The process, as documented: one path, four steps, ~24 minutes. This is what the binder says happens.
The process, as observed: 14 variants, 4 minutes to 30+ days. The documented path is followed in 9.2% of cases. Line thickness = case volume. Hover any line.
Section 02 · The Variant Ledger
Every variant, priced.
Each handling path observed, with volume, cycle time, and its estimated annual cost in labour, overpayment, and write-offs.
Variant
Observed path
Share of exceptions
Median cycle
Annual cost
Field note
Scroll the table sideways for annual cost and field notes →
Section 03 · Cycle Time Distribution
The same exception. A 10,000× spread.
Resolution time for the 4,914 observed exceptions. The SOP target is one hour; two-thirds of cases miss it.
Section 04 · Simulated Optimisation · Causal Twin
14 variants in. 2 sanctioned paths out.
The calibrated twin replayed this operation across 10,000 simulated quarters with exceptions consolidated onto two sanctioned paths: one automated, one human-reviewed.
Mean exception cycle time
Observed
2.6 d
Simulated
1.0 d
▼ −61% ± 6%
Handling cost per exception
Observed
£6.10
Simulated
£1.95
▼ −68% ± 9%
Annual margin leak
Observed
£412k
Simulated
£96k
▼ £316k ± £41k recovered / yr
Deployment is gated, not assumed. Candidate optimisations run in shadow mode alongside the live operation: the system decides, the human decides, outcomes are compared with statistical rigour. Nothing goes live without a measured confidence interval. Figures above are simulated projections on synthetic data, shown to illustrate the method, not a client result.
Section 05 · Autonomy Diagnosis
Where this operation sits. Where it can be in two quarters.
You are here · L0.8
Target · L3 in 2 quarters
L0
Manual
Humans are the glue. Tribal knowledge, inbox-driven exception handling.
L1
Assisted
Agents draft and route; humans execute every step.
L2
Partial autonomy
Agents execute defined paths; humans approve exceptions.
L3
Conditional autonomy
Agents own the process end-to-end; humans supervise by exception.
L4
Self-optimising
The operation runs, improves, and documents itself. Humans set direction.
The framework paper
Read how the method actually works.
Self-Optimising Operations sets out the whole architecture: variant discovery, the inconsistency measure, the causal twin, and the L0 to L4 autonomy scale. Section 8 is a worked example on synthetic data showing one documented process against the fourteen routes the same work really takes.