1. What a Stage-Gate Actually Is
A stage-gate model splits a long, uncertain piece of work into stages, and puts a decision point between each one. At each decision point somebody with authority looks at what is now known and decides whether to fund the next stage. That is the entire idea.
It is worth being precise about what that means, because the phrase gets used loosely. A gate is not a milestone. A milestone is a date on which something is expected to be finished; passing one is an observation. A gate is a decision, and a decision that can only go one way is not a decision. If the answer at a gate was never realistically going to be anything other than “continue”, the gate did not exist — a meeting happened.
Nor is a gate a quality review. Quality reviews ask whether the work was done properly. A gate asks whether the work should continue at all, which is a different question with a different answer. Work can be flawless and the program still worth stopping, because the world moved or because what the work revealed was unwelcome.
The model is old and its origins are industrial — it comes from new product development, where somebody noticed that most of the money in a failed product launch is spent after the point at which the failure was already knowable. The remedy is not better forecasting. It is spending the money later.
2. Why the Gates Sit Where They Do
The single most common mistake in applying the model is putting gates at calendar intervals — quarterly, or at the end of each financial year. That produces decision points where nothing new is known, and no decision point at the moments something is.
Gates belong where uncertainty resolves. Not where the calendar turns, not where a phase happens to end, and not where a governance body happens to meet. In this program that principle puts them here:
| Gate | Why here |
|---|---|
| Gate 1 — after candidate selection | The first point at which one molecule can be preferred over others on evidence. |
| Gate 2 — before the IND is filed | ⚠ Filing an IND is irreversible in reputational terms and starts a regulatory relationship that persists. It is also the last cheap moment. |
| Gate 3 — after Phase 1, before Phase 2 | Human safety and pharmacokinetics are known. The dose question is not. |
| Gate 4 — after Phase 2, before Phase 3 | ⚠⚠ THE decisive gate. Phase 2 resolves dose and signal; Phase 3 costs 58% of the program. *The largest commitment is made at the point of best evidence, which is the entire argument for the model.* |
| Gate 5 — before submission | Everything is known that will be known before the agency's view. The last point at which the sponsor decides anything alone. |
| Gate 6 — before launch | Approval is a fact; launch is still a decision, and they are not the same thing. |
Gate 4 is the one that carries the argument. Phase 2 exists to resolve two questions — what dose, and is there a signal — and Phase 3 costs 58% of the program. Putting the largest commitment immediately after the evidence that most reduces uncertainty is the whole model in one placement. Move that gate six months earlier and you are committing $125,800,000 on a guess; move it six months later and you have spent much of it anyway.
The corollary is that a gate in the wrong place is worse than no gate, because it consumes the credibility that a real one would need. People learn quickly which meetings decide things.
3. What a Gate Can Decide
Five outcomes, and the middle three are the ones that make the model work.
GO releases the next tranche and the program continues. KILL stops it, and the work already done is written off — not recovered, not repurposed, written off, which is why it is hard to say and why it has to remain sayable. HOLD pauses without deciding, usually while a specific piece of information is obtained; it is legitimate and it is frequently abused as a way of avoiding a KILL. RECYCLE sends the work back into the stage it came from to fix a specific deficiency, then returns to the same gate. And GO WITH CONDITIONS proceeds while requiring named things to be true by a named date.
That last one is where most stage-gate systems quietly fail, and this program used it — Gate 4 carried five in favor, none against, one abstention, with three conditions attached. A conditional GO is honest when the conditions are genuinely closeable and somebody owns each one. It is dishonest when it is a KILL that nobody wanted to say out loud, dressed as a qualified yes, with conditions written vaguely enough that they can later be declared met.
The test is simple and worth applying every time: can this condition be closed, by this person, by this date, in a way somebody else could verify? If not, the Committee has not attached a condition. It has recorded a worry.
4. Why the Money Is Released in Tranches
A gate with no money attached is advice. The tranche is what makes it a decision, and it is the mechanism people most often leave out when they adopt the vocabulary without the substance.
In this program the Board authorized $243,040,000 and released it in five tranches, the largest of which — $125,800,000 — was not released until Gate 4. At Gate 0 the company was not committing $243,040,000. It was committing $17,400,000 and buying the right to decide again later, with better information.
That distinction matters enormously to how the arithmetic reads. The risk-adjusted expected value of this program at Gate 0 was $96,000,000, against an authorized cost of $243,040,000. Read as a single commitment that says do not proceed. Read as a staged commitment it says something quite different: at each stage, the cost of the next stage was justified by the value of the information it would produce. The company was not buying a product. It was buying the option to keep going, and options are worth more when the underlying is uncertain.
The enforcement is unglamorous and essential. If work can begin before the tranche is released, the gate is retrospective — the Committee is asked to approve something already underway, and the honest answer becomes very expensive to give. The tranche is not an accounting convention. It is the thing that keeps the decision in front of the work.
5. Six Ways a Stage-Gate System Fails
Every one of these preserves the appearance of governance while removing its substance, and every one is common.
| Failure | What it looks like | The remedy |
|---|---|---|
| The gate everyone knows will pass | Criteria written after the evidence, or so loosely that any evidence satisfies them. | ⚠ Set the criteria at the PREVIOUS gate, before anyone knows what the evidence will say. |
| The gate held after the money is spent | Work continues into the next stage while the gate is 'being scheduled'. The decision is then about whether to waste what has already been committed. | The tranche is the enforcement. If funds are not released, work cannot start. |
| The conditional GO used as a soft no | ⚠⚠ Conditions attached because the Committee is uncomfortable but unwilling to say KILL. The program proceeds and the conditions decorate the minutes. | Every condition needs an owner, a due date and a named verification method — and a condition that is not closeable is a KILL that somebody was too polite to say. |
| The gate with no KILL in living memory | A body that has never killed anything is not deciding; it is processing. | ⚠ Record the KILL rate. A portfolio where nothing dies at Gate 0 is queueing, not selecting. |
| Criteria that measure activity, not evidence | 'Phase 2 complete' is an activity. 'Phase 2 demonstrated a dose with acceptable tolerability' is evidence. | Write criteria as statements about what is now KNOWN, not about what has been DONE. |
| The gate that reviews the plan instead of the uncertainty | ⚠ Reviewing a Gantt chart at a gate is comfortable and answers nothing. The gate exists because something was unknown at the last one. | Ask what we know now that we did not know then, and whether it changes the case. |
Criteria written after the evidence, and gates that review the plan rather than the uncertainty, both produce a meeting that cannot surprise anybody. The defense against both is to set the criteria for gate N+1 at gate N — before anyone knows what the evidence will say, and while it is still easy to be demanding.
This program does that, and it is visible in the Gate 5 criteria: they were set at Gate 4, twenty- four months before the gate, and the readiness assessment records that more criteria cannot yet be assessed than can. That is uncomfortable and correct.
6. Where the Model Is the Wrong Choice
A methodology guide that never says when not to use the method is advertising. Stage-gate is badly suited to at least four situations.
| Situation | Why the model does not fit |
|---|---|
| Work where uncertainty resolves continuously | ⚠ Software delivery, most operational change. Value arrives incrementally and a quarterly gate just adds latency to a feedback loop that is already working. |
| Work with a short, cheap failure mode | If finding out costs a week, run the experiment. Stage-gating is a way of buying information in expensive chunks, and it earns its overhead only when the chunks are big. |
| Work where the sequence is not forced | Stage-gate assumes stage N+1 genuinely depends on stage N. Where activities can proceed in parallel, gates impose a false serialization. |
| Anything where the honest answer at every gate is GO | ⚠⚠ If a KILL was never realistically available, the gates were theatre and the documentation cost was paid for nothing. |
The common thread is that stage-gate is a way of buying information in expensive chunks. It earns its considerable overhead only when the chunks really are expensive, when the sequence is genuinely forced, and when stopping is genuinely available. Drug development satisfies all three about as completely as any activity does: Phase 3 cannot start before Phase 2 finishes, it costs 58% of the program, and roughly half of the programs that reach it do not succeed.
Software delivery satisfies almost none of them, which is why importing stage-gate into it produces the worst of both — the latency of a heavyweight process and none of the protection, because the decision it defers was one that could have been made cheaply next week. Applying the model where it does not fit is not conservative. It is expensive in a way that is difficult to see, because the cost is latency and the benefit is imaginary.
7. What This Costs
Governance is not free and the honest accounting is rarely done. Each gate in this program consumes roughly six weeks of preparation across the functions, a Committee day, and the standing overhead of maintaining criteria, registers and evidence packages between gates.
Against that, the model's return is entirely contingent: it pays only when a gate produces a decision that would not otherwise have been taken. A program that would have proceeded anyway paid the full cost of the governance and received none of its value. That is not an argument against the model — it is the reason the KILL rate at portfolio level is the number worth watching, because it is the only evidence the system is doing anything.
This program has recorded no KILL, which is unremarkable for a single asset that survived. ⚠ The number that matters is portfolio-level: of candidates entering Gate 0, how many reach Gate 6?
There is a second cost, harder to see. A gate concentrates authority at a moment, and things that are not on the criteria list at that moment tend not to be discussed. This program's market access weakness was visible for two years and reached a gate only once somebody wrote a criterion for it. The remedy is not more gates; it is better criteria, and a standing willingness to add one when a program discovers it has been measuring the wrong thing.
8. Reading the Rest of This Suite
If you have come to this guide first, the four artifacts that will make most sense next are the charter, which authorizes the whole thing; the governance model, which says who decides what; the gate decision framework, which sets out how a gate is actually run; and the program story, which is the narrative of what happened when all of that met reality.
The gate decision records for Gate 5 and Gate 6 are worth reading as pairs with the readiness assessments that preceded them, because the interesting part of a gate is the distance between what was expected and what was tabled.
And one framing worth carrying into all of it: this program was executed well and missed its business case. Nothing in the suite describes a serious execution failure. The gates were run properly, the evidence was generated, the submission was clean, the approval came early. The commercial outcome was decided by a set of choices made years earlier, in documents that were not thinking about commercial outcomes. Stage-gate governance is very good at ensuring decisions are taken deliberately. It cannot make them right.