Framework
The agentic operating model
Your operating model was designed for a workforce that was entirely human. The agentic TOM replaces that assumption with three layers: bots that execute, agents that decide, humans who govern. Each layer is governed differently, fails differently, and answers for different decisions.
Governance layer
FINMA Guidance 08/2024: responsibility for decisions cannot be delegated to AI. Humans set every boundary below.
Task-volume shares are illustrative, for a redesigned end-to-end process. Select a layer to see how it is governed and how it fails.
Operating principles
Five principles carry the model
The layers describe the structure. The principles decide how it behaves under load. Each one is explored in depth in its own essay.
Escalation is designed by cost of error
The trigger for human involvement is the cost of being wrong, calibrated per decision type. A fee waiver and a CHF 2 million lending exception should never share the same path.
Humans govern outcomes, not tasks
People concentrate where judgement and accountability live: exceptions, envelope design, relationships. Routine work in the human layer is a design failure.
Every agent has a job description
Scope, policy envelope, escalation rules, a named owner, and a review cadence. An agent without a job description is unmanaged headcount.
Autonomy is earned in levels
No board approves a jump from supervised recommendations to autonomous decisions. Evidence at each level is the permission slip for the next.
The demand side sets the clock
When clients send agents, response time and unit cost stop being service metrics and become conversion drivers. The pace of redesign is set outside the bank.
The ladder
Autonomy is a progression
Borrowed from driving automation, applied to banking. Most Swiss institutions operate at Level 1 to 2. The structural advantage begins at Level 3.
Fully manual
Paper, spreadsheets, manual approvals.
Assistance
Bots extract, check, route. Deterministic.
most banks
Partial automation
AI recommends, a human reviews every output.
most banks
Conditional autonomy
Agents decide within envelopes, escalate beyond.
advantage begins
High autonomy
Agents run whole domains. Humans govern envelopes.
Full autonomy
Direction of travel. Regulatory, and later.
FINMA Guidance 08/2024 permits autonomous use once systems are “sufficiently reliable and this can ultimately be proven.” The evidence generated at each level is the permission slip for the next. How trust accumulates →
Worked example
What changes when you apply it
Take mortgage origination, the process every Swiss bank runs the same way. The three-layer model redesigns it without changing what FINMA requires.
Bot layer
Ingestion. Eligibility checks. Data enrichment. Sanctions screening.
Time: seconds. Cost: near zero.
Agent layer
Risk assessment. Pricing. Offer generation. Escalation decisions within the policy envelope.
Time: minutes. Cost: low.
Human layer
Exception review. Relationship calls. Policy governance. Quarterly performance review of agents.
Time: where it matters. Cost: where it matters.
Same process. Same FINMA requirements, designed in. Different cost structure, different speed, different accountability boundaries.
Read the full case sketch →Beyond lending
The same model, other domains
Next step
Where does your institution stand today?
Fifteen questions across the five dimensions this framework defines. About five minutes. Exportable as a PDF for your leadership team.
No registration. Individual answers stay in your browser.
Take the assessment