Maker-checker is a two-person control. One person (the maker) creates or changes a record, and a different person (the checker) reviews and releases it before it takes effect. It is meant to detect errors and fraud, but in most banks it behaves like a queue, because the checker approves almost everything and the delay is real.
A Kuwait bank cut its procurement cycle from 139 days to 57, a 59% reduction, by removing and simplifying steps before any automation. Handoffs, approvals included, were a large part of that wait. This article is about the same problem: a handoff that ages while being filed as a control.
What is the maker-checker process?
The maker-checker process splits one action across two identities so that no single person can both initiate and authorise it. The maker enters the payment, the beneficiary change, the limit increase, or the journal. The checker, with separate credentials and separate accountability, approves, rejects, or returns it.
In banking operations the pattern shows up everywhere:
- Payment release and high-value transfers
- Beneficiary and counterparty set-up
- Customer static data changes, such as address, signatory, or mandate updates
- Limit and exposure changes
- General ledger journals and reconciliation adjustments
- Know-your-customer (KYC) onboarding decisions
- Procurement and vendor approvals
Policies and audit reports also use the terms segregation of duties, four-eyes principle, dual control, and initiator and authoriser. They overlap but are not identical.
Maker-checker vs four-eyes principle vs dual control
Teams use these phrases interchangeably, and the first design mistake starts there. They describe different things.
| Term | What it requires | Typical question it answers |
|---|---|---|
| Maker-checker | Sequential: one person prepares, a second person reviews and releases | Did a second person see this before it took effect? |
| Four-eyes principle | Any decision involves at least two people, in sequence or together | Was this decision seen by more than one person? |
| Dual control | Two people act together on one sensitive step, such as holding split keys | Can one person alone complete this step? |
| Segregation of duties | Conflicting roles never sit with the same person | Can this person both create and approve? |
Maker-checker is the sequential form of the four-eyes principle. Dual control is stricter, because both people must act, not just one after the other. Segregation of duties is the underlying design rule that the other three implement.
A bank can satisfy all four on paper and still run a control that detects nothing. The label tells you who clicked, not what anyone decided.
Why do regulators care about maker-checker in Singapore and Malaysia?
The Monetary Authority of Singapore (MAS) and Bank Negara Malaysia (BNM) both expect licensed institutions to separate duties, restrict access to sensitive functions, and keep records that show who did what. That expectation sits within MAS technology and operational risk supervision and BNM's risk management in technology requirements.
When an examiner or internal auditor asks about a maker-checker control, they rarely stop at "does it exist". They ask you to evidence how it operated. Take the exact wording of the instrument that applies to your licence from your compliance team, not from any summary, including this one.
That evidence request is where most maker-checker designs fall apart.
Why does maker-checker slow everything down and catch nothing?
It was designed as a signature, not as a detection mechanism. Four failure modes follow.
1. The checker approves almost everything
Ask an operations manager how often the checker rejects an item. In most gates the honest answer is almost never. That is a finding about the design, not about the checker.
Most items are correct because the maker is competent and the system already validates the format. The checker sees a correct item, approves it, and the queue moves on. A control that approves almost everything it sees buys close to no detection and costs a full handoff.
2. The checker has a context deficit
The maker built the item and knows why it looks the way it does. The checker receives a screen. They did not see the customer email, the call, or the exception that explains the odd amount.
The checker makes a quick judgement on partial information, and the approval record presents it as a considered review. Only one of those is a control.
3. The approval record proves a click, not a judgement
A system log shows that user B pressed approve at a given time. It does not show what user B looked at, which criteria they applied, or why they were satisfied.
If you picked twenty items from last month, could you say what the checker saw that the maker did not? If not, you have a signature, not a control.
4. The latency is real and compounds
Every gate is a handoff. Handoffs wait for availability, for shift changes, and for a checker who is also a maker on another queue. A cycle-time analysis separates that waiting from processing time, and the gap is where cases age between steps.
The reframe: maker-checker is an evidence problem
You call it a control. Your queue calls it a handoff.
Maker-checker is an evidence problem wearing a signature's clothes. The control that detects is the exception trace: which items were stopped, by whom, on what rule, and what changed afterwards.
An auditor can test that trace, and it tells you whether a gate earns its place. A gate with a rich exception trace is doing detection work. A gate with no stopped items and no recorded reasoning is a ceremony.
Once you capture the exception path, the second signature stops being the safety mechanism and becomes a service-level agreement (SLA) question. How fast does the checker respond, and what happens when they do not? That is a cycle-time enforcement problem, and you can measure and manage it.
A worked example: the 14-gate matrix
Consider a team that maps its own maker-checker matrix as a process for the first time. This is a hypothetical, constructed example, not a client story.
The team lists every approval gate across payments, static data, limits, and journals. There are 14. It then pulls the last quarter of decisions for each gate and counts how often the checker stopped an item.
Eleven of the 14 gates never rejected anything. The checker approved every item.
Suppose the team then finds a familiar pattern. Some gates were added after an old incident and never reviewed. Two repeat a validation the system already performs. At least one exists because a policy names a checker role, and nobody has written down what that checker should look for.
The team does not delete 11 gates. It now has a map of which approvals involve a decision and which do not, and that map is the starting point for every sensible change.
The approval-matrix teardown
To separate a control from a ceremony, trace each gate across four columns.
| Gate | Decision actually made | Evidence produced | Latency added |
|---|---|---|---|
| Payment release over threshold | Compare beneficiary and amount against the instruction | Reason code on rejection; amount and beneficiary snapshot | Same day |
| Beneficiary set-up | Verify the account against supporting document | Document reference and verification result | Next day |
| Customer address change | None recorded; system already validates the format | Approve click only | Next day |
| Limit increase | Judge exposure against appetite | Free-text rationale (sometimes) | Two days or more |
| Journal over materiality | Confirm the supporting entry and the reason | Link to the supporting document | Same day |
| Vendor onboarding | None recorded; checks already done upstream | Approve click only | Two days or more |
The table is illustrative. Rows with an empty decision column and "approve click only" as evidence are the first candidates for redesign.
The four-question test for any approval gate
You can run this on your own matrix in an afternoon. Ask four questions of each gate.
- What did the checker see that the maker did not? If the answer is nothing, the second pair of eyes adds no information.
- What would make the checker stop this item? If nobody can state the criteria, the checker is guessing.
- When did this gate last stop something, and what happened next? If the answer is "never" or "we do not know", you cannot claim detection.
- What does the record show about the decision? If it shows only a name and a timestamp, you have evidence of a click.
A gate that passes all four is a control. A gate that fails two or more is a candidate for the scorecard below.
The scorecard: which approvals earn their latency
This is the one-page artefact to run on your own matrix. Fill it in from a quarter of real decision logs, not from memory.
| Gate | Items passed | Items stopped | Time added (median) | Evidence type | Verdict |
|---|---|---|---|---|---|
| Name the gate | Count from log | Count from log | Hours or days | Click only / reason code / document link / free text | Keep / Rewrite / Convert / Retire |
How to read each column
- Items passed counts every item the checker approved in the period.
- Items stopped counts every item rejected, returned, or queried, whether or not it was later approved.
- Time added is the median wait between the maker submitting and the checker deciding. Use the median, because averages hide the long tail.
- Evidence type is what the record holds about the decision. Rank it from weakest to strongest: click only, free text, structured reason code, then reason code plus a snapshot of what the checker saw.
The verdict rules
| If the gate shows... | Verdict | Rewrite rule |
|---|---|---|
| Stops items regularly, with structured reasons | Keep | Add the checker's criteria to the work instruction and set a response-time target |
| Stops items, but the record is click only or free text | Rewrite | Capture a reason code and a snapshot of what the checker saw, so the trace becomes testable |
| Stops nothing, and the risk is low or already checked by the system | Retire | Remove the gate, or replace it with a periodic sample review |
| Stops nothing, but the risk is high or the requirement is mandatory | Convert | Keep the second signature, define the criteria, add an SLA, and escalate when the checker is slow |
Two warnings apply. A gate that has stopped nothing may still deter, so look at the risk, not only the count. And where regulation or your licence conditions require dual control, keep it and fix the evidence and response time instead. Confirm with compliance before retiring anything.
How to rewrite the gates that fail
The E-S-S-A-M framework (Eliminate, Simplify and Standardise, Automate, Migrate) gives the order of work. Automate is the fourth action, deliberately. You design before you automate.
- Eliminate gates with no decision in them. This is the cheapest win, because the wait disappears entirely.
- Simplify and standardise the gates that stay. Write the checker's criteria once, in plain language, and show the checker what the maker saw.
- Automate deterministic checks such as format, threshold, and duplicate detection. Keep the human checker for judgement.
- Migrate the remaining approvals to the channel people already use, so response time stops depending on who is logged in.
The Kuwait bank procurement case followed this sequence. It is a real case, and it is a procurement cycle only. The bank went from 139 days to 57, a 59% reduction, with 82 days of work retired and the same headcount. Eliminate and Simplify delivered the result, and it came before any automation.
Do not read that outcome across to payments or static data, which carry different risks. What transfers is the method: map the handoffs, find the ones with no decision, and remove them before you build anything.
Where this fits in the 7-step AI Lean Cycle
ESSAM runs a seven-step cycle that repeats continuously: Baseline, Map, Analyse waste, Optimise, Document, Deploy, Improve. Maker-checker touches several of them.
- Map: ESSAM captures the maker-to-checker path across core systems, workflow tools, and email in a single conversation, so the real route appears, not the one in the policy.
- Analyse waste: each approval step is classified with E-S-S-A-M, and gates with no decision show up as Eliminate candidates.
- Document: this is the evidence-generating phase. The checker's criteria, the reason codes, and the exception path become part of the work instruction.
- Deploy: the instruction reaches the floor over WhatsApp, the channel Singapore and Malaysia teams already use, so makers and checkers work from the same current version.
- Improve: the before and after audit view shows whether cycle time moved and whether stopped items still get caught.
Process owners should hold the gate list and the scorecard, because a control nobody owns drifts back into a ceremony. For the wider role, see what an AI process engineer does, and the features page for how capture and deployment work.
Frequently asked questions
What is the maker-checker principle in banking?
It is the rule that the person who initiates a transaction or change cannot be the person who authorises it. A second, independent person reviews it before it takes effect.
Is maker-checker the same as the four-eyes principle?
Nearly. Maker-checker is the sequential form of four-eyes: one person prepares, another approves. The four-eyes principle also covers decisions made by two people together.
Does maker-checker satisfy MAS or BNM expectations?
It can support them, but the pattern alone is not the test. Supervisors look for evidence that the control operated: who reviewed, against what criteria, and what was stopped. Confirm the exact requirement with your compliance team.
How many approval gates should a process have?
As many as there are real decisions, and no more. A gate earns its place when it can stop something, you can show it did, and the latency it adds is justified by the risk.
Can maker-checker be automated?
The deterministic part can. Format checks, thresholds, and duplicate detection belong in the system. The judgement part, where context matters, stays with a person who has been given the criteria and the information to apply them.
Closing
A second pair of eyes is a sound idea. The fault is treating the signature as the control and never measuring what the second pair of eyes did.
Pick twenty items from last month and run the four questions. Fill in the scorecard for your busiest gate. By the end of the afternoon you will know which approvals are earning their latency.
Map your own approval matrix in a single conversation and see which gates stop work and which only slow it: apac.essam.ai
