a Chatham House summary

Executive Summary

The fireside sessions framed both the immediate and long-term imperatives for AI adoption. Enterprises are already capturing value through efficiency. Service desk automation, engineering acceleration, and finance process optimization all show that “anywhere there’s a person taking a request, doing some thinking, and replying — AI can do that very well.” Architecturally, the shift is toward agents orchestrated as microservices, MCP as the connective layer, and the unifying design principle to “build data platforms for agents, not people.”

The next wave of impact lies in vertical, domain-grounded agents. “Horizontal gives you coverage. Vertical agentic gives you transformation.” Early applications demonstrate the difference between generic coverage and embedded solutions that reshape business outcomes. Success requires trusted rollout, API modernization, and real-time data integration, with interoperability and registries serving as safeguards against sprawl. The future of enterprise applications will be conversational and agent-first, but “applications have to be agent-first, not agent-alone.”

Sovereignty has become non-negotiable. At the geopolitical level, “sovereignty at the national level is very serious — it’s a 10 out of 10. No country is willing to take risks.” Global technology stacks are diverging, and over-constraining supply chains risks accelerating alternatives. Enterprises must balance trust, transparency, and autonomy in their technology choices, recognizing that even local infrastructure may carry hidden dependencies through software updates. Sovereignty is also shaping industrial policy, with nations requiring reciprocal investment in local compute and AI infrastructure.

Looking to the scientific frontier, compute is becoming a tool for discovery — “telescopes and microscopes for discovery.” AI is advancing drug development and operational automation, while quantum computing opens possibilities beyond classical limits, from molecular simulation to new material design. Science demands formal verification, caution with synthetic data, and openness to breakthroughs arriving from unexpected directions. The trajectory points toward specialized compute — quantum, optical, thermal — combined with agents and reasoning systems, moving from machines that only answer questions to machines that also ask them.

Key Takeaways

1. Efficiency and Architecture

AI is already reshaping cost structures and productivity at enterprise scale, and the underlying pattern is clear: “anywhere there’s a person taking a request, doing some thinking, and replying — AI can do that very well.”

Three domains of impact stand out. Service desks and ticketing are being transformed through automated triage and resolution. Engineering productivity is rising as teams embed lightweight, purpose-built agents into daily workflows. In finance and back-office operations (e.g., bank reconciliation, revenue recognition, revenue variance analysis) are now handled first-pass by AI agents, cutting cycle times in half while improving accuracy.

The organizational implications are equally significant. Team composition is shifting toward a barbell model: fewer product managers and middle managers, more senior independent contributors and college hires who can grow up with new agentic workflows. The net effect is acceleration, not cost cutting — but with a very different mix of talent.

Architecturally, a new stack is solidifying. MCP is emerging as the new API strategy, with MCP everywhere becoming the equivalent of a modern API layer. Agents are most effective when designed as orchestrated microservices, not monoliths. Reasoning models represent a step-change on par with the leap from GPT-3 to GPT-4, enabling complex multi-step workflows; reinforcement learning is overtaking fine-tuning as the performance frontier. Retrieval itself is becoming agentic, with iterative reasoning loops replacing static lookups.

The unifying design principle is clear: “Build data platforms for agents, not people.” That means abandoning search-snippet APIs in favor of full-context, multi-modal content streams, curated semantic layers, and documentation written so that agents — not humans — can discover, traverse, and act.

2. Vertical Agents and Enterprise Transformation

The enterprise AI horizon is shifting from assistants to vertical, domain-grounded agents. “Horizontal gives you coverage. Vertical agentic gives you transformation.”

Examples illustrate the step-change: 80% of telco billing inquiries resolved autonomously; end-to-end lending compressed to 30 minutes; retail banking advisors scaled one-to-one at low cost; healthcare benefits verification accelerated with high reliability.

Three enablers stand out: trusted rollout through policy engines and CICD pipelines; API modernization to unlock orchestration; and unified data that merges transactional, big, and real-time streams.

The risks of AI sprawl are clear — overlapping agents, siloed deployments, and redundant investments. Interoperability, registries, and platform-first providers are the antidote. The future of applications will be agent-first and conversational, but not agent-alone. “Applications have to be agent-first, not agent-alone.” Agent-first means the primary interface is an agent; not agent-alone means those agents must still run on the scaffolding of applications — workflows, rules, security, and integration layers — to be safe, governed, and reliable.

3. Sovereignty and Geopolitics

Sovereignty has become non-negotiable at the national level — “sovereignty at the national level is very serious — it’s a 10 out of 10. No country is willing to take risks.”

Geopolitical dynamics shape choices: U.S. and China technology stacks dominate, while Europe pursues hybrid models. Over-constraining supply chains accelerates the creation of alternatives, with Huawei as precedent. Strategic co-dependencies — where providers and customers retain mutual levers — are proving more durable than unilateral controls.

For enterprises, sovereignty is a balancing act: which cloud to trust, what controls to demand, and how much autonomy to require. Even with local data centers, dependence on software updates remains a hidden vector of control. Providers that offer transparency and shared authority will build trust; those that do not risk long-term erosion.

Sovereignty is also shaping industrial policy. The U.S. re-industrialization push is compelling both domestic and foreign companies to invest reciprocally in local compute and AI infrastructure.

4. Science, Discovery, and Specialized Compute

Beyond the enterprise horizon, AI is evolving into a scientific instrument: “telescopes and microscopes for discovery.”

The landscape is already splitting into two distinct tracks. Clinical AI is accelerating discovery — from drug development to protein folding. Operational AI is automating the “factories” of documents, data, processes, and code.

Quantum computing opens a transformative but uncertain frontier, especially for problems classical machines cannot solve. For example, industrial ammonia production consumes ~2% of global natural gas annually via the Haber–Bosch process, yet bacteria perform the same reaction at room temperature. If quantum machines could replicate biological pathways like this, the efficiency and economic gains would be measured in trillions.

Principles emerging from this frontier include: science demands correctness, not just statistics, driving formal verification; synthetic data must be handled with caution to avoid drifting from physical reality; breakthroughs often come sideways, as when protein folding was solved not by quantum but by generative AI detecting symmetries; and boards must prepare now for quantum-resistant cryptography given long remediation timelines.