Sebastien Bietho
  A thesis on AI in financial services

Operating Model Zero

Most AI in financial services is incremental — a copilot bolted onto one task, a saving booked, the operating model left untouched. Operating Model Zero starts from a blank page: design the function around agents from day one, and what emerges isn't the old model run faster or leaner. It's a different model.

The argument
Incremental

Take today's middle office. Add a copilot to one step. Book the saving. The org chart, the controls, the cost base — unchanged. The existing model is slightly faster and slightly more efficient. The shape is the same.

Zero-based

Rebuild from zero. Redraw the function with agents native to the design, not bolted on after the fact: how many people, in what roles, with what controls. Incremental automation is the easy yes — it is not where the value is. The value is in the redesign.

The redraw

The same work, a different shape.

Todaytask by task
Trade confirmation
Reconciliation
Break investigation
Counterparty outreach
Front office queries
Amendments
Settlement chase
Static / SSI
redrawn →
Zero-basedtwo layers
Judgment · Exceptions · Accountability
A thin, senior human layer. Fewer people, harder calls.
Agent layer
High-volume, deterministic work — confirmation, matching, outreach, chase — run by agents, logged end to end.

The work is redistributed. Agents absorb the high-volume, deterministic load; the human layer gets thinner and more senior — judgment, exceptions, and the decisions that carry liability. Accountability for those decisions can't be automated away. It concentrates upward, into fewer and more senior hands. Zero-based means clean-sheet, not unmanned.

EXHIBIT 01  The Middle Office

A worked example, end to end.

Operating Model Zero, built and running. A working multi-agent system that reconciles FX trades end to end: it reads the counterparty's confirmation, matches it against the book, detects the break, queries the front office, chases the counterparty for an amendment, and validates the fix — leaving the humans to the exceptions and the calls that carry accountability.

ORCHESTRATORdeterministic state machine — routes, never guesses
EMAIL PARSERreads the counterparty VCON into structured economics
RECONCILERmatches economics, flags the break
EXCEPTION HANDLERworks the break to resolution — in the loop with traders, sales, and the counterparty's middle office

Built on a deterministic backbone: the agents do judgment, but the arithmetic and the audit trail are not left to an AI model. Every action is recorded — a control function you can't audit isn't a control.

Who

Sebastien Bietho. Fifteen years leading large-scale transformation inside regulated financial institutions, on a securities-finance trading background. Currently consulting on regulatory-reporting remediation for a Tier-1 Australian institution — and developing a clear, evidenced point of view on what AI actually changes in a financial institution's operating model.