DBS Agentic AI: What It Means for Enterprise Automation

TL;DR: DBS is deploying agentic AI—AI systems that can plan and execute multi-step tasks—to 350,000 corporate clients while keeping 10 million retail customers on gen-AI virtual assistants. The bank explicitly holds back on letting AI agents act autonomously, citing governance gaps. For developers building similar systems, open-model APIs offer a cost-effective way to prototype agentic workflows without locking into proprietary platforms.

What Is Agentic AI and Why Is DBS Deploying It?

Agentic AI refers to systems that don't just generate text but can reason, plan, and execute sequences of actions toward a goal—like fetching data, making decisions, and triggering follow-up processes. Unlike a chatbot that answers one question, an agent can break down a task like "reconcile these invoices" into sub-steps, call tools, and verify results.

DBS's rollout to 350,000 corporate clients marks one of the largest enterprise agentic AI deployments in banking. The move targets high-value, repetitive workflows—cash management, compliance checks, trade finance—where automation can cut hours of manual work. For retail customers, DBS keeps things simpler: gen-AI assistants that answer questions but don't act independently.

The distinction matters. Corporate clients have dedicated support teams and complex workflows, making them safer early adopters. Retail customers get conversational help but not autonomous action, reflecting the bank's stated caution about letting AI agents run without tight controls.

Why Is DBS Holding Back on Autonomous AI Agents?

DBS openly says controls lag capability—meaning the technology works, but governance, audit trails, and risk management haven't caught up. This is the central tension in agentic AI: the more autonomy you grant, the harder it is to predict, monitor, and correct behavior.

Three specific concerns drive this caution:

• Hallucination risk in actions: A wrong API call or an incorrect data interpretation can trigger real financial consequences, unlike a wrong chatbot answer that a human can ignore.
• Audit and compliance: Regulators require explainable decisions. An agent that chains 10 reasoning steps needs every step logged and reviewable.
• Liability boundaries: If an agent makes an unauthorized trade or payment, who's responsible—the bank, the client, or the model provider?

DBS's approach—deploying agentic AI to corporate clients while keeping human oversight—reflects a pragmatic middle ground. The agents assist, propose, and draft, but humans approve final actions. This "human-in-the-loop" pattern is currently the industry standard for high-stakes automation.

How Can Developers Build Agentic Workflows with Open-Model APIs?

You don't need a bank's budget to prototype agentic AI. Open-model APIs like those on TokShop let you build proof-of-concepts with pay-as-you-go pricing, using the same OpenAI-compatible SDKs you already know.

Here's a minimal agent loop in Python using the openai library:

from openai import OpenAI

client = OpenAI(base_url="https://tokshop.xyz/v1", api_key="sk-tok-...")

def run_agent(task, max_steps=5):
    messages = [{"role": "system", "content": "You are a task agent. Break down the task, suggest actions, and report results."}]
    messages.append({"role": "user", "content": task})
    
    for step in range(max_steps):
        resp = client.chat.completions.create(
            model="deepseek-v3.2",
            messages=messages,
            temperature=0.3
        )
        reply = resp.choices[0].message.content
        print(f"Step {step+1}: {reply}")
        
        # In a real agent, you'd parse this, execute tool calls, and append results
        if "[DONE]" in reply:
            return reply
        messages.append({"role": "assistant", "content": reply})
        messages.append({"role": "user", "content": "Continue or respond [DONE]."})
    
    return "Max steps reached"

run_agent("Check if Q3 revenue exceeds Q2 by 10%. If yes, draft a summary email.")

For production agentic systems, you'd add:

• Tool calling via function definitions in the API request
• State persistence across steps (databases, message history)
• Guardrails like output validation and human approval checkpoints

The key advantage of open models is cost. At TokShop's prices, DeepSeek V3.2 costs $0.42 per million input tokens and $0.63 per million output tokens. A 10-step agent conversation might consume 5,000 input and 2,000 output tokens—roughly $0.003. That's cheap enough to iterate aggressively.

What Are the Trade-offs Between Open and Proprietary Agent Models?

DBS uses a mix of proprietary and open models internally, but for developers, the choice matters for cost, control, and compliance.

Factor	Open Models (e.g., DeepSeek, GLM, Kimi)	Proprietary (e.g., GPT-4, Claude)
Cost per 1M tokens (in/out)	$0.42–$2.25 / $0.63–$11.25	$2.50–$15 / $10–$75
Data control	Self-host or API with clear logs	Vendor-hosted, data policies vary
Customization	Fine-tuning possible	Limited to prompt engineering
Ecosystem	OpenAI-compatible tools work	Native SDKs, but locked in

For agentic AI, the critical factor is token consumption. Agents burn tokens on every reasoning step, tool call, and error recovery. A single complex task might use 50,000+ tokens. At proprietary prices, that's $1–$5 per task. With open models, it's often under $0.50.

However, proprietary models sometimes show stronger reasoning on complex multi-step tasks. The honest trade-off: open models are cheaper and more controllable; proprietary models may reduce error rates on hard problems. For most agentic prototypes, open models are sufficient—and you can always escalate to a stronger model for specific sub-tasks.

How Do You Choose the Right Model for Agentic Tasks?

Model selection depends on your task's context window, reasoning complexity, and cost sensitivity. TokShop offers several open models with different strengths:

Model	Context (tokens)	Input $/1M	Output $/1M	Best For
DeepSeek V3.2	128,000	$0.42	$0.63	Cost-sensitive, high-volume tasks
GLM 4.6	200,000	$0.90	$3.30	Long documents, multi-step reasoning
Kimi K2	131,072	$0.855	$3.45	Balanced performance and cost
Qwen3 Coder	262,144	$2.25	$11.25	Code-heavy agent workflows

For agentic AI, context length matters more than raw benchmark scores. Agents accumulate conversation history, tool outputs, and intermediate reasoning. A 128K context model can handle most tasks, but if you're processing large codebases or long financial documents, the 200K–262K options prevent truncation errors.

A practical strategy: use DeepSeek V3.2 for high-volume, low-complexity steps (data extraction, formatting), and switch to GLM 4.6 or Qwen3 Coder for reasoning-heavy segments. This "model routing" cuts costs while maintaining quality.

FAQ

What exactly does "agentic AI" mean in banking?

Agentic AI in banking refers to AI systems that can independently plan and execute multi-step tasks—like reconciling accounts, drafting compliance reports, or analyzing trade documents—rather than just answering questions. DBS is deploying this to corporate clients while keeping retail customers on simpler gen-AI assistants.

How does DBS control the risks of autonomous AI agents?

DBS keeps humans in the loop for high-stakes decisions, meaning agents propose actions but don't execute them without approval. The bank also cites the need for better audit trails, explainability, and risk controls before granting full autonomy.

Can I build agentic AI without a large budget?

Yes. Open-model APIs like those on TokShop let you prototype agentic workflows for pennies per task. A 10-step agent conversation using DeepSeek V3.2 costs roughly $0.003, making experimentation affordable. Check the pricing page for current rates and docs for setup guidance.