358 bank agents: completion quality is the real metric

The number everyone will remember from this banking-AI story is 358. But at Real Summit 2026, Samsung SDS stated the core challenge of the deployment plainly — not “deciding what to build” but completion quality: getting agents to finish real business work, correctly and deliverable. Launching 358 agents at a bank is not a milestone; it is a stress test of whether you have the workflow redesign and measurement discipline to keep them from becoming 358 toys.

The constraint the vendor itself names

On September 8, Samsung SDS announced at Real Summit 2026 that it had deployed 358 AI agents across Woori Bank’s operations, spanning five domains — relationship management (RM), corporate credit, wealth management (WM), internal control, and customer service automation — with roughly 60 agents in RM and corporate credit and about 50 in WM (source: Gate News relay of the vendor announcement — https://www.gate.com/news/detail/018260/samsung-sds-deploys-358-ai-agents-across-woori-bank-operations-24126754 ; original coverage in Korea’s einfomax — https://news.einfomax.co.kr/news/articleView.html?idxno=4434071 ). All figures are vendor-reported, with no independent audit.

The detail worth pausing on is not 358 but how Samsung SDS frames the challenge: securing completion quality for actual business application, rather than determining what to build. Three years ago enterprises asked what agents could do; a year ago, which agents to build; now the bottleneck has moved to whether agents finish real business work, correctly. That shift in the binding constraint is the structural change.

Count answers “how many installed”; completion answers “how many done”

358 is a lagging indicator: it records how many were installed, not how many delivered. If every agent merely looks up information and offers suggestions, while a human still reviews each action and patches each output, then 358 agents and 35 agents are functionally the same — the same cognitive load spread across more surfaces. Completion quality is what separates a production agent from a pilot toy: whether the agent’s output is used directly, approved directly, and counted against a business KPI.

This is not an externally imposed standard. OOMeta runs dozens of production agents daily (the AQ task bus, the cron fleet, the social distribution engine), and we track completion rate, acceptance rate, and closure status — whether a task actually finished and whether its output was consumed. Count is dashboard decoration; completion quality is the language shared by operations and the business. In banking, where a single wrong entry is a regulatory conversation, completion quality should be the first metric: one error costs more than tokens.

The roughly 50 wealth-management agents carry the most signal

The most instructive part of the announcement is wealth management: more than 600 private bankers with uneven capability, causing lost sales opportunities. Samsung SDS deployed about 50 agents there — text-to-SQL so natural-language queries retrieve accurate data, plus consultation scripts that pull weaker performers up toward the top (source: same Gate News relay — https://www.gate.com/news/detail/018260/samsung-sds-deploys-358-ai-agents-across-woori-bank-operations-24126754 ).

This exposes a bias in most bank-transformation narratives. The default positioning of agents is “replacement”; this use case is “standardization.” The uneven capability is not in machines but in people — the spread between the best and worst of 600 private bankers is several orders of magnitude. Here agents lift the bottom above the acceptance line rather than replace the top. That is where completion quality truly lands for a buyer: the metric is not how smart the agent is, but how much the organization’s output distribution converges. Making “look up the data” available to everyone via text-to-SQL, instead of depending on a few experts, is where ROI becomes measurable.

Workflows must be redesigned for AI, not the other way around

Samsung SDS director Lee Jun-hyung gave a blunt verdict in the presentation: for enterprise AI transformation, work processes must be designed to suit AI from the outset. The direction is moving banks from “AI that asks and answers” to “AI that works and solves.” That is why completion quality is inevitably a workflow problem rather than a model problem: whether an agent finishes depends on the workflow giving it a clear input, a clear approval point, and a clear closure; drop it into an unrebuilt legacy process and it will only produce half-finished outputs.

Put this next to other banking deployments of the past two months. Citi’s Arc platform runs agents as a centralized operating system (180,000 employees using AI, over 100,000 agent-hours per week) — the same conclusion that running agents is a system capability, not a one-off project. The consensus forming is that the precondition for agentification is not a stronger model but rewired workflows. Any plan that treats agent count as the deliverable without touching workflows is mistaking the face for the substance.

Our judgment

First, agent count is a vanity metric and completion quality is the production metric — buyers should ask vendors for each agent’s completion rate, straight-through rate, and business-KPI linkage, not the deployment total. Second, a vendor naming completion quality as the core challenge confirms the industry bottleneck has moved to “getting agents to finish work”; whoever builds completion-quality measurement and workflows first moves from pilot to production first. Third, wealth management demonstrates agents as a capability-standardization tool — converging an organization’s output distribution is easier to quantify than replacing any single role. Fourth, workflow redesign for AI is a precondition, not an option; agents dropped into legacy processes only produce half-finished work.

Buyer action checklist

First, change the procurement and acceptance language from “how many agents” to “how is completion quality measured”: require the vendor to define each agent’s completion rate, straight-through rate, and business-outcome linkage. Second, pick one domain (service, wealth-advisor support, corporate due diligence) and establish a completion-quality baseline before talking about scale. Third, decide whether each agent’s role is replacement or standardization: for teams with uneven capability, use agents to converge the output distribution first — the returns are most direct. Fourth, put workflow redesign in scope: without rewired workflows, more agents is more face, not more substance.

The decision question left for you: if completion quality — not agent count — were written into the acceptance criteria of your bank’s AI program, how many of the agents in your current project would survive that test?

References: ①Gate News on Samsung SDS’s Real Summit 2026 announcement (vendor-reported, 2026-09-08) — https://www.gate.com/news/detail/018260/samsung-sds-deploys-358-ai-agents-across-woori-bank-operations-24126754 ；②Original coverage in Korea’s einfomax — https://news.einfomax.co.kr/news/articleView.html?idxno=4434071 ；③Citi Arc platform (OOMeta published article) — /en/insights/citi-arc-agent-operating-system-2026