The design is the monetary boundary plus sampled control lane.
AI value · Touchless share
Company-reported
About 25% of cheque volume processed with practically no human involvement
Executive disclosure through independent financial outlets; no error rate.
Before
Reads low-value cheques and verifies mandatory fields and compliance requirements. → Clears or routes the cheque after manual verification.
After
Reads and validates eligible cheques up to ₹10,000 and passes qualifying items without routine touch. → Samples AI-passed cheques, identifies misses, and determines whether retraining is needed.
Human boundary
AI/STP passes eligible cheques; bank sets threshold; CRU audits samples.
Why it matters
A bounded low-value tier can move to touchless processing while assurance shifts to sampling.
This case is company-reported. Use it for the operating-model shift; do not treat the numbers as independently measured.
How the work changed
Before
How the work ran before the change.
Step 1 of 2
Cheque-processing staff
Reads low-value cheques and verifies mandatory fields and compliance requirements.
ControlHuman review was the stated prior operating mode; sources do not disclose baseline time or error rate.
Step 2 of 2
Operations staff
Clears or routes the cheque after manual verification.
ControlBank payment controls and value limits apply.
What changed
A bounded low-value tier can move to touchless processing while assurance shifts to sampling.
Decision rightHuman sets the threshold; AI decides each instance
After
How the same work runs now.
Step 1 of 2
Vision-LLM/STP system
Reads and validates eligible cheques up to ₹10,000 and passes qualifying items without routine touch.
ControlBank-set ₹10,000 scope; approximately 25% of cheque volume.
Step 2 of 2
Control Risk Unit
Samples AI-passed cheques, identifies misses, and determines whether retraining is needed.
ControlHuman sampled assurance; above-threshold and exception handling details are not quantified publicly.
Process model built from the published workflow evidence for State Bank of India. Every step, actor, and control appears in full below.Every step, actor, and control
Exception path
Above-threshold or uncertain cheques remain human-reviewed.
Work removed
Per-cheque human review below ₹10,000
Decision authority
AI/STP passes eligible cheques; bank sets threshold; CRU audits samples.
Before
01
Cheque-processing staff
Reads low-value cheques and verifies mandatory fields and compliance requirements.
Control: Human review was the stated prior operating mode; sources do not disclose baseline time or error rate.
02
Operations staff
Clears or routes the cheque after manual verification.
Control: Bank payment controls and value limits apply.
After
01
Vision-LLM/STP system
Reads and validates eligible cheques up to ₹10,000 and passes qualifying items without routine touch.
Control: Bank-set ₹10,000 scope; approximately 25% of cheque volume.
02
Control Risk Unit
Samples AI-passed cheques, identifies misses, and determines whether retraining is needed.
Control: Human sampled assurance; above-threshold and exception handling details are not quantified publicly.
Work that left the path
Per-cheque human review below ₹10,000
Human role before
Cheque-processing staff read and validated mandatory fields on low-value cheques, then handed eligible items to payment processing under bank value and compliance controls.
Human role after
Control Risk Unit samples passed cheques, identifies misses and triggers retraining.
AI roleUses Vision LLMs to read and validate cheques up to ₹10,000.
Outcomes
Touchless share
Company-reported
Human cheque verification→About 25% of cheque volume processed with practically no human involvement
FY2026 · SBI cheque operations
Executive disclosure through independent financial outlets; no error rate.
What leaders can reuse
Anti-pattern
Do not infer accuracy or savings from touchless volume.
Questions
01Where is the operating threshold set and who can override it?
02What measured result would trigger rollback or retraining?
03Which residual decisions must remain human-owned?
Portability conditions
Monetary threshold
Exception path
Loss monitoring
Reputation risk
low
Evidence and authority
What the public record supports.
Current · updated
2 independent; publication outcomes are reported.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID d31443b5cf716855
