Five Times Return, or Do Not Start

Almost every prioritisation exercise I have been asked to run in the last two years ends the same way: a ranked list. Candidates down one axis, a score across the top, weights agreed in a workshop, and a winner. The list is usually competent. It is also incapable of telling you the one thing you most need to hear, which is that none of these is worth doing.

A ranking always produces a first place. It produces one when every candidate on the list is worth less than the effort of building it, and the organisation goes and builds that first place anyway, because the exercise told it to.

The alternative is a floor, and I have one on paper. It sits inside a proposal we put to a European food and consumer goods manufacturer for a first proof case. The proposal was never converted, so nothing here is a delivered outcome. What it is, is a qualifying test we were prepared to be held to before any work started: a candidate had to plausibly return five times the effort it consumed, or it was not a small win, it was a reason the programme would fail.

The sentence that does the work

The approach page in that proposal is organised as a set of obstacles: for each band of the work, two ways the thing dies before it produces anything. Under the opening band, where the question and the business case get formed, the two entries are these: there is no senior sponsor, and there is no real case identified, with the qualification set in brackets at five times return on effort.

Read those two as a pair, because the pairing is the argument.

Nobody needs persuading about the first one. A programme without a sponsor is dead and everyone in the room knows it; the only debate is how long it takes to notice. The second usually gets treated as a matter of degree: a thin case is a starter case, something to build confidence with. The page refuses that reading. A candidate that cannot plausibly return five times what it costs to build is not filed as a low-scoring option. It is filed under the same heading as having nobody senior who cares, and the heading is the reason your programme does not survive its first year.

That is a much harder thing to write into a proposal than a ranking, because it can disqualify the engagement you are trying to sell.

What the multiplier is measured against

The number is only usable because the thing it divides by is stated plainly on the same document. The commercial page offers two build shapes for that first proof case: a lean team of 1.3 full-time people at ten thousand euro a week, or 2.1 people at twelve and a half thousand a week with a supplier-provided analytics environment plugged into the client's own warehouses at no licence cost for the trial. Both are sized at about eight weeks, with a note that these exercises typically run four to eight weeks where the data environment is established. A scoping session sits in front at a one-time five thousand euro, waived if either build option is taken.

So the denominator is not a licence line or a model cost. It is people multiplied by weeks. Take the smaller shape at its quoted duration: about eighty thousand euro of effort, which puts the qualifying bar near four hundred thousand euro of benefit that a business owner would recognise and sign for. That arithmetic is the whole instrument: crude, done before anyone touches data, and done out loud in front of the person who would have to bank the benefit.

Most candidates do not survive that conversation, which is the point of having it early rather than in week seven.

Why a ranking could not have caught this

The opportunity page of the same proposal carries fourteen candidates across four executive owners, spanning demand prediction, promotional and trade spend, financial planning, asset and supply chain optimisation, customer view and marketing spend. The page marks itself illustrative and subject to change after discussion, which is honest. It was a conversation starter, not an assessment.

Now imagine scoring those fourteen. Impact, feasibility, data availability, weights argued for an hour. You would get a ranking, a top three, a slide that looks like a decision. Nothing in that method can return the answer "none of these", because its output is an ordering and an ordering has a top.

The threshold can return that answer, and something more useful with it: this candidate is fine, the benefit is real, and the benefit is worth about what the work costs, so build it later when the work is cheaper. That rejects the timing rather than the idea, and puts the case back in the queue with a condition attached rather than a score.

5x

Return on effort the proposed test required a candidate to clear

1.3 FTE

Lean team in the smaller offered shape, at ten thousand euro a week

4-6 weeks

Middle band of the offered shape, where the plan concentrates the effort

14

Illustrative candidates tabled, across four executive owners

The denominator is the part you can actually move

What makes the threshold more interesting than a hurdle rate comes out of the effort page rather than the approach page.

That page splits every activity between the supplier side and the client side, and the split is lopsided in a pattern. Getting raw data into a usable state, then reshaping it so the hypotheses can be tested against it, both run at ninety percent supplier. Agreeing which analytics platform to use runs at ninety percent client, as does clearing security before anyone touches the data. Taking the benefit case around the stakeholders is eighty percent client, and tracking whether the benefit lands afterwards is sixty percent client.

One week at the front, four to six weeks in the middle, one week at the end. The middle band is the longest and the most supplier-weighted, and it is almost entirely data work: sourcing, assessing quality, crunching, transforming, validating.

Which tells you where the multiplier is decided. The numerator is a property of the business and you cannot argue with it; a promotional spend case is worth what it is worth. The denominator is a property of your estate: how long it takes to find the data, get permission to use it, understand what the fields mean and get it into a shape a model can read. Every mature feature store, every pipeline that gets reused instead of rebuilt, every documented lineage path, every environment that stands up in a day rather than a quarter, takes weeks out of that middle band.

And when the denominator falls, the floor does not move but the population above it grows. Cases worth about what they cost become worth several times what they cost, with no change to the business case at all. That is the quiet argument for spending on the plumbing, and the reason RealAI sequences Platform work ahead of case selection rather than after it. The plumbing does not produce value. It changes which of your ideas are admissible.

A ranking always produces a first place. It produces one when every candidate on the list is worth less than the effort of building it, and the organisation goes and builds that first place anyway.

The discipline gets harder as building gets cheaper

The obvious reading of falling build cost is that thresholds matter less. If a pipeline takes days instead of months, why qualify anything at all.

I read it the other way. When the build was expensive, the expense did the qualifying for you: nobody spent two quarters on a case they could not defend, because the finance conversation happened whether you wanted it or not. Make the build cheap and that conversation stops happening by itself, and you end up with a portfolio of things built because they were buildable.

The proposal names the failure mode at the far end. Its last band of obstacles is not technical: the organisation lacks the discipline to run continuous benefit reporting, and lacks the habit of improving its fact base afterwards. Not one of the six ways that page says a first proof case dies is about whether the model would work. A floor is only a floor if somebody checks afterwards whether it was cleared, and cheap building removes the natural moment when anyone would have looked.

So the multiple becomes more load-bearing, not less. It is the one part of the method that does not get cheaper when the tooling does.

What I would take from it

Write the bar down before you see the candidates, so it cannot be tuned to admit whatever somebody has already decided to build. Express it against effort in people and weeks rather than a licence price, because effort is what your tooling reduces. Say the number out loud to whoever owns the benefit. Let the exercise return "nothing here qualifies yet", and treat that as a finding about your data estate rather than a failure of imagination. Then shorten the middle band and run the list again.

That last step is the one most organisations skip, and it is the one that changes the answer.

Drawn from a pre-sales proposal put to a European food and consumer goods manufacturer for a first proof case: its approach page, its effort split, its illustrative opportunity list and its commercial options. The engagement was offered, not delivered, so the threshold here is a qualifying test we proposed and no client result is claimed. Reading that threshold as a floor rather than a ranking is ours.