Not every take on AI at RoachFest London 2026 was about scale or speed. My conversation with Alistair on the floor was refreshingly grounded: AI is a bigger tool, but humans have to remain fully in control of it, the same as with any other tool.
Efficiency Gains Are Real
Alistair did not undersell what AI actually helps with. Efficiency gains from AI-assisted research are real and substantial — AI can dramatically cut the time spent researching a complex customer query, letting a support engineer or DBA arrive at a faster, more confident answer than digging through documentation and past tickets manually would allow.
Where Human-in-the-Loop Actually Belongs
The more useful part of the conversation was about where the human touchpoint needs to sit, and why “human in the loop” cannot mean a human rubber-stamping everything an agent produces. Agentic workflows can go off and research a problem independently — but decision points still need a human to step in and say, plainly, “OK, which way do we go?” The value of HITL is not oversight for its own sake; it is putting the human exactly at the point where the decision is genuinely ambiguous and the cost of guessing wrong is high.
A Concrete Example: MySQL to CockroachDB
The example that made the abstraction concrete: AI-assisted database migration from MySQL to CockroachDB. An agent can auto-convert schemas and data reliably for the overwhelming majority of a database — but it should pause at ambiguous type decisions rather than guessing silently. Should a given column become a UUID or stay a string? Should a flag column become a proper boolean or keep its legacy tinyint representation? These are exactly the decisions where a wrong automated guess does not fail loudly at migration time — it fails quietly, months later, when an application makes an assumption about a type the agent picked without asking.
Designing the Touchpoints, Not Distrusting the Tool
Alistair’s framing is worth holding onto specifically because it avoids both failure modes I see most often in this debate: uncritical automation that trusts an agent’s guess on every ambiguous decision, and reflexive distrust that keeps a human reviewing every trivial step an agent could safely handle alone. The right design question is not “how much do we trust AI” — it is “where exactly does this workflow need a human touchpoint, and where does adding one just slow things down for no safety benefit.” That is a database migration architecture decision as much as it is an AI governance one, and it generalizes well past migrations to any agent workflow with irreversible or hard-to-detect failure modes.
Related Reading
- RoachFest London 2026: Distributed SQL Meets AI Resilience
- Production Guardrails for AI Agents
- Inside Cockroach Labs’ AI Playbook: The Database Reckoning
About the Author
I am Luca Berton, AI and Cloud Advisor. I work at the intersection of distributed systems, platform engineering, and enterprise AI deployments. Book a consultation.
