After Agents Move Past the Trial Period in Financial Institutions, How Should the Value Account Be Calculated?

On October 28, Leiphone learned at a financial industry group interview during Tencent's Global Digital Ecosystem Summit that Hu Liming, Vice President of Tencent Cloud, used the phrase "earth-shaking" to describe the changes in the financial team over the past year. What sales teams discuss when they visit financial clients often most directly reflects what clients are concerned about: in the past it was cloud and infrastructure, but now it has almost entirely shifted to AI.

A more obvious change has occurred in the past six months. As Agents move from answering questions to executing tasks, financial institutions' attitude toward them has also shifted from small-scale trials to expanded procurement. The number of seats and Token consumption are growing rapidly, and Agents are beginning to enter more employees' daily work.

But after the scale expands, new questions also arise: How much value has the Agent actually created? Where are large amounts of Tokens being spent? And how can it move from a personal tool into real business processes?

From expanding procurement to beginning to calculate input-output, over the past six months, financial institutions' use of Agents has entered a new stage.

Financial institutions begin to expand Agent procurement

The change in financial institutions' attitudes first comes from changes in Agent capabilities.

Hu Liming recalled that last year and earlier, enterprise Agents were mainly Chatbots combined with fixed workflows. They could answer questions and also complete some simple tasks according to preset steps; once tasks became complex, fixed workflows were hard to handle.

This year, as foundation models' long-horizon task and autonomous planning capabilities improve, Agents have begun to shift from "answering questions" to "actually doing work." They no longer only generate a piece of text, but can break down tasks, call tools, and continuously handle more complex work.

The capability change quickly transmitted to the procurement side.

Tencent Cloud WorkBuddy was released in March this year, and its enterprise edition launched in June. Hu Liming said that WorkBuddy has now entered about 250 financial institutions, and leading financial institutions have basically shifted from small-scale trials to expanded procurement.

CICC's experience provides a more specific sample.

In the first half of the year, CICC began to gradually contact and trial a number of desktop agents. In the early stage of promotion, in order to mobilize employees' enthusiasm, CICC also held an internal AI innovation competition to encourage employees to develop agent applications according to their own business needs.

Xie Bisong, head of CICC's Information Technology Department, observed that employees accepted Agents faster than the company had originally expected. Among multiple products, WorkBuddy gradually became one of the products employees used more after two or three months online. Employees no longer just occasionally ask the large model questions, but have begun to treat it as a daily work entry point, letting the Agent call knowledge, process files, and even assist in completing a relatively complete task.

At present, CICC consumes 7 billion to 10 billion Tokens per day, and it is expected to reach 20 billion next year. The role of Agents at CICC has also begun to shift from an experimental tool for a few innovation teams to a productivity tool used by employees at scale.

After putting it to use, how to settle the account clearly

Agents entering financial institutions does not mean that financial institutions have completed AI transformation.

As the number of seats and Token consumption grow rapidly, some problems that were not obvious during the trial stage begin to surface. Xie Bisong said that the company is also beginning to pay attention to: with so many Tokens consumed every day, to what extent have employees improved quality and efficiency? Is there waste? Where does the waste occur?

These questions are a microcosm of financial institutions' use of Agents entering a new stage. Agents are moving out of the trial period, and financial institutions have also begun to seriously examine this investment.

Facing the company's questions about improving quality and efficiency, Xie Bisong admitted that at this stage it is difficult to precisely match AI input and output. If only short-term financial returns are considered, this account may well not add up. But in his view, AI is still a strategic investment at present and cannot be measured only by how much labor is saved or how much cost is reduced in the immediate term.

What really needs attention is whether employees have thereby formed new ways of working, and whether Agents can gradually enter the company's core business processes. CICC is adjusting its method for measuring AI value. Token consumption is still important, but it is no longer just a promotion metric where the higher the better. The company has begun to pay attention to who uses Tokens, in what scenarios they are used, what results they produce, and whether there are invalid calls and resource waste.

This is also a sign that Agents have moved from the promotion stage to the operations stage. However, for Agents to truly become productivity tools for financial institutions, spontaneous employee use alone is not enough.

Hu Liming divides the implementation of Agents within enterprises into three levels: "point, line, and plane."

"Point" is the efficiency improvement of an individual employee. For example, employees can use Agents to search for materials, summarize documents, and generate drafts. Such scenarios are the easiest to implement and are currently the most widely used part.

"Line" is the transformation of a business process. The Agent no longer just answers a question, but needs to call multiple tools and systems to continuously complete stages such as information acquisition, analysis, review, and output. At this step, AI begins to move from a personal assistant into specific business.

"Plane" means that the collaboration method of the entire organization changes. Agents in different departments, systems, and business processes can connect with one another, ultimately forming a new production system.

At present, most financial institutions are still in the stage of advancing from "point" to "line." Employees have begun to use Agents, but to make them penetrate business processes, issues of knowledge, data, and system connectivity still need to be solved.

Financial institutions often accumulate a large number of documents, policies, and business data internally, but this information is scattered across different systems, and permission boundaries are also very complex. Whether an Agent can find the correct knowledge and call data under appropriate permissions directly determines whether the answer is accurate and whether it can be used in real business.

In addition, the financial industry's requirements for security and compliance further raise the threshold for implementation. Which data can be read by Agents, which content can be processed by models, how the calling process is traced, and who is responsible for generated results all require clear rules to be established before expanding the scale of use.

Therefore, in addition to WorkBuddy, Tencent Cloud also needs to help financial institutions connect internal knowledge bases, business systems, and tools, and handle issues such as permission management, data isolation, and security auditing. What financial institutions buy is no longer just an AI entry point that can hold conversations, but an Agent system that can connect to existing organizations and business processes.

From encouraging employees to use more Tokens to beginning to ask where Tokens are spent, the logic of financial institutions' use of Agents is changing.

In Leiphone's view, this does not mean they have become cautious about AI; it precisely shows that Agents have moved from experimentation to broader use. Only when more and more people truly use them will issues such as input-output, organizational collaboration, and security compliance surface in concentrated form.

For financial institutions, the next challenge is no longer whether to use Agents, but how to turn an attempt spontaneously participated in by employees into a set of organizational capabilities that are truly measurable, manageable, and can be replicated at scale.