11-20 更新 实时
首页/ 公司动向/ 正文

银行为何押注 AI 编排而非自主性

公司动向 来源:Analytics India Magazine 2026-11-20 19:29 0 阅读 机器翻译
银行为何押注 AI 编排而非自主性

如果说银行经营有一条前提,那就是它的计算绝不能出错。一旦出错,后果可能从客户不满,到监管机构就某项财务决策是如何得出的展开审查。

这一显而易见却又至关重要的因素,影响着银行内部任何一项技术的采用。

面对 AI 模型和智能体系统层出不穷的自主能力,如今银行在将其投入生产环境时采取的是务实的做法。

它们并不允许模型独立执行业务关键操作,而是让模型用于理解请求、选择工具并传达结果,而底层计算仍留在传统的工程化系统之中。

这是近期一场 AIM Innovator Series 网络研讨会上的关键讨论话题之一,该活动与 Neysa 合作举办。

HDFC Bank 工程、GenAI 与创新执行副总裁 Amit Sehgal,以及 Jocata 产品管理高级总监 Karthigeyan S,与 Neysa 首席营销官 Sujit Janardanan 共同探讨了 AI 在印度金融服务行业的部署。

设想一个假设场景:一位客户符合贷款资格标准,但 AI 系统却拒绝了该申请。客户找到银行,随后监管机构要求银行解释这一决定。对于一个内部推理过程难以追溯的系统所给出的结果,银行该如何作出说明?

Sehgal 承认,对于大语言模型(LLM)而言,这是一个格外棘手的问题。

“目前没有标准化的方式来解释这个模型在做什么,”他说。

HDFC Bank 没有依赖模型为自身决策提供理由,而是采用了一种限制 LLM 能独立执行哪些操作的方法。

Sehgal 表示,尽管模型在编码和计算方面的能力不断增强,该银行仍不允许在生产环境中为业务用例生成代码。

与此同时,财务计算由预先构建、经过工程化设计的函数处理,这些函数通过 Model Context Protocol(MCP)工具对外提供。

LLM 负责解读请求并判断应调用哪个工具,但实际执行计算的是底层函数。模型的作用在很大程度上仅限于编排工作流,并将结果传达给客户或内部关系经理。

“模型主要做的是编排、决策,以及判断在什么情况下该用什么,”Segal 说。

随后,模型可以将结果传达给客户或内部关系经理。

这样一来,LLM 的语言与交互能力便与底层的财务计算相分离,而后者仍处于一个受到严格管控的工程化系统之内。

然而,这种做法并不能消除所有可能的错误来源。大语言模型仍可能误解请求、调用不合适的工具,或错误地传达返回的信息。

“我们的大多数机器人都构建在一个标准化的执行框架(harness)之上,其中包含批判或验证等步骤,”Sehgal 表示。

其大多数聊天机器人采用他所描述的框架系统(harness system),其中包含批判和验证步骤。

主模型生成的输出会接受事实准确性检查,包括所呈现的数字是否正确。这些检查可能涉及由另一个模型来评估编排模型给出的回答。

“让幻觉降到零是不可能的,”Sehgal 说。

不同模型的幻觉率可能不同,但即便很小的出错概率,在银行业也可能变得举足轻重。“在银行业,我们没有任何容忍幻觉的空间,”他说。

因此,可观测性与模型的选择同样重要。银行需要监控其 AI 应用在做什么、正在调用哪些工具,以及其输出是否通过了必要的检查。

“计算是以更确定性的方式完成的,因此解释一个函数比解释一个模型更容易,”Segal 说。“模型产出的任何内容——我们总有办法对其进行验证。”

经过严格映射、工程化的函数具有可检查的既定输入、输出和逻辑。然而,模型的内部推理要重建起来困难得多。

通过将重要计算保留在这些函数内,银行可以让 AI 辅助工作流中的各个步骤更易于检查和解释。

Karthigeyan 此前曾在文莱达鲁萨兰国中央银行工作,他就银行应如何确定赋予 AI 系统的权限提出了自己的看法。

“监管机构最终担心的是它对终端客户造成了什么影响,”他说。

与推荐金融决策的应用相比,检索文档或摘要信息的应用所呈现的风险级别不同。但当智能体被允许自主行动时,风险会进一步上升。

他以防欺诈场景说明了这一点。一位客户的信用卡在五分钟内先后在斯里兰卡和英国被使用。欺诈检测系统将这一活动识别为可疑,但无法立即确定哪笔交易是合法的。

在这类情况下,系统可以先将该活动标记出来交由人工核验,然后再冻结账户。

“它会冻结这笔交易,但人工会很快进行核验,而且你可以用确定性规则来判断它是否可接受,”他说。“也可以采用机器学习模型,利用标准进行分类、过滤并减少误报。”

不过,他补充说,最终决定必须由处于回路中的人来做出。

Karthigeyan 表示,被标记为可疑的交易中有一小部分可以被转入人工审核队列,让员工在该操作最终确定之前核实相关活动。

因此,无限制的自主决策——尤其是那些不可逆转的决策——比仅仅进行观察、检索信息或提出建议供人工审核的系统更应受到严格审查。

如今,银行正在加倍投入一种深思熟虑的做法,即为 AI 模型划定边界。这涉及了解何时以及如何让 LLM 负责解读、编排与沟通,同时将财务计算保留在工程化系统之内,并对模型输出施加额外校验。

金融服务行业所强调的,不是某项技术自动化某项任务的能力,而是是否有充分的理由和灵活性让它这么做。

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Why Banks Are Betting on AI Orchestration Over Autonomy)

If there is one premise a bank operates on, it is that it cannot afford to get its calculations wrong. The consequences can range from a dissatisfied customer to regulatory scrutiny over how a financial decision was reached.

This obvious, yet crucial factor influences any technology adoption within banks. 

With the autonomous capabilities of AI models and agentic systems galore, banks today are taking a pragmatic approach to deploying them in production. 

Rather than allowing models to independently execute business-critical operations, they are being used to interpret requests, select tools and communicate results, while the underlying calculations remain within conventional engineered systems.

This was one of the key discussions at a recent AIM Innovator Series webinar, held in partnership with Neysa. 

Amit Sehgal, EVP – Engineering, GenAI and Innovations at HDFC Bank, and Karthigeyan S, Senior Director of Product Management at Jocata, joined Sujit Janardanan, Chief Marketing Officer at Neysa, for a discussion on deploying AI in India's financial services industry.

The Explainability Challenge

Consider a hypothetical scenario: a customer meets the eligibility criteria for a loan, but an AI system rejects the application. The customer approaches the bank, and a regulator subsequently asks it to explain the decision. How does the bank account for an outcome produced by a system whose internal reasoning is difficult to trace?

Sehgal acknowledged that this is a particularly difficult problem with large language models (LLMs).

“There is no standard way to explain what this model is doing,” he said.

Rather than relying on the model to justify its own decisions, HDFC Bank has adopted an approach that limits what an LLM can execute independently. 

Sehgal said that, despite models' growing capabilities in coding and calculations, the bank does not permit code generation in production for business use cases.

Financial calculations, meanwhile, are handled by pre-built, engineered functions exposed through Model Context Protocol (MCP) tools. 

The LLM interprets the request and determines which tool to invoke, but the underlying function performs the calculation. The model's role is largely restricted to orchestrating the workflow and communicating the result to the customer or an internal relationship manager.

“What models do is mainly the orchestration, decisioning, and what to use when,” said Segal. 

The model can then communicate the result to the customer or an internal relationship manager. 

This separates the language and interaction capabilities of an LLM from the underlying financial computation, which remains within a tightly controlled engineered system.

Building the Validation Layer

However, this approach does not eliminate every possible source of error. LLMs can still misinterpret a request, call an inappropriate tool, or incorrectly communicate the returned information. 

“The majority of our bots are built on a stock harness, where we have steps like critique or validation steps,” Sehgal said.

Most of its chatbots use what he described as a harness system, which includes critique and validation steps. 

Outputs generated by the primary model are checked for factual accuracy, including whether the numbers being presented are correct. These checks may involve a separate model evaluating the response produced by the orchestration model.

“It is impossible to bring hallucination to zero,” says Sehgal

Different models may have different hallucination rates, but even a small probability of error can become significant in banking. “We have no scope for hallucinations in the banking world,” he says. 

Thus, observability is as important as the choice of the model. Banks need to monitor what their AI applications are doing, which tools they are invoking, and whether their outputs pass the required checks.

“The calculations are done in a more deterministic way, so it becomes easier to explain a function rather than explain a model,” Segal said. “Anything that the model produces—we always have means to validate it.” 

A tightly mapped, engineered function has defined inputs, outputs and logic that can be examined. However, a model’s internal reasoning is considerably harder to reconstruct. 

By keeping important calculations within these functions, banks can make individual steps in an AI-assisted workflow easier to inspect and explain.

How Much Authority Can You Give AI?

Karthigeyan, who previously worked with the Brunei Darussalam Central Bank, offered a perspective on how banks should determine the authority they give AI systems.

“The regulator finally worries about what impact it is creating on the end customer,” he says.

An application that retrieves documents or summarises information presents a different level of risk from one that recommends financial decisions. But the stakes increase further when an agent is allowed to act autonomously.

He illustrated this with a fraud prevention scenario. A customer's credit card is used in Sri Lanka and then in the UK within five minutes. A fraud detection system identifies the activity as suspicious, but it cannot immediately establish which transaction is legitimate.

In such cases, the system can flag the activity for human verification before freezing the account. 

“It will freeze the transaction, but a human will quickly verify it, and you can have deterministic rules to figure out whether it is acceptable or not,” he says. “It could also take a machine learning model, use criteria to classify, filter and reduce false positives.” 

However, he adds that the final decision must be made by the human in the loop. 

Karthigeyan said a small proportion of transactions flagged as suspicious could be moved to a human review queue, allowing an employee to verify the activity before the action is finalised.

Thus, unbounded autonomous decisions, especially those that cannot be reversed, warrant greater scrutiny than systems that merely observe, retrieve information or make recommendations for a human to review.

Banks today are doubling down on a thoughtful approach to building boundaries around AI models. This involves knowing when and how to let LLMs handle interpretation, orchestration, and communication, while keeping financial calculations within engineered systems and subjecting model outputs to additional checks.

The financial services industry places the onus not on a technology's ability to automate a task, but on whether there is sufficient reason and flexibility to let it do so. 

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 Analytics India Magazine(Why Banks Are Betting on AI Orchestration Over Autonomy),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗