9 min readEscalation and autonomy
99 Percent Is a Failing Grade
Ten steps at 90 percent complete 35 percent of the time. Chained reliability is arithmetic, and the strongest lever is a deterministic step that never degrades it.
ReadThoughts · Agentic TOM
The questions banks are not yet asking.
Essays on agentic AI, operating model transformation, and the structural shifts arriving faster than most institutions assume.
New here?
Three essays to start with
01The concept
What Is an Agentic Target Operating Model?
Three layers of workforce, escalation designed in, outcomes owned end-to-end.
02The pressure
The Tsunami Is Not Coming from Inside the Bank
When every client sends an agent, volume multiplies and revenue does not.
03The habits
The Six Percent Club
What the few companies with a measurable EBIT effect do differently.
By theme
Four threads run through the archive
The model
What an agentic operating model is, and who manages the agents.
Escalation and autonomy
How decisions return to humans, and how autonomy is earned.
The demand side
What happens when clients send agents.
In practice
The model applied to lending, compliance, software quality and the core.
All essays, newest first
4 min readThe model
The Six Percent Club
McKinsey puts the share of companies with a significant EBIT effect from AI at six percent. Three habits separate them: breadth, iteration, the discipline to stop.
Read5 min readIn practice
Your Software Will Be Tested at Machine Speed
Time from disclosure to exploitation fell from 2.3 years to three hours. Testing is the next function that must become agentic. In a bank, it is a control.
Read4 min readIn practice
Eighty-Four Reports a Day
MROS receives 84 suspicious activity reports a working day, four times the 2020 volume. Compliance is the first function where the case for agents is arithmetic.
Read3 min readThe demand side
The Complaint Machine
Clients rarely complain because the effort exceeds the payoff. AI agents remove the friction: every suitability gap surfaced as a structured complaint.
Read3 min readThe demand side
When Every Client Negotiates Like an Institution
AI agents give every client the information position of an institution. When fees can be benchmarked in real time, margins compress from the demand side.
Read3 min readThe demand side
The Family Office for Everyone
The family office model was always gated by wealth. AI agents remove the gate. What happens to wealth management when every client gets institutional attention?
Read4 min readIn practice
The Easy Part Is the AI
Calling a model is a commodity. Closing the loop into the core banking system, with audit and control intact, is where the real work sits.
Read2 min readThe demand side
The Window Narrowed This Morning
Every bank COO has a planning horizon of 18 to 36 months. The day Anthropic released Claude Managed Agents, that horizon got shorter.
Read6 min readThe model
Who Manages the Agents?
Who is responsible for a workforce that is not human? Neither IT nor HR can own it alone. It is a design decision most banks have not yet made.
Read4 min readEscalation and autonomy
From ABS to Autopilot: The Five Levels of Banking Autonomy
Banking autonomy arrives in levels, each earned with evidence. Most institutions sit between Level 1 and 2, mistaking early automation for readiness.
Read7 min readIn practice
Case Sketch: Mortgage Lending in an Agentic TOM
Mortgage origination redesigned across bots, agents and humans, grounded in what FINMA actually requires. Same rules, different cost structure, different speed.
Read5 min readThe demand side
The Tsunami Is Not Coming from Inside the Bank
When every mortgage seeker sends an AI agent, inbound volume multiplies tenfold and revenue does not. The shift that hits first comes from outside the bank.
Read1 min readThe model
The Job Description No One Has Written
Every bank deploying AI agents has skipped the most basic step in workforce management: defining what the agent is actually hired to do.
Read2 min readEscalation and autonomy
Escalation by Design: Why Human Oversight Is an Architecture Decision
The failure mode of most automation projects is not that machines make wrong decisions. It is that nobody designed when and how the decision returns to a human.
Read3 min readThe model
The Swiss Agentic Advantage
FINMA's principles-based framework, data sovereignty and relationship banking give Swiss institutions a structural position to lead agentic deployment.
Read6 min readThe model
What Is an Agentic Target Operating Model?
Most AI discussions in banking pick tools. The consequential question is how the organisation itself must change to absorb what the tools make possible.
ReadStay in touch
Questions, disagreement, or a case from your institution: find Marc on LinkedIn. New essays also appear in the RSS feed.