The Copilot era proved AI could draft text and answer questions. The Agent era is about something harder — and more valuable: AI that decides, acts, and delivers outcomes inside real business workflows. This keynote walked through what that shift demands of enterprise architecture, organisational design, and trust.
What I argued
For most enterprises, the first wave of Generative AI delivered productivity gains at the individual level — a faster email, a better draft, a quicker summary. The next wave is workflow-level automation, where agentic systems own end-to-end outcomes: a claim resolved, a customer onboarded, a policy underwritten — without human-in-the-loop on every step.
That shift sounds incremental but it isn't. It changes what we build, how we govern it, and who owns the result.
“Copilots augment a person. Agents replace a process. Most enterprises haven't reckoned with what that second statement actually means.”
The four moves that matter
- Architecture: moving from monolithic LLM calls to multi-agent systems with clear roles, MCP-based interfaces, and observability built in from day one — not bolted on after an incident.
- Trust: agentic systems make decisions; that demands new governance frameworks for explainability, auditability, and graceful failure modes that don't exist in most enterprise risk registers today.
- Talent: the skill mix shifts from "ML engineer who can fine-tune a model" to "system designer who can compose agents, evaluate trajectories, and reason about reliability under distribution shift."
- Adoption: the bottleneck isn't model capability anymore — it's integration debt, change management, and the unglamorous work of meeting business processes where they actually live.
What it means for BFSI
Insurance and banking are uniquely positioned for agentic AI because the workflows are well-defined, the documentation is rich, and the cost of mistakes is measurable. Claims triage, fraud investigation, KYC verification, underwriting support — each of these is a sequence of decisions an agent can own, with the right guardrails. The opportunity is real, but so is the risk: regulators won't accept "the model said so" as an audit response.
We're building toward a world where an agent can handle 80% of a claim end-to-end, route the genuinely complex 20% to a human with the full context, and learn from the resolution. Getting there is less about the model and more about the operating model around it.