Every piece of broker infrastructure eventually accumulates a configuration surface too large for any one person to hold in their head: routing rules, pricing policies, symbol mappings, account-group overrides. The traditional fix is a dashboard with enough screens to cover every setting. The newer fix, now showing up across broker tooling, is a chat interface backed by a large language model that understands the current configuration and can answer questions or draft changes in plain language — an AI copilot for the infrastructure itself.
Done carelessly, this is a liability: a model with write access to live trading configuration is one ambiguous prompt away from an outage. Done well, it is a genuine step change in how fast a dealing desk can operate. The difference comes down to a handful of design decisions that have nothing to do with which model is behind the chat window.
The question to ask first: can it act, or only propose?
The single most important property of an infrastructure copilot is whether it can apply a change directly or whether every change is staged as a proposal for a human to approve. An assistant that calls a `set_execution_rule` function and commits it immediately is a different risk category from one that calls the same function in a sandboxed read of the current state, then writes a proposed diff for review. The second pattern is the only one that belongs anywhere near live trading configuration — it keeps the LLM's job strictly to drafting, and keeps the decision with a human who understands the consequences.
What a well-built proposal actually contains
A proposal worth approving shows the old value and the new value side by side for every field it touches — not a prose summary of what will change, but the literal before/after. It should be traceable back to the exact request that produced it, reviewable by someone who was not in the chat conversation, and reversible: if a published change turns out to be wrong, there should be a previous version to roll back to, not just a new chat message asking the assistant to undo itself.
- Old value → new value, for every field touched, not a paraphrase
- Tied to the original natural-language request, visible to a reviewer who wasn't in the conversation
- Versioned on publish (e.g. v114 → v115), so a bad change rolls back cleanly
- Logged permanently — every question answered and every proposal staged, approved or rejected
Why the audit log matters more than the model
Brokers evaluating these tools tend to ask which model is behind the assistant. That matters less than whether every interaction — including the questions that never produced a proposal — lands in the same audit trail as a manual dashboard change. Compliance and risk teams don't need to trust the model; they need every change, however it originated, to be exactly as traceable as if a human had clicked through the dashboard themselves. An assistant that is clever but invisible to the audit log is strictly worse than no assistant at all.
Scope it, don't open it
A copilot that reads and proposes changes to execution policy should not also have access to, say, client PII or withdrawal processing, even if the underlying model is the same one powering both. Scoping the assistant's tool access to exactly the configuration surface it is meant to help with — and nothing else — limits the blast radius of an unexpected tool call to a diff someone will review anyway, rather than to a system it was never supposed to touch.
What this looks like in BridgeQ
BridgeQ's Copilot follows exactly this model. It reads live policy, routing and pricing configuration through the same APIs the dashboard uses, answers questions and takes requests in plain language — in any language — and stages every proposed change as a diff. Nothing is applied until an admin reviews and approves it; publishing increments a version number with the approver and timestamp recorded, and every question and every proposal sits in the same audit log as a manual change. The assistant is scoped to BridgeQ's own configuration surface — it drafts, a human decides.