Margin & Exception Intelligence

Find the issues hidden by aggregate reporting.

What changed that should not have changed—and what deserves management attention?

Demonstration project using synthetic data designed to simulate a realistic operating company.

The business problem

Halbrook Industrial Supply grew revenue to $38.40M and lost 1.59 points of gross margin over two years. The monthly reporting package showed both facts and explained neither. It reported one blended margin, in absolute dollars, so a decline spread across twenty-four months never looked like an event and nothing in it said which accounts, which products or which branches had moved.

The company has 142 accounts and 278,169 invoice lines across 36 months. Somewhere in it are the reasons. The question is not whether the data holds them — it is whether anyone can be pointed at the few that matter without reading all of it.

Why standard reporting is not enough

A blended margin is an average of two businesses. Distribution sells parts at 22.1%; Installed Service sends technicians out at 41.8% — nearly double. Both moved a little: Distribution gave up 1.30 points, Installed Service gained 0.60. Weighted, that is 0.81 points of the fall. The other 0.79 is Installed Service dropping from 26% of revenue to 22% — the same margins, on a worse blend.

That is a different problem from either line getting worse, and it needs a different decision. Pricing discipline does not recover it; selling more service does, on a longer timescale. A report that shows only the blend cannot tell the two apart.

Analytical approach

Rules run before statistics, because a rule management already understands produces an explanation rather than a flag. A line billed below what it cost is wrong on its face. An account two points past its segment’s discount schedule is a conversation someone can have. Neither needs a reader to know what a robust deviation is.

Statistical detection then covers what no rule anticipated — a segment sitting away from its peers, or a series that changed level on a date. Attribution runs alongside both, because a movement is only actionable once it is attributed, and at account level attribution is the route that works: it never asks whether an account is unusual, only how much of the movement it accounts for, which has an exact answer.

What the solution found

629 detections across six ways of cutting the book, aggregated to 122 rows, of which 22 are promoted to management. Each carries what it departed from, by how much, what it is estimated to cost, and who should act — and the action follows from the evidence rather than from the measure, so a concession past a configured schedule is renegotiated while the same measure away from its peers is a question.

Management use

What went wrong is not the same as what someone must do, so every issue carries an action and the actions go to different people. A cost that rose and was not passed on needs repricing. A concession past the configured threshold needs renegotiating. Volume leaving an account at an unchanged margin is a demand question, not a pricing one. A mis-billed line needs correcting, and nobody needs to negotiate anything.

On this dataset the most severe row for 5 of 6 seeded problems carries the action the register says was needed. That is the strict reading, and it is the one worth publishing. A weaker question — whether the right action appeared anywhere among the rows for a problem — answers 6 of 6, because five of the six produce more than one action and being right somewhere is not the same as leading with it. The mix shift is the one to read carefully: it is routed to monitor rather than investigate, because a shift between lines is not recovered by chasing anybody.

Methods and technical note

Every accuracy figure on this page is measured against a ground-truth register, not asserted. The dataset is generated with its problems deliberately seeded and recorded, so what the engine found can be checked against what was put there. The register never ships with the results it scores.

180 of 180 mis-billed lines were found, and 6 of 6 segment-level problems. The one problem deliberately not in the data was not reported. 14 of 14 individual segments were identified. The ones missed are the smallest members of the affected sets, whose contribution to the decline is correspondingly small.

Coverage is only half of the question, and it is the flattering half: a detector can find every planted problem by flagging a great deal. The other half is how much irrelevant work the queue creates. Two figures answer it, because they are not the same question. A row lands on a seeded population if its segment carries a registered condition at that row’s grain — 30% of the whole queue does. Issue precision also asks whether the row carries the action that condition needs, which is the question a manager has when they open it, and that is 17%. The gap between them is rows that found the right account and described a different thing about it.

So the queue is tiered, and only one tier is put forward as work. 22 rows reach the management tier — on decisive evidence, or because the route that found them has measured precision of at least 70% at that grain — and they run at 82% issue precision. The other 100 are review signals: explained, checkable, and not presented as something to act on. This page shows the management tier. Calling all 122 of them issues was a promise these same figures disproved.

4 of the 7 measurable routes never fired on a held-out account. The rest are reported as they measure, because a detector that flags clean accounts at the rate they occur in the book is not avoiding them, and saying so is more useful than a figure that implies otherwise.

The median estimate came in at 2.08x the seeded magnitude and the worst at 3.33x, on account cost not passed through. It runs high because a comparison against peers attributes all of a segment’s difference to the problem, when some of it is the segment being genuinely different. That is what the figure is worth, and publishing it beats calling the estimate accurate.

Generated 2026-09-11 from 278,169 invoice lines. Every figure on this page is read from the published artifacts; none is written by hand.