Six Patterns That Worked, One That Didn't: The Enterprise Agent in 2026

01Banking: The Research Acceleration Agent at JPMorgan

Pattern: Knowledge retrieval over fragmented internal systems.

The most measurable production deployment in financial services this year is JPMorgan's LLM Suite, an orchestration layer the bank has been pushing into the workflows of its research and portfolio management teams. The outcome numbers are the headline. JPMorgan's adoption of the LLM Suite for agentic orchestration has delivered 83 percent faster research cycles for portfolio managers, automation of over 360,000 manual hours yearly, and rapid production of investment banking documents.

The architecture is the part to study. This is not a customer-facing chatbot. It is an internal agent that sits between an analyst and the bank's research corpus, internal models, and external data feeds, and answers the question "what do we already know about X" faster than a human analyst can navigate the systems where the answer lives. The pattern is identical at Bradesco in Brazil, whose Bridge agent on Microsoft Azure has produced 83 percent resolution rates for digital service and a 30 percent reduction in technology costs.

The architectural lesson for leaders is that the first agent that earns its keep inside a bank is almost never the one that talks to customers. It is the one that gives your highest-paid employees their morning back.

02Healthcare: The Outreach Agent at Hackensack Meridian

Pattern: Time-sensitive outreach where consistency outperforms intermittent humans.

Hackensack Meridian Health, a regional health system in New Jersey, deployed an agent called Erin in 2026 to handle a single high-value workflow: post-discharge follow-up. The AI agent checks how patients are feeling, whether they have scheduled follow-up appointments, whether they are taking medications, and if they need to speak with someone immediately.

The architectural choice deserves attention. Hackensack did not deploy a general-purpose patient assistant. They picked one workflow, post-discharge follow-up, that has a known clinical-outcome consequence (avoidable readmissions) and is reliably under-served by humans because it falls between shifts, between departments, and between visits. The agent does what a coordinator would do if the coordinator existed and never slept. Senior Vice President and Chief AI Officer Sameer Sethi describes the approach as "automation plus," targeting unmet needs rather than replacing existing work.

The same pattern explains why Color Health's breast cancer screening assistant works. Color Health partnered with Google to leverage agentic AI to make breast cancer screening more accessible with the Color Assistant agent, automating the first steps of breast cancer risk assessment and screening for women age 40 and older. Risk-factor branching logic that is too unwieldy for static decision trees becomes tractable when an agent can ask the next question conversationally.

Healthcare's working pattern in 2026 is the outreach agent, not the diagnostic one. Diagnostics are still humans plus tools. Outreach is becoming agents plus humans.

03Manufacturing: The Execution Agent at Siemens

Pattern: An agent that writes and validates code inside the system of record, not advice about it.

In April 2026, at Hannover Messe in Germany, Siemens commercially launched an industrial agent called Eigen. The product delivers up to 50 percent efficiency gains in automation engineering and is among the first commercially available AI systems that can plan and execute industrial automation engineering tasks. The launch forms part of Siemens' one billion euro investment in industrial AI.

The architectural distinction Siemens drew at the launch is the one to internalize. Unlike AI tools and co-pilots that generate advice, the Eigen Engineering Agent operates within real engineering systems to plan, execute, and validate tasks. It understands its projects, writes automation code, configures systems, and iterates until pre-defined performance benchmarks are met.

The agent is not generating suggestions for a human to apply. It is operating inside Siemens TIA Portal, the platform six hundred thousand industrial engineers already use, and producing validated automation programs that meet benchmarks the engineer specified up front. Pilot customers included Austrian ANDRITZ Metals, China's CASMT, and US-based Prism Systems.

The pattern matters because it answers the question every CFO eventually asks: when does an agent stop being an assistant and start being headcount. The answer is when it operates inside the system of record, with measurable acceptance criteria, in a loop. Eigen is the first commercial example most leaders have ever seen of that shape.

04Retail: The Super-Agent at Walmart

Pattern: A small set of orchestrators routing to many specialized agents.

Walmart's 2026 strategy is the most architecturally ambitious of the six, because the company is not deploying one agent. It is building an orchestration layer over many. Walmart's CTO has publicly described an agentic approach in which many specialized agents are organized under a smaller set of unified "super agents" serving distinct audiences.

The execution detail that matters is the constraint Walmart's leadership has put on each constituent agent. "Our approach to agentic AI at Walmart is surgical. Extensive early testing proved that, for us, agents work best when deployed for highly specific tasks, to produce outputs that can then be stitched together to orchestrate and solve complex workflows."

That sentence is the operating principle. One agent is narrow. Many narrow agents, composed by a few super-agents, cover the surface. The strategy is paying. Walmart has applied AI to scan over half a billion third-party marketplace listings for counterfeit and policy violations, monitor HVAC and kitchen equipment with industrial digital twins (producing a thirty percent reduction in emergency maintenance costs), and run digital twin simulations across supply chain infrastructure.

The architectural insight for leaders outside retail is that the question "should we build one big agent or many small ones" has now been answered by the largest retailer in the world. Many small ones. Composed.

05Logistics: The Perception Agent at DHL Express

Pattern: Computer vision in the user flow, replacing a manual entry step.

On May 7, 2026, DHL Express launched what is, in retrospect, an obvious deployment. DHL Express introduced AI-powered item identification for international shipping, a first in the global express logistics industry. The feature is now live across eight markets: Canada, Germany, Hong Kong, Netherlands, Singapore, South Africa, Spain, and United Arab Emirates.

The user flow is the architecture. A customer photographs the item they intend to ship using any standard smartphone. The AI system processes the image via a server-side computer vision model, classifies the object, and generates a structured, customs-compliant item description aligned with international documentation standards, all within seconds. The suggested description is then presented to the customer, who can easily review, edit, or override the entry before submitting.

This is one of the cleanest design patterns of the year. The agent does not replace the customer. It removes the cognitive load of describing an object in language that satisfies a regulator. The customer keeps the decision and the edit rights. The agent does the part that was friction.

DHL is not alone. The company is already deploying HappyRobot AI agents that autonomously handle appointment scheduling and carrier coordination across hundreds of thousands of emails and millions of voice minutes annually. The pattern, again, is to identify the highest-volume repetitive interface in the operation and put an agent across it, with a human override one click away.

06Law: The Multi-Step Reasoning Agent at A&O Shearman

Pattern: Agentic decomposition of expert workflows, with the firm's expertise as part of the stack.

In February 2026, A&O Shearman, the global law firm formed by the merger of Allen and Overy with Shearman and Sterling, announced something that would have been unthinkable in 2024. A&O Shearman and Harvey launched a series of agentic, multi-step reasoning AI agents for complex legal tasks. The initial agents focus on antitrust filing analysis, cybersecurity, fund formation, and loan review, high-value areas requiring deep legal expertise and multi-step reasoning. These AI agents incorporate expertise from A&O Shearman's leading practices into agentic systems that can conduct research and engage in multi-step reasoning over matter-specific documents and curated data.

The architectural move worth naming is decomposition. These agents break down complex issues into actionable plans, finish them interdependently, and combine intermediate outputs into complete work products with transparency and oversight. The earlier generation of legal AI was a search-and-summarize tool. The 2026 generation plans, executes, and assembles, with the firm's senior practice expertise baked in as part of the system, not as a wrapper around it.

The economic model is the second story. A&O Shearman is selling these agents externally, including to other law firms. The firm has gone from being a buyer of AI to being a vendor of it, packaged with the credibility of its name. That move is the precedent every professional services firm with deep proprietary expertise will study this year.

07Customer Service: The Replacement Agent at Klarna

Pattern: The one that didn't work. Agent-as-replacement, with the human escalation layer removed.

The canonical 2026 cautionary tale is Klarna, the Swedish buy-now-pay-later company that has spent the last eighteen months publicly unwinding what was, in 2024, the most-cited AI customer-service deployment in the industry.

The original move looked surgical. Between 2022 and 2024, Klarna eliminated approximately 700 positions, primarily in customer service and support, and replaced them with an AI assistant developed in partnership with OpenAI. At its peak, Klarna claimed that its AI systems managed two-thirds to three-quarters of all customer interactions. The CEO framed it as efficiency. Analysts framed it as the proof that AI displacement of white-collar work had arrived.

By early 2026 the framing had inverted. Customer satisfaction data had deteriorated on complex service interactions. The cost savings projected in the original announcement had not fully materialized. The company began rehiring customer service staff to handle the interactions the AI could not manage well.

The diagnostic detail is the part most leaders miss. The agent was not broken. Klarna built an AI agent optimized to close tickets fast, not to actually solve customer problems. Repeat contacts jumped 25 percent, the fully-automated model was reversed within 18 months, and the CEO had to publicly admit the strategy failed. The agent did exactly what it was instrumented to do. The instrumentation was wrong. Tickets closed is not problems solved, and a system that optimizes the former erodes the customer relationship that produces the latter.

The CEO's own framing, in his own words, was direct. "We went too far," Sebastian Siemiatkowski said, noting that the focus on efficiency and cost ultimately reduced the quality of the company's offerings and eroded trust with customers.

Klarna is not isolated. Nearly three-quarters of enterprises that deploy AI customer communications agents later roll them back or shut them down, according to new Sinch research published in May 2026. The starkest finding is the 74 percent rollback or shutdown rate for deployed AI customer communications agents tied to governance failures. Klarna is, in other words, the visible top of a very large iceberg.

The architectural lesson is the inverse of the previous six. The six wins all share one property: a human, or a clearly human-adjacent metric, remained in the loop somewhere. JPMorgan's analysts use the agent. Hackensack's clinicians get escalated to. Siemens' engineers set the acceptance criteria. Walmart's super-agents route to specialized agents that route, eventually, to associates. DHL's customer reviews and edits. A&O Shearman's lawyers supervise. In each case the agent narrowed a workflow. It did not own it.

Klarna's deployment removed the human layer entirely from the workflow the company most needed to keep it in. The replacement pattern fails for the same reason every comparable replacement pattern in industrial history failed before it. It optimizes for the visible metric and erodes the invisible one. The visible metric was support cost per ticket. The invisible one was trust. By the time the invisible metric showed up in the data, eighteen months had passed and seven hundred employment relationships had been unwound.