Inventory adjustments and release overrides: computing the gating threshold

Every argument about whether to add an approval step is really an argument about a number nobody has computed. Here is the number, and the formula that produces it.

Part of Freight, cargo and customs identity

The argument nobody wins

Operations says an approval step on inventory adjustments will slow the floor and the floor is already short-staffed. Finance says unexplained shrink is running above tolerance and somebody should be accountable. Both are correct, and the argument recurs quarterly because neither side has a number.

The number exists. It is the adjustment value above which the expected loss prevented exceeds the operational cost of requiring a signature. Below it, gating destroys value. Above it, not gating destroys value.

The formula

Let c be the fully loaded cost of the signing gesture in dollars — seconds of operator time, converted. Let p be the probability that a given adjustment is illegitimate or erroneous in a way the signature would deter or surface. Let v be the value of the adjustment.

Gating is worth it when:

p × v  >  c

therefore   v*  =  c / p        (the threshold)

with  c = (gesture seconds / 3600) × loaded hourly rate
      p = estimated from your own adjustment investigation history

Work an example. A three-second gesture at a fully loaded rate of $38 per hour gives c ≈ $0.032. If one adjustment in four hundred proves problematic on investigation, p = 0.0025, and v* ≈ $12.70.

That result surprises people. The threshold is low because the gesture is cheap, which means the honest answer is usually to gate more than operations expects and far less than finance would like to mandate.

Estimating p without deceiving yourself

This is where the exercise goes wrong. p is not the fraction of adjustments that were fraudulent; it is the fraction that a signature would have deterred or surfaced. Those are different, and the second is smaller.

• Start from investigated adjustments, not all adjustments. You only know outcomes where someone looked.
• Exclude adjustments whose root cause was a system or process error — a signature would not have changed those.
• Include adjustments where the person performing them could not later be identified, because attribution is part of what the signature buys.
• If you have no investigation history, say so and use a range. Report v* at p = 0.001, 0.0025 and 0.01, and let the committee choose.

The distribution does most of the work

Pull twelve months of adjustments and plot value on a log scale. In every warehousing operation I have seen the shape is the same: a very large number of small corrections and a thin tail of large ones.

A threshold set at the third band gates five percent of adjustments and covers eighty-five percent of the value. That is the entire argument, expressed as arithmetic rather than as a position.

Where the formula generalises

Nothing in it is specific to warehousing. Substitute the action and the loss mechanism and it applies to purchase order approvals, credit memos, refund authorisations, price overrides, and write-offs of any kind.

It also applies in reverse, which is the use most organisations neglect: if you are currently gating an action whose v* exceeds its typical value, you are spending operator time on a control that cannot pay for itself. Removing controls is a legitimate output of this analysis and it is the one that buys you credibility with operations.

The second-order effect

There is an effect the formula does not capture and it is worth naming. A thresholded control changes behaviour at the threshold: adjustments cluster just below it.

Monitor for that. A histogram spike immediately under your threshold is not a coincidence, and it is a more interesting finding than anything above the threshold.

What to hand the committee

1. The adjustment value distribution for twelve months, log scale.
2. c, measured rather than assumed, from a timed gesture on your actual hardware.
3. v* at three values of p, with the estimation method stated.
4. Coverage at the proposed threshold: percent of count gated, percent of value covered.
5. A commitment to re-plot the distribution after two quarters and look for clustering.

Why value alone is not enough

A gate on value alone is defeated by splitting, which is why the frequency axis matters. The practical rule is a value threshold plus a cumulative-per-actor threshold over a rolling window.

Objections and honest limits

“Operations will route around any gate.” If it fires often, yes. At the 99th percentile it fires on roughly one adjustment in a hundred, which is a rate operations can absorb.

“Our WMS cannot do per-value authorisation.” Most cannot natively. The gate sits in front of the adjustment endpoint, which is smaller than a vendor feature request.

Setting the threshold

1. Export twelve months of adjustments. With value, reason and actor.
2. Set the value gate at the 99th percentile. Not at a round number someone guessed.
3. Add a cumulative-per-actor gate. So splitting does not defeat it.
4. Re-derive annually. Distributions drift with volume and mix.

Terms used here

Inventory adjustment

A correction to recorded stock, which can conceal shrink or theft when unreviewed.

Override

Bypassing a system control, typically granted broadly because operations need it.

Cumulative threshold

A limit on total value adjusted by one actor over a rolling window, which defeats splitting.

Frequently asked questions

What if we have no investigation history at all? Report a range rather than a point estimate and state the assumption explicitly. A decision made on a stated range beats one made on an unstated intuition.

Does this apply to negative adjustments only? Both directions. Unexplained positive adjustments are as diagnostic as negative ones and are frequently the earlier signal.

Should the approver be a second person? That is a separate decision. The threshold answers whether to require a signature; whether it must be someone else's is a segregation-of-duties question with its own analysis.

How often should the threshold be recalculated? Annually, or after any material change to the gesture cost or the operation's mix. Recalculating quarterly produces noise.

How should the threshold be set? From your own twelve-month distribution. The 99th percentile fires on roughly one adjustment in a hundred.

Why add a cumulative threshold? Because a value-only gate is defeated by splitting one large adjustment into several small ones.

What if the WMS cannot gate by value? Put the gate in front of the adjustment endpoint rather than requesting a vendor feature.

Sources and further reading

• Warehousing and inventory accuracy audit standards.
• Third-party logistics contractual inventory accuracy and shrink provisions.
• Uniform Electronic Transactions Act (ULC)
• NIST SP 800-53 Rev. 5 — access enforcement