done is to reconstruct business processes around human-machine collaboration, otherwise AI will always be able to assist. "The key to AI competition in banks in the future is not who invests more, but who can release AI value at a more reasonable cost and in a safer and more controllable way." Cai Guoya pointed out that in accordance with the requirements of the "Guiding Opinions on the Safe Development and Application of Artificial Intelligence in the Banking and Insurance Industry", the first is to prioritize value and advocate AI applications guided by value rather than blind construction. Supervision encourages institutions to build their own scenario data sets for their own use, and encourages industry associations and peer platforms to jointly build and share common industry data sets in accordance with laws and regulations to further share capabilities. The second is to require the establishment of a red line for manual intervention. In capital transactions, asset evaluations, credit approvals, underwriting and claims, risk management, etc., as well as generative artificial intelligence scenario applications that are directly related to customer interests and directly affect the conclusion of financial contracts, they should be regarded as high-risk applications, and manual review links should be retained, and automatic decision-making by models is prohibited. The third is to strengthen refined management. Banks need to reduce token consumption and increase the value output per unit of AI investment through large and small model collaboration, model routing optimization, and intelligent agent governance.
Wang Runshi pointed out that the core competitiveness of bank AI is not who has the strongest model, but who can embed AI into business processes at the lowest cost. In this regard, he put forward five suggestions. First, establish a model classification system. Small models are used for 80% of daily tasks, and expensive models are only reserved for high-value scenarios such as credit approval, compliance review, and transaction analysis. This is resource allocation, not saving money. Second, set a budget limit for Token. Let each department know how much money it has burned, and any excess will be subject to approval. Without cost perception, there is no efficiency awareness. Third, don’t pursue precision in ROI assessment, start with a run. On the cost side, the Token consumption, API fees, and computing power consumption are calculated; on the revenue side, hard indicators are used first—how many man-years are saved, how many working hours are shortened, and how much the defective rate is reduced. The attribution problem can be put aside until the A/B test generates data. Fourth, small and medium-sized banks should not think about investing heavily in self-construction. Buying a vertical scenario agent worth millions can produce tens of millions of outputs - speeding up approvals, improving risk control accuracy, and increasing delivery volume. These are all proven paths. Large banks output computing power and models, while small and medium-sized banks focus on their own business scenarios and draw on their own strengths. Fifth, and the most easily overlooked - don't let AI adapt to old processes, let the processes adapt to AI. At present, most banks' AI applications are still at the shallow level of "auxiliary substitution". The fundamental reason is that the process has not changed. What really needs to be done is to reconstruct business processes around human-machine collaboration, otherwise AI will always be able to assist. "The key to AI competition in banks in the future is not who invests more, but who can release AI value at a more reasonable cost and in a safer and more controllable way." Cai Guoya pointed out that in accordance with the requirements of the "Guiding Opinions on the Safe Development and Application of Artificial Intelligence in the Banking and Insurance Industry", the first is to prioritize value and advocate AI applications guided by value rather than blind construction. Supervision encourages institutions to build their own scenario data sets for their own use, and encourages industry associations and peer platforms to jointly build and share common industry data sets in accordance with laws and regulations to further share capabilities. The second is to require the establishment of a red line for manual intervention. In capital transactions, asset evaluations, credit approvals, underwriting and claims, risk management, etc., as well as generative artificial intelligence scenario applications that are directly related to customer interests and directly affect the conclusion of financial contracts, they should be regarded as high-risk applications, and manual review links should be retained, and automatic decision-making by models is prohibited. The third is to strengthen refined management. Banks need to reduce token consumption and increase the value output per unit of AI investment through large and small model collaboration, model routing optimization, and intelligent agent governance.