a Chatham House Summary

Executive Summary

This session underscored that agentic AI is no longer a theoretical construct but an operating reality, with enterprises deploying it across clinical trials, engineering, logistics, procurement, manufacturing, and customer engagement.

The productivity impact is material and measurable: up to 80% faster document analysis, 1.7x higher sales conversions, 5–10% reductions in call-handling times, and 20–30% improvements in manufacturing throughput. Yet the strongest lesson is clear. Broad horizontal rollouts often retrench, with sustained adoption closer to 15–20% of pilots. The most consistent results come from focused, workflow-embedded deployments directly tied to business outcomes. As one member summarized, “Horizontal gives you generic answers; vertical transforms the business.”

The central challenge has shifted from data availability to data governance. With only five percent of enterprise data structured, the emerging principle to deal with the size of unstructured data is to “leave data where it is — clean at time of use.” Enterprise architectures are being re-engineered to evolve systems of record into systems of insight, while choices between thin-core ERP models and costly platform consolidations will carry long-term consequences.

Trust and governance remain the decisive factors in adoption. In regulated domains, “if the answer is wrong, there is no reprieve,” making continuous guardrails, explainability, and auditability essential. Cybersecurity threats are escalating, with one enterprise noting “we are attacked two billion times a year — all AI-enabled.” Organizational operating models are also being reshaped, as agents increasingly resemble digital employees that must be managed as part of the workforce. Cultural change is equally critical: early horizontal deployments built a “curiosity muscle” that now enables more targeted vertical applications.

Enterprises should treat agentic AI not as a technology overlay but as a driver of operating model redesign, workforce transformation, and measurable business outcomes. Priorities should be clear data governance, explainable trust frameworks, and vertical use cases that deliver demonstrable value.

Key Takeaways

1. Use Cases Across Industries

AI adoption is manifesting in a wide range of practical applications, spanning industries and functions.

Clinical operations — accelerating trial site selection and regulatory documentation.
Engineering — documentation retrieval, code generation, and simulation of complex systems (e.g., nuclear boilers).
Predictive maintenance and supply chain optimization, particularly in asset-heavy industries.
Document analysis at massive scale, including RFPs, contracts, and regulatory filings.
Embedded AI within enterprise applications such as SAP for quality control and reporting.
Digital twins of both software development lifecycles and logistics operations, with robots operating warehouses with minimal human oversight.
Retail personalization in merchandising and recommendations, as well as customer service quality evaluation.
Regulated industry automation — SOP adherence, rewriting of standard operating procedures, and autonomous claims evaluation.
Investor relations, IT demand management, and digital twins for supply chain modeling.
Inbound email classification and workflow routing, with hand-offs between agents.
Procurement transformation through “procurement brain” and multi-agent sourcing frameworks.
Product lifecycle acceleration — design generation, engineering documentation, and code review-as-a-service.
Employee-facing agents for HR, benefits, payroll, and finance.
AI embedded in manufacturing workflows for anomaly detection, documentation updates, and supply chain predictability.
Network operations optimization, IT self-monitoring and self-healing, and customer-facing chat interfaces.

2. Horizontal and Vertical Approaches

Industry is pursuing both horizontal platforms and vertical, domain-specific solutions.

Horizontal platforms promise scale and speed, but often deliver limited value when detached from business context.
Vertical approaches embed agents within specific domains, combining industry-specific data, models, and processes. These are showing higher impact in areas such as healthcare, procurement, and finance.
The trade-off is clear: horizontals build reach, while verticals deliver depth. As one leader summarized, “Horizontal gives you generic answers; vertical transforms the business.”
Industry is converging on starting with horizontal deployments to build literacy and governance muscles, then shifting toward high-value vertical use cases where business outcomes are more tangible.

3. Data Foundations

Data is both the enabler and the constraint for AI. The challenge is not lack of data, but the ability to govern, access, and use it meaningfully.

The open question remains: “Do we bring all data together, or leave it messy and let AI sort it out?”
Only a small fraction of enterprise data is structured; the vast majority is siloed and poorly labeled. “Five percent of our data is structured; ninety-five percent is not.”
Traditional centralization into warehouses or lakes has proven costly and slow when it come to unstructured data. The shift is toward “leave data where it is — clean at time of use.”
Knowledge graphs, entitlement models, and discovery agents are emerging to map metadata, enforce access rights, and guide retrieval across petabytes of data.
Versioning problems remain acute — spreadsheets and multiple data copies create confusion over which source to trust.
Governance and ownership are as critical as technology: enterprises must balance access, security, and compliance with the risks of vendor lock-in.

4. Enterprise Architecture

AI adoption is forcing a rethink of core enterprise systems and architectures. The focus is shifting from monolithic consolidation to flexible, interoperable, real-time infrastructure.

Architecture choices made today will have long-term consequences; openness, interoperability, and clarity on agent deployment models will shape competitiveness for years.
Systems of record must evolve into systems of insight, analytics, and AI. Accuracy, metadata, entitlements, and real-time access are now essential design principles.
As one architectural principle put it: “Keep ERP super thin — general ledger — and put intelligence outside.”
Distributed models such as data mesh, knowledge graphs, and semantic layers are emerging to balance local ownership with enterprise-wide coherence.
Enterprises are pursuing two main strategies: keeping ERP thin (limited to essential functions like general ledger) with intelligence layered outside, or consolidating into a single ERP and lakehouse despite the cost. Both approaches reflect the same priority — ensuring architectures are ready for AI.
Vendor platforms create both opportunities and risks. Success requires coexistence between vendor-provided agents and custom-built ones, while avoiding dependence that locks in future choices.
Cloud migration, entitlement enforcement, and API openness remain practical hurdles.

5. Trust & Governance

Trust is the single greatest enabler — and limiter — of agentic AI adoption.

In regulated settings, accuracy cannot be compromised: “If the answer is wrong, there is no reprieve.”
Governance must evolve from one-time oversight to continuous monitoring with clear guardrails, explanations, and auditable trails.
Baseline safeguards now include toxicity detection, entitlements, and granular access controls.
Without interoperability and lifecycle management, agent sprawl risks overlap, inconsistency, and unmanaged behavior; registries and performance monitoring are becoming necessary.
Culture and trust are as critical as controls; in high-stakes domains, human-in-the-loop remains the default until systems prove reliability.
Looking ahead, several leaders aligned that “In 3–5 years, agents will operate like digital employees.”

6. Operating Models

Technology progress is outpacing organizational readiness, making operating model redesign the defining challenge.

Legacy workflows and fragmented structures often blunt the impact of AI. Without redesign, deployments risk adding complexity rather than removing it.
Many organizations are experimenting with agents at the edge — in IT, finance, or customer service — but find scaling limited by entrenched processes and accountability gaps.
Success depends on integrating agents into established workflows rather than layering them on top. Organizations that start with workflows achieve higher adoption and faster scaling.
The shift from assistive to agentic requires rethinking governance, talent, incentives, and even management models. “Agents are beginning to look like digital employees — which means we need to manage them as part of the workforce.”
Human–agent collaboration is becoming central: SOPs are being rewritten, escalation paths redefined, and productivity gains reframed as shared responsibility between people and AI.

7. Culture and Change Management

Adoption is as much cultural as it is technical. Enterprises are rethinking workforce engagement, leadership roles, and governance structures to unlock value.

“Curiosity muscle” has become an important outcome — early horizontal deployments, even when they did not deliver long-term business value, built awareness and experimentation capacity that now fuels more targeted vertical use cases.
Change accelerates when business leaders see AI not as a technical overlay, but as an enabler of new ways of working.
Governance structures are evolving: centers of excellence, cross-functional AI leadership teams, and workforce transformation programs are emerging to balance innovation with compliance and security.
Business users are beginning to code their own AI-driven automations, creating productivity benefits but also raising challenges of security, oversight, and enterprise readiness.
Organizational transformation requires more than technical enablement — it depends on leadership engagement, cultural adaptation, and workforce trust in how AI is embedded.
Traditional operating models are proving too rigid. New approaches emphasize hybrid teams where humans and agents work together, often with humans providing oversight as AI rewrites or enforces SOPs.

8. Security & Regulation

As AI adoption accelerates, enterprises face rising risks across cybersecurity, compliance, and regulatory oversight. Managing these effectively is foundational to scaling agentic AI.

Cybersecurity threats are escalating, with one enterprise noting “We are attacked two billion times a year — all AI-enabled.”
Data protection and entitlement remain central challenges. As agents access sensitive systems, granular access controls and real-time monitoring are essential.
Regulated industries highlight a divide: in some processes “accuracy must be 100%” (e.g., claims processing), while in others a “good enough” answer is acceptable (e.g., lead scoring). This distinction shapes where AI can be safely applied.
Transparency and explainability are vital for building trust. Many organizations are developing mechanisms for audit trails, monitoring, and human-in-the-loop oversight.

9. Productivity Gains

AI is already delivering measurable productivity improvements across industries, often in double-digit percentages. The impact is seen in both efficiency and speed.

Document analysis, legal review, and RFP processing are showing productivity gains of up to 80%.
Sales agents guided by AI are delivering 1.7x higher conversion rates.
Customer service has seen 5–10% reductions in average call handling time, with faster summarization and improved knowledge access.
Manufacturing and logistics are reporting 20–30% improvements in throughput and anomaly detection savings, in some cases cutting onboarding time from six months to as little as four to six weeks.
Design and engineering functions are achieving 3x acceleration in output, particularly in product design generation and code reviews.
Even incremental improvements — such as 20% reductions in days of inventory outstanding — translate into millions of dollars in financial impact.

10. Adoption Lessons

As experiments mature, enterprises are beginning to measure the real-world impact of AI deployments. Results vary widely, but clear lessons are emerging.

Enthusiasm is tempered by uneven returns. Broad horizontal rollouts that initially reached thousands of users often retrenched after business value proved elusive, with sustainable adoption closer to 15–20% of pilots.
Success appears most consistent when AI is deployed in narrow, workflow-embedded scenarios tied directly to measurable business outcomes, rather than broad platform-wide rollouts.

AI Agents & Agentic AI