Variance Analysis and Cost Control
How to read cost in three sentences a month

The month-end cost meeting usually goes like this.

"Why did labor cost come in so high last month?" "Volume was up. And we onboarded some new hires, plus there was overtime."

All three are probably true. But all three are missing a size. Without how much came from volume, how much from new hires, and how much from overtime, the meeting trades circumstances for thirty minutes and ends with no instructions. Next month, the same meeting convenes again.

Part 1 built the measuring stick called standard cost, Part 2 attached cost to shippers, and Part 3 covered how to measure time. This installment is about the system that lets an organization digest the output every month — variance analysis and cost control. The calculation was already finished in Part 1. Part 4 is not about calculating; it's about operating. When to pull the numbers, how to slice them, which variances to investigate, who each one goes to, and when to revise the standard itself.

First, a one-table refresher on the Part 1 skeleton. Our example center for this series — standard time 5.5 minutes, standard rate KRW 14,760, overhead absorption rate KRW 7,000 per hour, base volume 5,000 hours a month.

You often hear that variance analysis is an old tool. It is. But the surveys tell a different story. Standard costing adoption ran at 85–86% in US studies (1985, 1988), 65% in Japan (1991), and 71% in China (2016) — across decades, "we're about to abandon it" has never been the majority answer. A tool that is old yet unreplaced usually has a reason. For variance analysis the reason is simple: there is still no better way to split a total into units of responsibility.

Flexible budgets — where "versus budget" creates an illusion

Before diving into variance analysis, look at the problem with the method most warehouses use today: actuals versus budget.

Say the annual plan assumed 20,000 orders a month and you budgeted outbound labor accordingly. At the Part 1 standard of KRW 1,353 of labor per order, the budget is about KRW 27.06 million. This month, actual labor came in at KRW 25.5 million. KRW 1.56 million under budget — a good month?

If actual volume was 17,000 orders, the story flips. Processing 17,000 orders at standard should have cost about KRW 23 million. KRW 25.5 million is not a saving — it's a KRW 2.5 million overrun. A month that looked favorable against the static budget turns unfavorable the moment volume enters the picture.

Accounting calls this adjustment a flexible budget: re-basing "what this volume should have cost" by multiplying actual volume by standard unit values. If you built the standard cost from Part 1, the flexible budget comes free — standard hours allowed is the flexible budget.

Why this matters is that the illusion points in the dangerous direction. A slow month shrinks the total and looks well-run; a busy month swells the total and looks sloppy. Reality is often the reverse — in slow months you can't shed labor fast enough and efficiency slips, while in busy months the line heats up and efficiency improves. A static budget tends to become a device for scolding the good months and praising the bad ones.

Totals hide offsets — the activity matrix

Part 1's example covered a single activity, outbound. Now open up the whole center. The same center with a 5,000-hour base volume, this month — actual direct hours 4,300, actual rate KRW 15,610 (+5.8% versus standard).

As a total, the efficiency variance looks like this: standard hours allowed 4,113, actual 4,300. 187 hours — KRW 2.76 million unfavorable. A 4.5% overrun — the kind of size that invites "within control range" and a page turn.

Open it by activity and a different picture appears.

Inside that mild 4.5% total, outbound was sounding an 11.8% alarm. The favorable variances in putaway and value-added work were covering for it. And as we'll see shortly, the favorable variance in value-added work is itself a candidate for investigation — the work may not have gotten faster; an inspection step may have been skipped.

So the basic form of a variance report is not a one-line total but a matrix of activity × variance type. Rows are activities (five to seven is plenty — the same activity list as Part 2), columns are rate and efficiency, and one volume line at the bottom. This single page is the skeleton of the monthly cost report.

Slicing has limits too. When Part 3 covered ledger regression, the rule was to merge any combination with fewer than 30 observations — the same rule applies here. A variance in a thinly observed cell is noise, not signal. A matrix with forty activity rows looks precise, and half its cells swing on coincidence every month.

Open it once more, by shipper — warehouse's fault, shipper's fault, market's fault

If the activity matrix answers "which process leaked," the next question is "because of whom." Part 2 attached a shipper to every work event, so efficiency variances open up by shipper too.

Split outbound's KRW 3.2 million efficiency variance by shipper.

Shipper A's KRW 450,000 is a variance the warehouse made. The new-hire learning curve is passing through the 70–85% band we saw in Part 1, and it resolves in a few weeks. That's the floor's job.

Shipper B's KRW 2.1 million is different in kind. The workers didn't slow down — the orders changed. Lines per order rose from 1.8 to 3.1, so the same "one order" became a longer piece of work. Here Part 3's time equation creates a fork. If the standard time stands as a formula — base + per-line coefficient — this change is absorbed into standard hours allowed automatically, and the efficiency variance never appears. If you're on a single-constant standard — and the rate card is a flat per-order price too — this KRW 2.1 million is being recorded as the floor's failure when it is actually a pricing problem.

Repeat this classification monthly and variances settle into three piles.

An efficiency variance made by the shipper is not a performance problem — it's a pricing problem. That may be the most practical sentence in Part 4. It is not a variance you fix by pressing the floor harder; it's one you carry to the negotiating table. And at that table, "Shipper B's order profile changed and is adding KRW 2.1 million of cost per month" is a sentence with force — because it came out of the ledger.

Compress it into three sentences a month

Build everything above — flexible budget, activity matrix, shipper split — and the report gets thick. Thick reports don't get read. So a final process step is needed: compression.

The top of the monthly cost report needs only three sentences. One for rate, one for efficiency, one for volume. Each sentence carries four parts — direction and size (how favorable or unfavorable), concentration (which activity, which shipper), one cause (only the biggest), and one next action.

Our example center's month, written as three sentences:

Compare the bad version: "Labor cost rose 8% over the prior month. Volume growth and increased overtime are analyzed to be the causes. Continued cost-reduction efforts are needed." No direction, no size, no concentration, no action. The third sentence means nothing at all. A sentence without a next action gives the reader no work to do, and a report that assigns no work meets the same numbers again next month.

Speed is part of the format too. In financial-close benchmarking (APQC, roughly 2,300 organizations), the monthly close takes a median of 6.4 days, with the top quartile at 4.8 days or less and the bottom quartile at 10 days or more. The cost report should ride the same rhythm. A provisional variance on day 5 beats a precise one on day 15 — corrective action can still land inside the month. Close provisionally first; if the gap between provisional and final proves persistently large, fix the aggregation process then. And once a month is closed, freeze it. If last month's numbers move this month, the whole time series becomes a negotiation.

Which variances to investigate — management by exception

Build the matrix and you have dozens of cells; you cannot investigate all of them every month. A plan to investigate everything ends as investigating nothing. What you need is an investigation threshold. Management by exception — formalized in the era of scientific management — is the operating rule of variance analysis: let what's inside the fence pass, and put only what crosses it on the desk.

"When is a variance worth investigating" is also an old academic question. Robert Kaplan's 1975 survey paper in the Journal of Accounting Research framed it: investigation has a cost, variances contain random variation, so investigation is justified only where expected benefit exceeds the cost of looking. You don't need the statistical models to act on the insight — a practical compromise is three thresholds.

The third row is the most often forgotten and the most important. A cell that swings ±6% randomly every month never trips a threshold, but +4% three months running is not chance. This is exactly the distinction Walter Shewhart formalized with control charts at Bell Labs in the 1920s — common-cause versus special-cause variation. React to every common-cause wiggle and the organization chases noise; miss the special causes and structural problems set like concrete. If your ledger holds full-population data as in Part 3, you can even compute each activity's natural variation band and set control limits from data. Even if you don't, the "three months running" rule alone does 80% of a control chart's job.

One more thing — favorable variances get investigated too. A practice of investigating only unfavorable ones misses three patterns.

① A favorable rate that transferred into unfavorable efficiency. More cheap day labor makes the rate variance favorable, while lower skill makes the efficiency variance more unfavorable than the saving. Viewed separately, procurement gets praised and the floor gets scolded; viewed together, the center lost money.

② Favorable efficiency borrowed from quality. That KRW 1.03 million favorable in value-added work is the suspect here. Skip an inspection step and the efficiency variance improves — and the bill returns next month as mis-shipments, returns, and claims, with interest the variance account never shows.

③ A loose standard. If everyone beats the standard comfortably every month, it likely means the yardstick is generous, not that the team is heroic. Rates built on generous standards price high in the market and lose competitive quotes quietly.

The politics of variances — ask only what can be controlled

In warehouses where variance analysis fails, the calculation is rarely what broke. It usually collapses when variances become a tool for hitting people.

The principle of responsibility accounting is simple: hold people accountable only for what they control. Part 1 said each variance has a different owner; in operation you need one more layer, because having an owner and being controllable by that owner are different things. Charging the volume variance to a floor supervisor is obviously wrong — and charging the outbound team for an efficiency variance created by Shipper B's order-profile shift produces the same result. An organization graded on numbers it cannot control starts manufacturing the numbers — cherry-picking easy orders, deferring hard ones, gaming scan timing. This is why Part 3 tested every driver for manipulation resistance.

There is a famous objection at this exact point. In his fourteen points, W. Edwards Deming argued for eliminating work standards (quotas). Most variation comes from the system, not the individual — so managing system-made variation with individual quotas leaves only fear and manipulation. The reason a series advocating standard costing cites this critique is that Deming's target was not the standard itself but the use of the standard. The same standard time used as a personal scorecard fulfills Deming's prophecy. Used for cost attribution and pricing evidence — this series' use — it sidesteps the problem.

Hence three rules for the variance meeting.

1. Report in sentences. Tables are attachments; the body is three sentences. No reading numbers aloud.
2. No personal names. Causes are written in the language of activities, structures, and order composition. Not "the outbound team was slow" but "Shipper B's line-count growth sat outside a single-constant standard."
3. One next action, verifiable next month. "Raise cost awareness" is not an action. "Prior approval for overtime" is verified by next month's overtime hours.

Standards are consumables — the revision discipline

The last use of variances is to watch the standard itself. If a variance runs the same direction beyond ±15% for three straight months and no floor-level cause can be found, one suspect remains: the standard has gone stale. Part 1 said ±30% is normal for a first standard — but that tolerance belongs to the first quarter. A settled standard that keeps missing in one direction is a measurement problem, not a floor problem.

Put revision triggers in writing.

• Immediately when the process changes. Conveyor, batch picking, put walls — the day layout or method changes, the old standard is scrap from that day. As Part 3 noted, best practice for recalibration is every 18–24 months, but a process change doesn't wait for the cycle.
• Rate standards every January. In a Korean warehouse's calendar this is a constant. The 2026 minimum wage is KRW 10,320 (+2.9%); wages move by index, and overtime premiums and social-insurance burden rates move with them. Centers that post a large January rate variance are, most often, centers that didn't revise the standard rate.
• Time standards when the variance says so. The three-months-running rule above is that signal.

And when you revise, leave a record. Revision date, old value, new value, one line of reason. Without it the time series collapses. If you're comparing last year's efficiency variance to this year's and nobody knows whether the yardstick changed in between, that comparison is fiction, not analysis. A standard is not a monument erected once; it is a consumable replaced on a schedule, and consumable swaps get logged.

2026 — the environment for reading variances has changed

① The volume variance stepped out of the books and became nationwide physical fact.

Per market research on Seoul-metro logistics centers (RSQUARE), average vacancy in H1 2026 stands at 12.3% for ambient and 33.7% for cold storage. Both are coming down from the H1 2024 peak (16.8% and 41.2%), and new supply has collapsed to about 580,000 m² — roughly a fifth of H1 2023 — but regional spreads remain wide: cold-storage vacancy in the northwest corridor reaches 70.0%. A vacancy rate is, in the end, the market's aggregate volume variance. Nobody absorbs the fixed cost of an empty position for you, and as the storage-costing piece showed, it demands honesty about the denominator. A center that doesn't pull its volume variance into a separate line each month lets this pressure seep into unit costs somewhere — and walks into renewal negotiations not knowing why a given shipper's quote no longer works.

② Volume stagnation became the default, not the exception.

In the annual US 3PL benchmark survey (Extensiv), 22% reported flat or declining order volumes, and 65% are running below 90% of capacity. In a growth phase, the efficiency variance is the protagonist of cost management; when volume stalls, the volume variance takes the lead. The floor can beat standard every month while unabsorbed fixed cost erases the gain — in this phase the cost meeting must be a volume-and-price meeting, not a floor-improvement meeting. Variance analysis is the device that tells you, in numbers, when to switch.

③ The rate variance became an annual event.

Labor cost pressure hits 70% of 3PLs in the same survey, and in Korea the minimum-wage link resets the rate base every January. On the transport side, the reinstated Safe Trucking Freight Rate System institutionalized a rate floor. Which means much of the rate variance is now a calendar variable, not a negotiation variable. Calendar variables can be forecast, and what can be forecast can be designed for — revising standards and repricing on schedule. The difference between a center surprised by its January rate variance and one that adjusted standards and rate cards in December is not information. It's discipline.

④ The cadence of variances is moving from monthly to daily.

Monthly variance analysis is accounting's rhythm, not operations'. If the ledger accumulates in real time, efficiency variances can be read daily, with AI anomaly detection running on top. But Part 1's conclusion holds here too — there is no anomaly detection without a baseline. Without a standard, the only way to compute "today looks off" is the historical average, and the historical average learns yesterday's problems as normal. The practical deployment is a double rhythm: efficiency on a daily/weekly dashboard, rate and volume at monthly close. Daily numbers are provisional; closed monthly numbers are frozen.

⑤ Automation is changing the composition of variances.

Automation converts variable cost (labor) into fixed cost (depreciation), as Part 1 showed. In the language of variance analysis: the stage for rate and efficiency variances shrinks while the stage for the volume variance grows. Human efficiency recedes as a management object and equipment utilization takes its seat. Cost control in an automated warehouse shifts from "faster" to "volume to fill it with" — and securing that volume is sales-and-pricing work, not floor work. The more you automate, the more the variance report's addressee moves from the floor to the boardroom.

What you can do in the first month

1. Fix the close calendar. Provisional close at D+5 every month, marked provisional. The goal isn't an accurate day 15 — it's a usable day 5.
2. Build one activity × variance matrix. Five activities × two columns (rate, efficiency) + one volume line. With Part 1's formulas, a spreadsheet is enough.
3. Adopt the three-sentence format. Direction and size, concentration, one cause, one next action. Keep the format even if the first month's sentences feel stiff.
4. Put the three investigation thresholds in writing. Absolute amount, percentage, three months running. Agree together that variances below threshold don't reach the meeting.
5. Pick one favorable variance and investigate it. Do this once in the first month and the equation "investigation = blame" breaks. Precedent sets the practice.
6. Start a standard revision log. Date, old value, new value, reason. The first entry is today's date, recording the current standard.

The precondition for all six is the one thing this series has repeated throughout: actual records at the activity level. Without the ledger, the flexible budget, the matrix, and the shipper split have nothing to compute.

Control is design, not explanation

"Cost control" is easy to mishear. It sounds like explaining last month's numbers. But last month cannot be controlled. It has already happened.

Variance analysis qualifies as control because its output is not an explanation but next month's changes. The overtime approval regime, the rate discussion with Shipper B, the standard-rate revision — the actions at the end of the three sentences are the actual control devices, and variances are the instrument panel that tells you where to install them. Read the panel as precisely as you like; if you never turn the wheel, the car goes where it was going.

And the most valuable discovery in this series usually comes out of the second sentence: the efficiency variance made by the shipper — the point where the price, not the floor, is wrong. Once numbers like Shipper B's KRW 2.1 million start coming out of the ledger every month, the next question asks itself: how do you fold that cost back into the rate card? Part 5, "From Cost to Price," takes that question — the final piece connecting rate cards, contracts, and profitability to cost.