Diagnosis, Then a Recommendation

A diagnosis is a mechanism you could intervene on. It is not a restatement of the symptom with more decimal places.

“Dealer inventory rose 14 percent while fill fell 2 points. Item-level variance is high. Forecast error is 31 percent. Expedites are up.” Those are observations, and they may all be true. “The replenishment rule treats a water pump and a floor mat as the same kind of uncertainty, so we hold the wrong cash and still miss the jobs that stop a vehicle” is a diagnosis, because it names a mechanism with a location, an owner, and a plausible change.

The test is whether the sentence implies an intervention. If it does not, you have measured rather than diagnosed, and any recommendation attached to it is a guess in professional clothing. The distinction between an observation and a diagnosis is drawn early in the book, in Chapter 4; at the point of writing it decides what can sit beneath the recommendation.

How to produce a mechanism, which is the part nobody teaches

Saying that a diagnosis must be structural is easy. Producing one is not, and it is not done by staring at the observations until a mechanism appears.

The move that works is to invert the question. Instead of asking what the observations mean, ask what would have to be true of the process for these observations to be its normal output.

Take three findings that most teams report as three findings: inventory rose, fill fell, and expedites are up. Treat them instead as symptoms of one machine. A rule holding too little of everything gives you low fill and low inventory. A rule holding too much of everything gives you high inventory and high fill. Getting high inventory and poor fill at once requires a rule that is holding the wrong things, which is a statement about the incumbent rule rather than about demand. The rising expedites are then not a fourth finding but a consequence: the organization is buying its way out of the mismatch one air shipment at a time.

Notice what that did. It converted four observations into one sentence, and the sentence names something somebody could change.

A diagnosis makes a prediction, which is what separates it from a story

The mechanism above implies something checkable. If the rule is holding the wrong things, then the items driving the misses and the items holding the cash should be largely different sets. If they turn out to be the same set, the diagnosis is wrong and a class-based rule cannot help.

That is the property to insist on. A diagnosis hands you a specific way to be embarrassed. A story does not.

Two habits make this easier. Prefer mechanisms that live in a rule, a constraint, or an incentive, because an owner can change those; mechanisms that live in culture or alignment usually restate the problem in a softer register. And check the mechanism with somebody who runs the process, who will usually tell you within minutes which part of it you described wrongly. When two mechanisms fit the evidence equally well and lead to the same recommendation, say so and stop, because a diagnostic question that cannot change the pick does not need resolving.

The recommendation has two layers, and gluing them together is the common error

The first layer is the approach: treat items by their economics rather than by a uniform target. The second is the set of actions that make that approach real this cycle: three classes, an owner, a priced exception path.

The approach without the actions is a slogan. The actions without the approach are a portfolio of initiatives that will spend the year undoing one another.

Write the approach as a single sentence, then ask of every proposed action whether it follows from the same mechanism. An action that cannot be traced to it, such as a dashboard added because someone asked for one, does not belong in the recommendation. Coherence is the test that the diagnosis did some work. Without it the room can like the actions and still not know what they are for, so the first objection that arrives, a dashboard or a fill target or a forecast, gets added as a friendly amendment.

The objective also has to be reachable, not merely correct. The exception path is unpriced, therefore the organization should install an end-to-end cost-of-service model: true, and not a decision anyone can take this quarter. Pricing expedites for one region with a crude standard cost is a worse model and a better decision than pricing every exception correctly in eighteen months, because the first will exist this year and will make the second easier.

A missing diagnosis shows on the page

It shows in three ways: abstract language doing the work of a concrete claim, a target restated as though it were a plan, and silence about the one thing that makes the decision hard.

The progression applies well outside inventory. “The delay model has good discrimination” is an observation; “late parcels cluster on a few lanes the current expedite heuristic cannot see” is a diagnosis; “add a delay score to that heuristic above a named threshold” is a recommendation. “The experiment is significant” is an observation; “the treatment works on new customers and does nothing on the base” is a diagnosis; “roll out to new customers only, with a kill switch” is something an owner can install.

One ordering note. Diagnosis comes before the recommendation in the work. In the write-up, the recommendation comes first and the diagnosis sits beneath it as its leading reason. Work backward, present forward.

Make an actual recommendation

The recommendation is a single sentence that completes a thought, and it has a recognizable shape.

We recommend [action], for [scope], starting [when]. We expect [consequence]. The major tradeoff is [this]. We would change our minds if [this].

At Meridian Motors: we recommend replacing the single replenishment rule with a three-class stocking rule for dealer-facing items, starting at the next planning freeze, owned by Parts Planning. We expect lower working capital on the cheap class and fewer missed jobs on the protected class. The major tradeoff is operational complexity and a fight with anyone who wanted a uniform 20 percent cut. We would change our minds if dealer bypass of the protected class cannot be closed, or if missed-job costs are far more uniform across items than last year’s traces show.

Each clause does work. Scope belongs in the sentence because the ranking often depends on it: a class-based rule at two central distribution centers, extended if bypass stays inside a trigger, is a different commitment from a company-wide change. The tradeoff belongs there because an owner who discovers the cost later will distrust the next recommendation. The last clause is what separates a recommendation under uncertainty from a claim of certainty, and it is where hedging belongs, attached to the uncertainty that would change the ranking rather than spread across every sentence. Confidence, where you express it, is operational: we are willing to run this for ninety days with a kill switch, and we would not stake the cash event on it.

Why teams refuse to recommend

The refusal is rarely stated. It shows up as a slide titled “Options” with three balanced columns, or a conclusion that “the data suggests” something without saying what to do. Teams that can write the sentence privately decline to put it on the page because it can be rejected in a way a survey of results cannot, because they may be wrong, and because it may damage someone’s scorecard.

All three are honest, and none is an argument. Withholding the sentence does not lower the chance of being wrong. It lowers only the chance that the process changes, which means the incumbent wins by default. Uncertainty is handled by stating the conditions that would reverse the pick, which is more honest than a balanced menu, not less, because a menu transfers the judgment to someone with less information than you. A damaged scorecard is handled by naming the damage as the tradeoff. And recommending is not deciding: the owner decides, and withholding a recommendation deprives them of the input they asked for.

Recommending the incumbent

Keeping the current rule is a legitimate recommendation and it deserves the same sentence. The current rule may suit economics flatter than anyone assumed, the challenger may lose its advantage to the exception path, or installing it may cost more than it earns in the period anyone can see.

It needs more care than a change, not less. The incumbent usually has no written description, so recommending it means documenting it as it actually runs, informal expedite path included, and that document is occasionally the most valuable thing the project produces. A recommendation to change nothing is easily heard as permission to stop paying attention, so it needs an owner, a condition under which the answer would change, and a date for the next look. The condition also exposes the dishonest version. An author who can say precisely what would justify a change is recommending the incumbent on the evidence. An author who cannot is avoiding a fight.

The test

Write the top sentence of your current write-up in the shape above, all four parts. Then read it to someone and ask whether they could disagree with it and say why. If they could not, you have a finding, not a recommendation. If you cannot fill the last clause, go back to the diagnosis: the sentence cannot be written yet, and the problem lies upstream of the writing.