Insights

The modernization mandate: Why healthcare, financial services, and retail cannot wait to re-architect for agentic AI

Adnan Masood, PhD, Chief AI architect, UST.

Key takeaways

1. Agentic AI breaks traditional enterprise architectures Autonomous agents don’t bolt onto human-paced, application-centric systems—they require re-architected, event-driven workflows designed for machine-speed decision-making.
2. Most AI failures are architectural, not model-related “Pilot purgatory” persists because organizations wrap agents around broken processes instead of modernizing workflows, data foundations, and governance structures.
3. Modernization is now a competitive prerequisite In healthcare, financial services, and retail, legacy architectures actively block measurable AI ROI—modernization is no longer optional transformation work but a baseline requirement.
4. Winners standardize, platformize, and govern early Enterprises seeing EBITDA impact separate control from compute, adopt agent interoperability standards (MCP, A2A), and build shared platforms where new agents are configuration—not reinvention.
5. Regulation and governance must be designed in, not bolted on Successful agentic systems treat compliance, human-in-the-loop controls, and observability as architectural inputs, enabling autonomy without sacrificing trust or safety.

DIVIDER

Agentic AI is the shift from generative models that answer prompts to autonomous software that pursues goals — perceiving its environment, reasoning through multi-step decisions, invoking tools across enterprise systems, and learning from the results without a human in every loop. Where a chatbot tells you what to do, an agent does it: files the claim, reorders the inventory, reconciles the ledger, escalates the edge case.

The enterprise implication is structural, not cosmetic. Autonomous actors don't bolt onto application-centric workflows designed for human pacing, which is why most deployments stall, while a few quietly rewrite how entire industries operate.

Gartner says 33% of enterprise applications will be natively agentic by 2028 — a 33-fold jump from 2024. The AI market has crossed $391 billion. Yet Forrester's 2025 survey of 1,400 executives found only 13% can point to a positive EBITDA impact from AI deployments, and MIT puts enterprise pilot failure at 80-95%. The industry calls it pilot purgatory.

The uncomfortable truth: the bottleneck isn't the model. Healthcare, financial services, and retail were architected for passive applications and human-paced workflows, not for autonomous software that perceives, reasons, acts, and learns across thirty-year-old siloed systems. Modernization isn't a transformation line item anymore. It's the prerequisite for staying competitive.

DIVIDER

The failures are architectural

McKinsey's review of 50+ large-scale agentic deployments is blunt: success never depends on a clever agent. It depends on whether the underlying workflow got transformed. The dominant failure mode is the elegant-agent trap — wrapping a chatbot around a broken process and calling it transformation. The cautionary file is full: UnitedHealth's algorithmic claims system with a 90% appeal-overturn rate and "the model said so" as its only defense. Replit's rogue agent that deleted a production database. Arup's $25M deepfake heist. McDonald's drive-thru AI that added butter and ketchup to ice cream. Each one is an architectural failure dressed up as an AI failure.

Three architectural moves separate the 13% from everyone else:

Separate control plane from compute plane. Governance lives centrally; execution happens wherever the data lives. 94% of enterprises are multi-cloud and 79% are reverting some services to on-prem. Agents hard-coded to one hyperscaler can't move.

Standardize on MCP and A2A. Anthropic's Model Context Protocol and Google's Agent-to-Agent protocol killed the bespoke-connector era in late 2024, the TCP/IP moment for agents. A legal agent on AWS Bedrock can now delegate to a financial agent on Google Vertex AI.

Build a shared platform. If the tenth agent is mostly configuration, you have a platform. If it costs as much as the first, you have a craft project. Tata Steel deployed 300 specialized agents across global operations in nine months on that model.

DIVIDER

Retail: Real-time or nothing

Retail runs on velocity and thin margins, and end-to-end agentic workflows are delivering 5-15% revenue growth and up to 30% cost reduction. The winning pattern isn't customer chatbots; it's autonomous supply chain. Walmart's agents are wired into shelf sensors and computer vision cameras; when stock drops below dynamic thresholds, the system bypasses human procurement and triggers restocking. During a recent heatwave, a grocery chain launched targeted promotions on water and sunscreen within 90 minutes.

The emerging reference is the AAIPS framework — specialized Inventory Monitoring, Demand Forecasting, Reorder Decision, Negotiation, and Trend Discovery agents operating under a stochastic optimization problem that mathematically bounds their decisions.

Requirements: event-driven architecture (batch ERP is incompatible), harmonized SKU data, MCP-compliant connectors, and hard monetary guardrails on autonomous actions.

DIVIDER

Financial services: From capital repository to intelligence platform

90% of NVIDIA's financial services survey respondents report a positive revenue impact from AI. The strategic repositioning is visible: American Express's April 2026 acquisition of Hyper embedded autonomous expense processing into its corporate offering. JPMorgan's Cash Flow Intelligence helped Domino's cut manual accounting by 90%. Autonomous fraud agents clear 100,000+ alerts in seconds versus the 30-90 minutes a human analyst needs per alert.

Requirements are non-optional: FFIEC-aligned architectures, SOC 1/2 and PCI DSS controls at the infrastructure layer, attribute-based access control, and cryptographic verification for high-stakes transactions. The Arup heist proved that visual and auditory identity confirmation are dead. Microsoft Azure's reference pattern for contract workflows deploys a clause-customization agent, a regulatory-compliance agent, and a risk-assessment agent in sequence — each specialized, each audited, each handoff logged. That's what makes regulators comfortable with autonomous workflows.

DIVIDER

Healthcare payers: Rewiring the point of decision

Healthcare generates massive amounts of data and uses only 3% of it. 93% of physicians report care delays from prior authorization; 82% say patients abandon treatment because of the bureaucracy. The CMS-0057-F mandate now forces automated APIs and strict decision timelines. Oklahoma and Indiana legally require human review of adverse AI coverage decisions.

AWS Bedrock AgentCore's reference workflow compresses prior authorization from a multi-day ordeal to under ten minutes: an eligibility agent verifies coverage, a document agent extracts clinical notes from FHIR storage, and a PA agent submits the request.

Requirements: HIPAA-grade infrastructure, ABAC instead of RBAC, automated PHI sanitization at every boundary, FHIR as the interoperability anchor, explainable decisions with confidence scores, and — critically — mandatory human-in-the-loop interrupts at denial-of-care decisions, implemented via patterns like AWS's Strands "Agentic Loop Interrupts." Treat legislative pressure as a design specification, not a constraint.

DIVIDER

The multi-modal pivot and the governance spine

The deeper shift is from manual data processing to multi-modal insight generation — agents ingesting text, images, telemetry, and voice simultaneously. Tata Steel uses PaliGemma vision models to monitor factory floors. Walmart's shelf sensors drive procurement. Fraud agents fuse transaction streams with device fingerprints and behavioral biometrics in real time. This demands vector databases, knowledge graphs (GraphRAG, not just vector similarity), and durable agent memory — not yesterday's data warehouse.

None of it works without Propose-Enforce-Verify observability: the agent proposes an action, the platform enforces permission boundaries, and the system verifies the resulting state. Traditional APM returns a 200 OK for a hallucination; agentic observability catches it. LangSmith, Arize Phoenix, and Arthur AI exist because server uptime tells you nothing about whether an agent drifted from policy.

DIVIDER

The question that decides the decade

The 13% delivering real EBITDA impact didn't find better models. They made better architectural decisions — separating control from compute, standardizing on interoperability protocols, building shared platforms, and treating regulatory requirements as design inputs.

For CIOs and Chief Architects in healthcare, financial services, and retail: is your current infrastructure a launchpad for autonomous agents, or the cage they will eventually outgrow?

Organizations that act in the next 24 months will define the decade. Everyone else spends it catching up.

DIVIDER

7 FAQs

1. What is agentic AI, and how is it different from generative AI?

Agentic AI refers to autonomous software systems that can perceive context, reason through goals, take actions across enterprise systems, and learn from outcomes—without requiring constant human prompts, unlike traditional generative AI.

2. Why are so many enterprise AI pilots failing to scale?

Most failures stem from legacy architectures built for passive applications and manual workflows. Without modernized data, workflows, and governance, agents cannot operate safely or deliver business impact.

3. Why are healthcare, financial services, and retail most affected?

These industries rely on high-volume decisions, thin margins, and strict compliance. Legacy systems limit real-time action, while agentic AI demands continuous, cross-system autonomy to create value.

4. What architectural changes are required for agentic AI?

Key requirements include event-driven architectures, separation of control and execution layers, shared agent platforms, interoperable protocols (like MCP and A2A), and built-in observability and guardrails.

5. How does agentic AI improve outcomes in retail?

Retail leaders use autonomous agents for inventory monitoring, demand forecasting, and real-time supply chain decisions—driving revenue growth, faster response to market signals, and significant cost reduction.

6. How can regulated industries deploy agentic AI safely?

By embedding governance at the infrastructure layer—using attribute-based access control, cryptographic verification, explainability, full audit trails, and mandatory human-in-the-loop decision points.

7. What should CIOs and enterprise architects do now?

Assess whether current infrastructure enables, or constraints, autonomous systems. Organizations that modernize architectures in the next 24 months will lead the agentic era; others will spend years catching up.
DIVIDER

Discover how leading enterprises are re-architecting for agentic AI—before pilot purgatory becomes permanent. Talk to our experts about building an agent-ready enterprise.