Method note 03

Pipeline supply

Optimization is easy. The pipeline is not.

A portfolio committee will spend six weeks this quarter deciding, with considerable rigor, between the things that happened to turn up.

It will weight its criteria. It will run scenarios. In the better organizations it will solve the selection properly, to provable optimality, with a defensible objective and an audit trail. And it will do all of this over a set of candidates that nobody designed, nobody commissioned, and nobody has looked at as a whole.

Here is one such set. Sixty-seven initiatives, $16.0m, assembled the usual way — a submission window, a form, and whoever chose to fill it in. The organization had already declared six strategic drivers and weighted them with some care.

The two priorities carrying 73 percent of the declared weight attracted 37 percent of the proposed spend. Two carrying 17 percent between them attracted 60 percent. And security compliance, declared at 2 percent — small, but declared, which means somebody argued for it and won — had nothing behind it at all. Not an underfunded initiative. Not a small one. Zero. It existed in a spreadsheet.

The organization declared what mattered and then funded what turned up. The two lists have almost nothing to do with each other, and no one noticed, because nothing in the process looks.

Maximizing a garbage can

No optimizer improves this. Solve it exactly, solve it with duals, solve it under whatever objective you prefer, and the answer will be the best available subset of a candidate set that is close to inverted relative to the strategy. A driver with no initiatives cannot be served at any budget. The mathematics has nothing to work with, and it will not say so — it will return an answer.

That is the trouble. An optimum over the wrong set is still an optimum. It arrives clean, reproducible, defensible under questioning, and beside the point. Nothing about it looks wrong, which is precisely why nobody checks.

The arithmetic underneath is not subtle. Optimization returns a maximum over a set. Improving the optimizer moves you toward the best element of the set you already have. Improving the set moves the maximum itself. The first is bounded by the second, and the second is almost always the larger number.

The field has spent thirty years on the first. Better scoring, better weighting, better workflow, better algorithms, better dashboards. Enormous ingenuity applied to choosing well from whatever is in the bag, and next to none applied to what goes in the bag. We have refined the step where the least value was available and left untouched the step where the most was.

Under the streetlight

Why did a serious discipline spend its ingenuity on the smaller of the two numbers? Not carelessness. Because it could.

Selection is a well-posed problem. It has inputs, a constraint set, an objective and a provably best answer. It can be automated, benchmarked, demonstrated in a sales meeting, published in a journal, and improved in measurable increments. Every one of those properties makes it attractive to work on.

The pipeline has none of them. Improving it means going to the parts of the organization that proposed nothing and asking why — slow, political work that yields no artifact, no benchmark, no demonstration, and no algorithm. There is no version of it you can put on a slide with a number attached.

So the profession looked where the light was good. Thirty years of genuine ingenuity, applied to the half of the problem that would hold still, while the half that determined the answer was left to whoever felt like filling in a form.

I write this as someone who built the solver. My own implementation solves the selection exactly, to provable optimality, with duals on the binding constraints — and I am telling you it is the smaller lever. Getting the selection right is necessary. It is nowhere near sufficient, and the effort the field has put into it stands in almost inverse proportion to the value it can add.

Intake is a word for waiting

The industry's own vocabulary gives it away. The process begins with intake — a form, a deadline, a submission window. Intake is something you receive.

So proposals arrive from the parts of the organization that have the capacity to write proposals, from sponsors who understand how the process works, and from functions that submitted last year and expect to submit again. The candidate set is shaped by the distribution of proposal-writing energy, which has no reason whatever to resemble the distribution of strategic need.

That is how system availability draws 27 percent of proposed spend against a 4 percent weight: infrastructure teams are practiced at proposing infrastructure work. It is how customer onboarding speed draws 7 percent against 25: whoever would own that work is presumably busy doing it. Nobody decided any of this. Nobody would defend it if it were put to them in these terms. It is simply what arrives when you wait.

And the gap is invisible by construction, because the reporting starts after selection. A portfolio review shows what was funded, sometimes what was rejected, and never what was never proposed. The most consequential absence in the entire process is the one thing the process does not record.

What the comparison cannot tell you

The measurement is easy. Reading it is not, and a coverage figure presented as a verdict would be worse than no figure at all.

A driver with declared weight and nothing behind it admits at least four readings. The organization may be neglecting something it said mattered. The driver may have been aspirational, argued into the strategy without anyone intending to fund it. The work may already be happening outside the project portfolio, in operations or run-rate spending that never appears as an initiative. Or the declared weight is simply wrong, and the absence of proposals is the organization telling you so.

Those are not variations on one finding. They call for opposite responses, and nothing in the comparison distinguishes them. The inversions are no easier: a driver drawing more spend than its weight suggests may be over-served, or expensive to serve, or carrying work that serves other drivers and is credited only to itself.

Attribution is its own difficulty. An initiative rarely serves one driver, and splitting its cost across several — evenly, by stated impact, or not at all — changes the picture materially. Any of those conventions can be defended. None is a fact.

So the figure supports a question, asked earlier than it usually is: is there anything behind what we said mattered, and if not, why not? It does not support a score, a rating, or a compliance percentage. Those would give the appearance of an answer to something the data cannot settle, which is the error this note is about.

Nor can ideas be commanded. Direction can be set and attention pointed, but generation happens in people's heads, outside the system, and a process that pretends otherwise collects compliance rather than ideas.

The narrow claim is this: the value available from a portfolio decision is fixed by the candidate set long before any algorithm touches it; that bound is visible using data every organization already holds; and a discipline that has spent its attention on the selection step has been optimizing the smaller of two numbers.

A great pipeline with an average selection beats a poor pipeline with a perfect one. The reverse is not close.

Cite as: Gruia, M. (2026). Pipeline Supply: Optimization Is Easy, the Pipeline Is Not. Frontnovo Method Notes.