Real-World Agentic AI

The word “agent” becomes useful only when we can point to the work being delegated, the tools being used, the human controls, and the measured result. This module examines real deployments through those four questions instead of treating every AI feature as an agent.

Business problem → agent responsibility → tools and controls → measured result

What You Will Learn

• How organizations use agents in software, support, research, security, and professional work.
• Which parts of each case are autonomous and which remain human-controlled.
• What evidence supports the reported outcome and what the evidence does not prove.
• Which architecture and safety lessons transfer to other applications.
• How to separate a useful production pattern from marketing language.

Twelve modules of principle deserve to meet reality directly, which is what this module is for. Every example here is real, sourced, and current — no hypotheticals, no invented statistics.

Anthropic’s 2026 State of AI Agents report gives a useful sense of scale before diving into specific companies. It draws on responses from more than 500 technical leaders. Among those surveyed, 57% reported using agents for multi-stage workflows, 16% for processes spanning multiple teams end to end, 86% for production coding, and 42% trusted coding agents to lead development work with human oversight. These are results from Anthropic’s survey sample—not measurements of every organization worldwide. (Anthropic, 2026 State of AI Agents)

This is no longer an experimental technology at the margins. It’s, measurably mainstream — and, as you’ll see below, uneven in how well it’s working.

Legal, tax, and audit — Thomson Reuters

The problem: legal, tax, and audit professionals spend enormous time on document review and compliance work that follows recognizable patterns but requires professional judgment at every step.

Why agentic AI: the actual sequence of what needs checking in a given document or filing depends on what’s found along the way — precisely Module 1’s original test for when a task needs an agent rather than a fixed workflow.

Approach and agent responsibilities: Thomson Reuters’ CoCounsel platform runs agentic workflows for document review and multi-step legal research; a separate product automates tax return preparation by processing source documents and prior-year filings into review-ready output. The company’s own Platform Engineering team — responsible for internal service reliability — separately built an agentic system on Amazon Bedrock AgentCore to automate what had been manual operational workflows.

Human involvement: deliberately central, not incidental. The stated goal throughout is producing “review-ready” and “audit-ready” output — a licensed professional retains final judgment on every output, precisely Module 9’s approval-gate principle applied to an entire product line’s design philosophy.

Result: Thomson Reuters reports “significant time savings through automation of repetitive tasks” and “enhanced accuracy,” combining reasoning models with their own proprietary, trusted content base.

Engineering lesson: this is what a well-scoped deployment looks like — a high-value, structured domain, with the agent handling labor-intensive extraction and drafting while a human professional retains the final call. That combination shows up repeatedly across the examples that follow. (Thomson Reuters press release, AWS Machine Learning Blog)

Enterprise automation and customer support — Salesforce Agentforce

The problem: varied across three real deployments — slow sales proposal turnaround at Uber, seasonal customer support spikes at academic publisher Wiley, and attendee-question volume at the World Economic Forum’s Davos meetings.

Approach: Uber’s Agentforce deployment turns 20-page RFP documents into structured sales opportunities and autonomously follows up with engaged leads. Wiley’s agent handles customer self-service independently during the start of each academic semester rather than only assisting a human agent. The WEF’s agent answered live logistics and content questions for roughly 3,000 attendees.

Human involvement: each case explicitly frames the agent as handling first-line volume, with complex or unresolved cases routed to humans — none described as fully replacing the underlying function.

Result: Uber reports processing RFPs 83% faster with a 3x increase in onboarding conversions. Wiley reports self-service efficiency improved over 40%, with 213% ROI on the related integration. The WEF reports a 75% first-time case resolution rate.

Limitations, stated honestly: these figures are self-reported by the vendor and customer together, worth reading with the same skepticism you’d apply to any customer success story rather than an independent audit.

Engineering lesson: every one of these deployments narrowed scope to a specific, well-defined task — RFP processing, semester support spikes, event logistics — rather than “handle all customer interactions.” Narrow scope is a real, recurring success factor across this entire module, not unique to Salesforce. (Salesforce customer stories, CX Today case study roundup)

Customer support, revisited — Klarna’s reversal

It’s worth placing a limitation directly against the successes above, because customer support is exactly the domain where the autonomy-versus-oversight trade-off from Module 9 played out publicly and expensively.

What happened: Klarna deployed an AI customer service agent in 2024 that handled 2.3 million chats in its first month — comparable to roughly 700 human agents — and paused most customer-service hiring for over a year. By 2025, the company reversed course and began rehiring. CEO Sebastian Siemiatkowski’s own words: “We focused too much on efficiency and cost. The result was lower quality, and that’s not sustainable.”

Limitations, precisely: the AI handled routine, high-volume queries well while struggling with emotionally sensitive situations and multi-step problems — exactly the harder tail of real support work where a human’s judgment still outperforms an agent’s.

Engineering lesson: Klarna measured and optimized for volume and cost, and only discovered the quality cost of that autonomy level after scaling to real, full production traffic. This is a direct, public instance of exactly the metric-selection problem Module 12 warned about — measuring what’s easy to measure instead of what matters. (Forbes, Entrepreneur)

Software engineering — mainstream adoption, and its sharpest failures

Coding is the domain where agentic AI adoption is furthest along — 86% of enterprises using coding agents for production code, per Anthropic’s own survey data — which makes it also the domain where this course’s most serious documented failures occurred, covered in full depth back in Module 10: Replit’s agent deleting a production database despite an explicit code freeze that existed only in the prompt, and Google’s Gemini CLI destroying a user’s project files after never verifying its own actions had succeeded.

Rather than re-walking those incidents here, the lesson worth restating in this module’s context is the honest juxtaposition: this is simultaneously the domain with the highest real adoption and some of the most severe publicly documented failures — proof that widespread adoption and reliability are two separate achievements, not one.

Cybersecurity — eSentire

The problem: security operations centers face an overwhelming alert volume, with real attackers moving fast once they have valid credentials — eSentire’s own threat research found active exploitation can begin within 14 minutes of credential compromise, far faster than a traditional, queued human-triage process can respond.

Why agentic AI: the correct next investigative step depends on what’s found at each prior step — querying identity logs, correlating endpoint telemetry, cross-referencing threat intelligence — exactly the dynamic, multi-step decision-making Module 1 identified as the real signal an agent is warranted.

Architecture and agent responsibilities: eSentire’s Atlas platform is explicitly a multi-agent system — “task-specific Atlas Agents” investigating threats with what the company describes as full transparency and explainable outcomes, rather than one undifferentiated agent handling everything.

Human involvement — stated with precision: eSentire calls its own operating model “Controlled Autonomy,” and its own language is worth quoting directly because it states Module 9’s exact principle in a real company’s own operational terms: “high-impact actions require full human control: shutting down production systems, wiping devices, modifying security policies.” Lower-stakes work — signal triage, investigation, correlation — proceeds autonomously; the consequential actions are gated.

Result: the company reports a mean time to engage a signal of under 30 seconds, autonomous triage of 100% of signals with no queue delay, full threat context assembled in under five minutes, and 99.99% threat containment at the initial host.

Engineering lesson: this is Module 9’s entire autonomy-gating argument, independently arrived at by a real security operations company under regulatory and insurance scrutiny — proof the principle isn’t an academic preference, it’s what a company accountable to boards, regulators, and cyber insurers built. (eSentire, Agentic AI in Cybersecurity, Help Net Security)

Enterprise productivity, and the security cost of that same convenience — Microsoft 365 Copilot

It’s worth pairing eSentire’s defensive success against the specific production system Module 11 already covered as a real, disclosed attack target, because the contrast is instructive. Microsoft 365 Copilot’s entire value proposition — reading emails and documents automatically, in the background, to save real employee time — is precisely what made it exploitable: EchoLeak (CVE-2025-32711) achieved zero-click data exfiltration through a single crafted email, no user interaction required, by defeating the same kind of automated, classifier-based guardrail eSentire uses successfully elsewhere in this module.

Microsoft’s actual remediation — a server-side patch plus defense-in-depth measures including data-loss-prevention policies and content restrictions — is covered in full in Module 11.

Engineering lesson, stated once more in this module’s context: the exact convenience that makes an agent valuable — reading content automatically so a human doesn’t have to — is frequently the same property that makes it exploitable. This tension doesn’t have a clean resolution; it’s a, ongoing trade-off every team deploying an agent against untrusted content has to actively manage, not solve once and forget.

Healthcare and life sciences — Novo Nordisk

The problem: clinical study documentation and device verification work at a major pharmaceutical company involves enormous manual effort, in a domain where errors carry real regulatory and patient-safety consequences.

Approach: using Anthropic’s Claude models through Amazon Bedrock, Novo Nordisk built a system that generates clinical documentation — work that previously required 40 to 50 people and up to 15 weeks, reduced to minutes for a team of three, according to AWS’s own case study. Separately, the company reports a 95% reduction in resources required for device verification protocols. Beyond documentation, Novo Nordisk has built what it calls a “Co-Scientist” system — a multi-agent architecture where specialized agents analyze clinical data and independently consult external medical literature, working together rather than one undifferentiated agent handling every task.

Human involvement, in the company’s own words: Novo Nordisk’s own Digital Transformation Officer, Stephanie Bova, states the underlying philosophy directly: “in a hyper-regulated environment, speed alone isn’t enough; we have to be certain our agents are operating exactly as designed.” Medical professionals continue to review and validate every generated output — explicit, structural human review, not an optional check.

Result: dramatic time reduction on documentation work, now extending from administrative use cases into actual research and drug-discovery workflows through further partnerships announced with AWS and NVIDIA.

Engineering lesson: this is a real-world instance of Module 9’s core argument that different actions within the same system warrant different autonomy levels — documentation generation runs with substantial agent autonomy, while anything touching clinical conclusions retains mandatory human validation, in a regulated industry where getting this boundary wrong carries consequences far beyond a bad customer interaction. (AWS Case Study, Celonis, How Novo Nordisk speeds up drug development)

Retail — L’Oréal

The problem: conversational analytics tools meant to help non-technical staff query business data reliably enough to trust the answers.

Result: per Anthropic’s 2026 State of AI Agents report, L’Oréal improved conversational analytics accuracy from 90% to 99.9%, scaling to roughly 44,000 monthly users producing 2.5 million messages per month.

Engineering lesson, stated honestly given the more limited detail available for this example specifically: the jump from 90% to 99.9% accuracy is the important number here, not the user count — directly reinforcing Module 12’s argument that a system generating large volume with mediocre accuracy is a worse outcome than a system generating less volume reliably. Scale without reliability isn’t success; it’s a bigger version of the same problem.

What’s still maturing — finance and research

It’s worth being honest, in the same spirit as the rest of this course, about where verified, deeply-documented individual case studies are harder to find yet, rather than manufacturing false depth.

Finance and research are both domains with real, growing agentic AI investment — the aggregate adoption statistics opening this module apply to them as much as any industry covered above — but the specific, publicly detailed, outcome-reported case studies available for those domains specifically are, as of this writing, less mature and less numerous than what exists for the seven examples above. This is worth knowing as a, current gap, not glossed over with a weaker example dressed up as a strong one.

Looking across all seven examples at once

It’s worth stepping back and noticing what these seven real companies have in common, because the pattern is more informative than any single case study alone. Thomson Reuters, eSentire, and Novo Nordisk — the three examples in this module with the clearest, most sustained success — all operate in domains with real regulatory or professional accountability: legal work, security operations accountable to insurers and boards, and pharmaceutical development under real regulatory oversight. None of that is a coincidence.

Operating under real external scrutiny appears to have forced exactly the discipline this course has argued for from Module 9 onward — explicit autonomy gating, mandatory human review at consequential points — not as an afterthought, but as a condition of being allowed to deploy at all. Klarna and the software-engineering failures from Module 10, by contrast, operated in domains with lighter external accountability, and that absence tracks directly with where this module’s clearest reliability problems occurred.

This isn’t proof that regulation causes good engineering — but it’s a real, observable pattern worth taking seriously: the discipline this entire course has taught isn’t an academic ideal external to how real companies operate. In the examples where it was present, it correlates directly with the examples that are working.

How to Judge a Claimed Success Story

Ask what task was delegated, what baseline it was compared with, how success was measured, how many cases were observed, which humans remained involved, and who published the number. A vendor case study can be useful evidence, but it is not automatically an independent experiment.

Common Misconception

Incorrect idea: A company using the word “agent” proves that its product operates autonomously.

Why it is incorrect: Product language is not an architecture diagram. Look for model-directed decisions, tool use, observations, repeated steps, controls, and measurable outcomes.

Key Takeaways

• Real agentic AI adoption is mainstream, not experimental — Anthropic’s 2026 survey found 57% of organizations using agents for multi-stage workflows and 86% deploying coding agents in production — and uneven in reliability at the same time.
• Every success in this module shares a real pattern: narrow, well-defined scope, plus explicit human review at the specific points where errors carry real consequence — Thomson Reuters, Novo Nordisk, and eSentire each independently arrived at this same structure.
• eSentire’s own “Controlled Autonomy” language — full human control required specifically for high-impact, hard-to-reverse actions — is Module 9’s exact principle, arrived at independently by a company accountable to regulators and insurers, not an academic preference.
• Klarna’s reversal is a direct, public instance of Module 12’s warning: measuring volume and cost without measuring quality can hide a real problem until it’s already expensive to fix.
• Software engineering has simultaneously the highest real adoption and some of the most severe documented failures in this entire course — widespread use and reliability are separate achievements.
• The same property that makes an agent valuable — automatically processing content so a human doesn’t have to — is frequently the same property that makes it exploitable, as Microsoft 365 Copilot’s EchoLeak incident demonstrates directly.
• Not every domain has equally mature, equally documented case studies yet — finance and research show real aggregate investment without the same depth of individually verified outcomes available for law, healthcare, cybersecurity, and customer support as of this writing.

Think Like an AI Engineer

• Compare eSentire’s “Controlled Autonomy” language against Klarna’s actual 2024 deployment. Using Module 9’s vocabulary precisely, what specific autonomy-gating decision do you think Klarna was missing that eSentire’s architecture explicitly builds in?

• Novo Nordisk’s Co-Scientist system uses specialized agents for different tasks rather than one undifferentiated agent. Based on what you learned about single vs. multi-agent trade-offs, what benefit does specialization offer in a clinical research context specifically, beyond what a single well-designed agent might provide?

• Pick any two examples from this module and design a realistic evaluation metric, following Module 12’s framework, that would have caught a real problem before it reached the scale Klarna’s did.

• This module was honest that finance and research lack the same depth of documented case studies as the other domains covered. What would you look for as a signal that a specific company’s finance-agent deployment is mature and reliable, versus early and unproven, before trusting it as a real example yourself?

Module 14 turns from single agents to what happens when several of them need to work together — a brief, conceptual bridge into multi-agent coordination, deliberately light, since the full depth of that subject is exactly where this course’s next learning path begins.