Petty cash and small spend — the costliest transactions you own

Consider a RM 35 taxi receipt.

Somebody keeps it, remembers to submit it, fills in a form, attaches it, sends it for approval. An approver reviews it. Finance checks it against policy, codes it, posts it, reconciles it, and files it.

Being generous, that is fifteen minutes of collective attention. Costed at any reasonable rate, processing exceeded the value of the transaction.

This is why small spend is worth automating even though the amounts are trivial. The amounts are the point: the cost is entirely in the handling, so removing the handling removes almost all of the cost.

What makes small spend different

The volume-to-value ratio is inverted. Hundreds of transactions, negligible individual value, identical processing cost to a large one.

The documentation is the worst you receive. Crumpled thermal receipts, photographs taken in a car, faded print. Whatever handles this has to cope with genuinely poor input.

Policy questions are constant and small. Was this within limits? Does it need a receipt? Is it allowable? Each takes seconds and there are hundreds.

It carries disproportionate risk. Not because anyone loses much on one claim, but because small spend is where control weakness normalises. A culture where the rules are loosely applied on RM 35 items does not stay confined to RM 35 items.

What automation handles well

Reading a bad photograph. Extracting merchant, date, amount and tax from a receipt captured on a phone in poor light. This was not practical with template-based extraction and is now routine.

Immediate capture. Photograph at the point of spend rather than a shoebox at month end. This single change removes most of the pain, because the reason expense claims are painful is that they are assembled retrospectively from things people half-remember.

Policy checking. Amount limits, category rules, receipt requirements, duplicate detection — applied consistently to every claim rather than depending on whether the approver is busy.

Coding. Small spend is highly patterned. The same merchants recur, the same categories, the same people.

Duplicate detection. The same receipt submitted twice, or claimed by two people who shared a meal. Genuinely hard to catch manually and easy for a system.

The control that actually matters

Not approval. Approval of small items is largely theatre — an approver receiving forty claims of RM 20 to RM 200 approves them in bulk, because examining each is not a rational use of their time.

The controls that work on small spend are:

Policy applied at submission. Reject or flag at the point of claim rather than after. Prevention rather than review, and it also educates — people learn the rules by encountering them.

Anomaly detection across claims. The useful question is not "is this claim valid" but "is this person's pattern unusual". Claims consistently just under the receipt threshold. Weekend claims from someone who does not work weekends. Sudden change in frequency. None of these are visible when reviewing claims one at a time, and all are visible in aggregate.

Sampling with real scrutiny. A small number of claims examined properly, including the underlying receipt, beats every claim glanced at.

That combination catches more than an approval workflow does, at less cost to everybody.

Petty cash specifically

Physical cash is a separate problem from card spend, and the honest advice is to reduce it rather than automate it.

Cash has no independent record. A card transaction exists in the bank feed whether or not anyone submits a receipt; a cash payment exists only if someone writes it down. That asymmetry is why cash floats are reconciled by counting rather than by matching, and why differences are so hard to explain.

Where cash is unavoidable, the useful controls are a small float, frequent reconciliation, and a single named custodian. Where it is avoidable — which is most places — a card removes the entire class of problem, because the transaction records itself.

What good looks like

• Receipt photographed at the moment of spend
• Data extracted and coded automatically
• Policy checked immediately, with the reason given if flagged
• Routine claims reimbursed without an approval step
• Approvers see exceptions and pattern anomalies, not every claim
• Sampling done properly on a small number

The outcome is that claims stop being a monthly event people dread and become invisible — which is also what makes people submit them promptly, which is what makes the numbers right.

Common questions

Is it worth automating small expense claims?

Yes, and the small amounts are precisely why. Processing cost is roughly the same regardless of transaction value, so a claim worth RM 35 can consume more in collective handling time than it is worth. Automating capture, coding and policy checking removes almost the entire cost, because with small spend the cost is nearly all in the handling.

Can AI read a poor-quality receipt photograph?

Generally yes — extracting merchant, date, amount and tax from a photograph taken in poor light is routine for modern document reading, which is a change from template-based extraction that required consistent layouts. Very faded thermal receipts remain difficult, which is an argument for capturing at the point of spend rather than weeks later.

What is the best control over expense claims?

Applying policy at the point of submission rather than reviewing afterwards, combined with anomaly detection across a person's claims over time and proper scrutiny of a small sample. Approving individual small claims is largely ineffective, because an approver facing forty small items will clear them in bulk regardless of the process.

Should we automate petty cash?

It is usually better to reduce it. Physical cash has no independent record — a card transaction appears in the bank feed whether or not anyone submits a receipt, while a cash payment exists only if someone writes it down — which is why cash differences are so hard to explain. Where cash is unavoidable, keep the float small, reconcile frequently and have a single named custodian.

Related: expense claims without the headache · approval workflows people actually follow · controls that survive automation