Shinhan AI agent platform development is entering a new phase as Shinhan Securities works with Google Cloud to connect generative AI agents directly with internal financial systems while maintaining the security controls required in South Korea’s heavily regulated financial sector.
Shinhan Securities has built a next-generation enterprise AI agent platform on Google Cloud using the Gemini Enterprise Agent Platform. The system is designed to move AI beyond isolated chatbot experiments and connect agents with internal business data, databases and financial workflows.
The project is particularly notable for its scale. Shinhan Securities says employees have already created more than 500 AI agents through an internal initiative, while the company is now preparing a closed-beta version of the integrated platform for a broader launch.
The company is also reporting measurable results from early deployment. In one over-the-counter derivatives workflow, processing a roughly 50-page unstructured contract previously required three or four employees to review the document and manually enter around 50 fields into an internal system.
With the new agent-based workflow, Shinhan says the process can be completed in about five minutes instead of roughly two hours, with employees reviewing the final output rather than manually performing every step.
That example illustrates why AI agents are attracting growing interest across financial services.
The opportunity is no longer limited to asking an AI model questions. Financial institutions increasingly want AI systems capable of participating in actual workflows — retrieving documents, analyzing information, interacting with enterprise systems and returning results to employees.
But allowing AI to operate inside a financial institution creates a much more difficult problem than deploying a public chatbot.
Accuracy, security, data governance and regulatory compliance become central parts of the architecture.
Shinhan AI Agent Platform Goes Beyond Chatbots
The first generation of enterprise generative AI was largely conversational.
Employees opened an AI interface, entered a question and received a response.
That model can improve productivity, but it remains separate from the systems where actual business operations occur.
An AI agent takes the concept further.
Instead of only producing text, an agent can potentially retrieve information, analyze documents, invoke enterprise applications and complete multiple steps toward a business objective.
Shinhan Securities is attempting to build this capability directly into its operational environment.
The company established a dedicated AX organization in 2026 to accelerate capabilities in AI and digital assets. It also selected AX coordinators across departments, encouraging employees closer to individual business processes to develop agents for their own workflows.
That employee-led approach reportedly resulted in more than 500 agents.
The challenge then became turning those individual experiments into a sustainable enterprise platform.
An isolated agent can be useful for one task.
A financial institution needs something more systematic: controlled access to data, standardized integration with business systems, security policies, monitoring and reusable infrastructure.
The Google Cloud project is intended to provide that foundation.
Gemini Enterprise Provides the AI Layer
Shinhan Securities is using Google Cloud’s Gemini Enterprise Agent Platform as part of the architecture.
Rather than allowing a large language model to control every part of a workflow, however, Shinhan designed its own hybrid orchestrator for financial operations.
This is an important architectural choice.
LLMs are powerful because they can interpret natural language, reason across unstructured information and respond flexibly to unfamiliar inputs.
Those same characteristics can create problems when a task requires completely predictable execution.
A financial institution cannot treat every process as an open-ended reasoning exercise.
Transferring money, for example, requires precise rules.
A system needs to know exactly which conditions must be satisfied and what action is allowed.
Contract analysis is different.
Legal and financial documents contain unstructured language that can benefit from LLM reasoning.
Shinhan’s architecture attempts to separate those two types of work.
Rules Handle Deterministic Financial Tasks
For operations requiring accuracy and consistency, Shinhan says agents are configured to follow predetermined rules rather than allowing an LLM to independently decide what action to take.
This approach can reduce one of the biggest risks associated with enterprise AI: using probabilistic models for deterministic processes.
Generative AI models do not behave like traditional business rules.
They generate outputs based on learned statistical patterns.
That is extremely useful for understanding text or generating explanations.
It is less appropriate when a system must execute exactly the same validated process every time.
Financial infrastructure contains many such processes.
Payments, transfers, account operations and regulatory checks can require strict validation.
Shinhan’s hybrid architecture therefore treats the LLM as one component rather than the entire decision system.
That distinction may become increasingly important as financial institutions move from AI assistants toward operational agents.
LLMs Are Called When Reasoning Is Needed
The other side of the architecture uses Gemini and similar LLM capabilities for tasks requiring interpretation.
Contract analysis is a strong example.
Financial contracts can contain dozens of pages of unstructured language.
Employees may need to identify specific clauses, extract important values and transfer information into structured internal systems.
Traditional automation struggles with this type of document because wording and formatting can vary.
LLMs can interpret more flexible language and identify information based on meaning rather than exact document structure.
Shinhan’s approach is therefore to invoke an LLM where reasoning provides genuine value rather than using generative AI indiscriminately across the entire workflow.
This can potentially improve both reliability and cost efficiency.
Large models consume computing resources.
If a deterministic rule can perform a task reliably, calling an LLM may be unnecessary.
Hybrid orchestration allows the system to reserve AI reasoning for the stages where it is useful.
Contract Processing Falls From Two Hours to Five Minutes
The clearest evidence of the platform’s potential comes from Shinhan Securities’ derivatives-contract workflow.
The previous process involved an unstructured document of around 50 pages attached to an email.
Three or four employees could be involved in reviewing the document.
After identifying the relevant information, they manually entered roughly 50 fields into an internal business system.
Shinhan says the entire process took around two hours per case.
The agent-based workflow changes that process.
Employees can invoke the agent through the business system and select the relevant email.
The system then analyzes the contract and enters the extracted information into the internal platform.
A human employee reviews the result.
Shinhan says processing time has fallen from approximately two hours to five minutes per case.
That represents a reduction of about 96% in elapsed processing time based on the figures provided by the company.
The result should still be interpreted as a company-reported example rather than an independently benchmarked performance study.
Nevertheless, it demonstrates a concrete enterprise use case for agentic AI.
Human Review Remains in the Workflow
One important detail is that employees still verify the output.
That matters considerably in financial services.
Automation does not necessarily mean removing humans entirely from a process.
Instead, AI can shift employees from repetitive data extraction toward supervision and exception handling.
In the contract example, employees no longer need to manually read every page and enter every field.
They can focus on verifying whether the extracted information is correct.
This human-in-the-loop structure can provide an additional control layer for tasks where errors could have financial or regulatory consequences.
It also illustrates a more realistic near-term model for financial AI.
Rather than fully autonomous systems replacing entire workflows, agents may initially perform much of the mechanical work while employees retain final oversight.
Shinhan Targets More Than 70% Agent Automation
Shinhan Securities wants to expand the model significantly.
Based on successful use cases such as contract processing, the company says it plans to raise agent-based automation above 70% for workflows considered suitable for automation.
That does not mean 70% of all jobs or all activities at Shinhan Securities will be automated.
The company’s statement specifically refers to tasks identified as suitable targets for agent-based workflow automation.
That distinction is important.
Financial institutions contain many activities requiring human relationships, judgment, accountability or regulatory responsibility.
But repetitive information-processing tasks can represent a substantial portion of internal operations.
Automating those processes could allow employees to spend more time on analysis, customer service and higher-value decision-making.
The success of that strategy will depend on whether additional workflows can reproduce the productivity improvements reported in the contract-processing example.
BigQuery Supports a Data Flywheel
The Shinhan AI agent platform also incorporates a data feedback architecture.
Results generated by business systems and feedback from users can be stored and updated in databases including BigQuery.
That information can then become part of the internal knowledge base available to agents during future tasks.
Shinhan describes this as a data flywheel.
The concept is important because enterprise AI quality depends heavily on organizational knowledge.
A general-purpose model may understand financial terminology, but it does not automatically know a company’s internal procedures, historical decisions or proprietary data.
An enterprise agent becomes more useful when it can retrieve relevant internal information.
The flywheel attempts to make that knowledge improve continuously as the system is used.
Completed workflows generate information.
Employees provide feedback.
That information returns to the knowledge system.
Future agents can then use the updated data.
Enterprise AI Needs Organizational Memory
This creates something closer to institutional memory for AI agents.
Traditional employees accumulate knowledge through experience.
They learn which systems contain particular information.
They understand how processes work.
They remember how unusual cases were handled.
When employees leave, some of that knowledge can disappear.
Structured enterprise AI systems can potentially preserve more operational knowledge digitally.
That does not mean AI replaces institutional expertise.
But it can make information easier to retrieve and reuse.
For financial institutions with complex processes, this can be particularly valuable.
Regulations change.
Products evolve.
Internal procedures are updated.
A useful enterprise AI platform therefore needs a mechanism for keeping its knowledge current.
The data flywheel is intended to provide that mechanism.
Security Is the Hard Part of Financial AI
Building capable AI agents is only one side of the challenge.
The other is allowing them to operate securely.
Financial institutions manage some of the most sensitive data in the economy.
Customer identities, transaction histories, investment information, contracts and internal financial records all require strict protection.
South Korea also imposes regulatory requirements around financial cloud services and network separation.
Shinhan and Google Cloud therefore had to design the platform around both AI functionality and regulatory compliance.
According to the announcement, Shinhan completed procedures including approval for innovative financial services, cloud service provider safety assessments, cloud-use reporting and security reviews.
These requirements help explain why deploying enterprise AI inside a securities company is fundamentally different from launching an ordinary consumer AI application.
Network Separation Remains Part of the Architecture
Shinhan says dedicated connections between its internal network and Google Cloud are used alongside VPC Service Controls to maintain network-separation requirements.
The company also separates areas where LLMs can be used.
This prevents a general-purpose AI model from automatically receiving unrestricted access to every part of the enterprise environment.
Data Loss Prevention technology is also applied as an additional security control.
These measures reflect a broader principle emerging in regulated AI deployments.
Security cannot simply be added after the AI system is complete.
It needs to shape the architecture from the beginning.
An AI agent with access to enterprise tools potentially has more operational power than a chatbot.
That means identity, permissions, network boundaries and data controls become essential.
Financial AI Agents Need Permission Boundaries
An ordinary AI assistant can produce an incorrect answer.
An operational AI agent could potentially perform an incorrect action.
That difference changes the risk model.
If an agent can interact with business systems, organizations need to define exactly what it is allowed to do.
Some agents may only read information.
Others may prepare actions but require employee approval.
A smaller group might be authorized to perform carefully constrained automated operations.
This resembles permission systems already used in enterprise IT.
AI does not remove the need for those controls.
It makes them more important.
Shinhan’s hybrid orchestrator reflects this philosophy by separating deterministic processes from tasks requiring LLM reasoning.
The Platform Was Built in Three Months
Speed is another notable part of the project.
Shinhan Securities says the integrated agent platform was completed in approximately three months.
According to the company, building a comparable platform using a conventional development approach could have taken at least a year.
Shinhan attributes part of the shorter development period to Google Cloud’s managed services and integrated AI environment.
Managed infrastructure can reduce the amount of time internal engineering teams spend building foundational components.
Instead, developers can focus on integration, data pipelines and the company’s own business requirements.
The claim about the one-year alternative is Shinhan’s estimate rather than an independently verified development benchmark.
Still, the three-month implementation illustrates one of the main attractions of modern cloud AI platforms.
Companies can assemble existing infrastructure components rather than developing every layer themselves.
Cloud Platforms Are Becoming AI Operating Systems
This points toward a larger transformation in cloud computing.
Cloud platforms originally became popular because companies could rent computing and storage rather than purchasing physical servers.
The next phase added managed databases, analytics and application services.
Generative AI is creating another layer.
Cloud providers increasingly offer models, agent frameworks, data infrastructure, security controls and orchestration as integrated platforms.
For enterprises, this means AI projects can increasingly be constructed from managed components.
The competitive question shifts from simply having access to a powerful LLM toward integrating that model with enterprise data and workflows.
Shinhan’s project demonstrates this transition.
Gemini provides reasoning capability.
BigQuery supports data infrastructure.
Cloud security services help control access.
Shinhan’s own orchestration layer determines how those components interact with financial systems.
Financial Services Are a Major Test for Agentic AI
If AI agents can operate successfully in finance, the implications could extend well beyond the sector.
Financial institutions combine several difficult requirements.
They process enormous amounts of data.
They contain many repetitive workflows.
They operate under strict regulation.
They require extremely high reliability.
And errors can have significant economic consequences.
This makes finance both an attractive and challenging environment for AI agents.
The potential productivity gains are substantial.
The tolerance for uncontrolled behavior is extremely low.
Hybrid architectures that combine deterministic rules, AI reasoning and human oversight may therefore become particularly important.
Shinhan’s implementation provides an example of that model.
South Korea Is Testing AI Under Financial Regulation
The project also reflects changes in South Korea’s regulatory approach to financial AI.
Shinhan Securities has previously participated in the country’s financial regulatory sandbox for cloud-based software use on internal networks. Official sandbox records list Shinhan Securities among financial institutions approved for cloud SaaS use involving Microsoft 365 and Copilot.
That broader regulatory experimentation matters because financial AI cannot expand through technology alone.
Institutions need a legal framework allowing new services to operate while protecting customers and financial stability.
South Korea’s financial sector has historically maintained strict network-separation requirements.
Cloud and generative AI services challenge some assumptions behind traditional IT architecture because they often depend on external computing environments.
The challenge for regulators and institutions is therefore to permit innovation without weakening security.
More Than 500 Agents Create a Governance Challenge
Creating 500 agents is an impressive adoption metric, but it also creates a management problem.
Who owns each agent?
What data can it access?
How is its performance evaluated?
What happens when an underlying business process changes?
How are outdated agents retired?
These questions become increasingly important as organizations move from a handful of experiments to hundreds or thousands of AI agents.
Agent proliferation without governance could recreate problems already seen in enterprise software.
Departments may develop overlapping tools.
Access permissions can become inconsistent.
Old systems may remain active long after they are useful.
A centralized platform can help impose standards.
Shinhan’s move from employee-generated agents toward an integrated enterprise architecture can therefore be understood as a governance project as much as a technology project.
AI Governance Will Become a Financial Priority
Shinhan Securities and Google Cloud executives met in Seoul in July to discuss longer-term cooperation around the financial agent ecosystem as well as AI governance and security.
Governance becomes especially important when agents gain access to operational systems.
Organizations need to know which model produced a result.
They need audit trails.
They need access controls.
They may need to explain how a decision was reached.
They need processes for responding when an agent behaves unexpectedly.
Financial institutions already have mature governance frameworks for conventional IT systems.
AI introduces additional questions because model behavior can be probabilistic.
A well-designed governance layer therefore needs to combine existing financial controls with new AI-specific monitoring.
Agents Could Reshape Financial Back Offices
Much of the immediate impact of enterprise AI may occur where customers never see it.
Back-office financial operations contain large volumes of document processing, data entry, reconciliation and reporting.
These activities are often necessary but repetitive.
AI agents can potentially combine several automation technologies into one workflow.
A document arrives by email.
The agent retrieves it.
An LLM identifies relevant information.
Rules validate the extracted values.
The system writes approved information into an enterprise application.
An employee reviews exceptions.
That is essentially what Shinhan’s contract-processing example demonstrates.
When repeated across many processes, this architecture could materially change how financial back offices operate.
Productivity Gains Depend on Workflow Design
The technology alone does not guarantee those gains.
A poorly designed workflow can simply automate an inefficient process.
Organizations therefore need to examine how work is performed before adding agents.
This is one reason Shinhan’s department-level AX coordinators are notable.
Employees working directly with business processes often understand operational bottlenecks better than a centralized technology team.
Allowing those employees to identify agent opportunities can produce more practical use cases.
The central technology organization can then provide security, data access and platform standards.
This combination of decentralized experimentation and centralized governance may become a common enterprise AI model.
The Goal Is AI Transformation, Not AI Installation
Shinhan Securities CEO Lee Sun-hoon framed the project as part of a broader transformation rather than simply another technology deployment.
According to the company, genuine AI transformation requires changing how an organization works rather than merely introducing AI tools.
That distinction is important.
Giving employees access to an AI chatbot can produce incremental productivity improvements.
Rebuilding workflows around agents can create larger changes.
But deeper integration also creates greater risk.
The more important an AI system becomes to business operations, the more carefully organizations need to manage security, reliability and governance.
The Shinhan project is therefore a useful example of the trade-off facing enterprise AI adoption.
Greater integration creates greater potential value.
It also requires stronger controls.
Google Cloud Gains a Regulated-Industry Case Study
For Google Cloud, the project provides a valuable reference case in a highly regulated industry.
Enterprise AI adoption increasingly depends on proving that cloud-based models can operate within existing security and compliance requirements.
Financial institutions are particularly important customers because they manage sensitive data and complex infrastructure.
Google Cloud Korea President Ruth Sun described Shinhan’s implementation as an example of how AI agents can be integrated into actual business systems within a regulated industry.
If similar architectures prove successful, they could be applied across banking, insurance, asset management and other financial businesses.
AI Agents Could Become Enterprise Infrastructure
The larger significance of the Shinhan AI agent platform is that agents are beginning to move from experimental tools toward enterprise infrastructure.
The first wave of generative AI asked whether employees could use AI.
The next wave asks whether AI can become part of how the organization itself operates.
Shinhan’s platform is designed around that second question.
More than 500 agents have already been created.
A centralized platform is connecting them with business systems.
A hybrid orchestrator separates deterministic execution from LLM reasoning.
BigQuery supports a continuously updated knowledge base.
Security controls maintain boundaries between internal financial systems and cloud AI services.
Employees remain involved in reviewing sensitive outputs.
These components collectively provide a more mature model of enterprise agent deployment than simply giving employees access to a chatbot.
Shinhan AI Agent Platform Sets an Ambitious Target
Shinhan’s next challenge is scale.
The company is currently preparing the platform for formal rollout after closed-beta testing with selected employees.
Its target of more than 70% agent-based automation across suitable workflows is ambitious.
Achieving that goal will require more than creating additional agents.
The company will need to demonstrate consistent accuracy, secure integration, effective governance and measurable productivity gains across many different processes.
The contract-processing example provides a promising early result: a workflow reportedly reduced from around two hours to five minutes.
But enterprise transformation will ultimately be measured across the organization rather than through a single use case.
Still, Shinhan Securities’ approach illustrates an important direction for financial AI.
The future may not be a single all-powerful financial chatbot.
Instead, financial institutions may operate ecosystems containing hundreds of specialized agents, each with carefully defined permissions, connected to shared enterprise data and coordinated through secure orchestration layers.
For South Korea’s financial industry, Shinhan Securities and Google Cloud are now testing what that architecture looks like in practice.
And if the model scales successfully, the most significant impact of generative AI in finance may occur not on a customer-facing chat screen, but deep inside the workflows that keep financial institutions operating every day.
