Topline and Bottom Line in the Same Programme

The decision that settles whether a data programme survives its second year is usually taken in the executive summary, and at the time it does not look like a decision at all. It looks like scope.

I have a proposal in front of me, put to a large European retail banking group, that takes that decision cleanly and then declines to finish it. The offer is described in the document's own words as two-fold. On one side, data-driven analytics support, aimed in the document's own phrase at both topline growth and bottom-line efficiencies. On the other, an operational excellence programme taking cost, possibly a larger chunk of it, out of a newly consolidated group servicing centre for payments and loan processing. Growth and efficiency, named in the same paragraph, put to the finance and operations sides of the house rather than to anyone in technology.

Then the sentence that decides the whole thing. The cost half is to be proposed in a separate document. That document is not in the pack.

Two mandates in one paragraph

Take the pairing on its own terms first, because it is right, and most programmes I review do not attempt it.

An efficiency-only programme runs out of political air. Every unit of saving it delivers is taken from somebody's line, and the coalition that has to fund year two is the coalition the programme spent year one shrinking. By the second cycle sponsors are defending headcount rather than volunteering processes.

A growth-only programme runs out of patience instead. Revenue attribution arrives late and arrives contested, the sponsor has an eighteen-month attention span, and the board asks what changed. Nothing has changed yet that anyone can point at, so the programme is asked for an interim number, and the only interim numbers available are activity counts.

Running both at once is a hedge against both failure modes. The cost half funds the growth half and buys it time; the growth half gives the cost half a reason to exist beyond subtraction. That reading is mine, not the document's. What the document supplies is the structure, and the structure is unusually clear about it.

The half that can measure itself

The two halves are not symmetrically hard, and that is the part nobody prices.

A servicing centre that has recently been consolidated is close to the most measurable object in a bank. Consolidation itself forced the instrumentation: volumes had to be counted to size the centre, handling times had to be measured to staff it, unit costs had to be derived to justify it. Cost takeout therefore lands in a currency the finance function already keeps, against a baseline that already exists, in a ledger nobody has to be persuaded to open.

Growth has no such luck. Every incremental revenue claim is a claim about a counterfactual: what this customer would have done had the model never scored them. Answering it needs a holdout the commercial teams agree not to touch, a baseline frozen before the intervention rather than reconstructed after it, and a written definition of incremental. None of that is technically hard. All of it is politically expensive, and all of it has to be arranged before the work starts, when nobody wants to spend the negotiating capital.

So one workstream shows up to the review with a meter already attached, and the other shows up with a request for one. Which is why deferring the cost half to a separate proposal is not the administrative act it appears to be.

Two documents means two decisions

One document produces one decision, in one meeting, with both halves on the same page and one person accountable for the comparison. Two documents produce two decisions, on two dates, in front of two review cycles, quite possibly with two sponsors. Whichever half can produce a number produces it first, and that is not the same as whichever document arrives first. The document in front of this buyer is the growth half, and the number, when it comes, will come from the half still to be written. Once one half has a signed-off benefit case the other is no longer a peer. It is an unfunded ambition sitting beside a funded plan.

I do not think anyone intended that here. The deferral reads like an ordinary scoping choice: different discipline, different authors, another week. That is precisely how it happens. Nobody decides that growth loses. Growth loses because it was asked to wait, and the thing it was waiting behind arrived quantified.

What the pack admits about itself

The document is honest by accident, which is the most useful kind.

The page carrying the market narrative has no argument on it at all, only a note assigning me to describe it tomorrow, pasted over and over until it filled the box. Its commercial page sets out two proof cases, each carrying a fifty-fifty investment split, and both case descriptions carry the same placeholder wording and the same case number, so what looks like a two-case construct is one case pasted into a layout and not yet edited. Where the return should be quantified sits a question addressed to nobody in particular about whether the firm takes fees or a share of the cash, and at what ratio. It sits there twice, once per case.

None of this is a delivered outcome. It is a proposal, unfinished in the copy that survives, and I read it as calibration rather than as an indictment. The polished artefact a buyer eventually sees is a short interval of work away from a page that reads "what ratio?". Anyone evaluating a proposal should ask which pages were written last, because those are the pages where the thinking had not yet closed.

Here the page still open was the commercial page. Which is to say: the price of the growth half, and the split of its upside.

The same shape, with agents in it

The pairing question has not aged. It has acquired a harder version of itself.

The current two-fold pitch in a lending operation looks like this. On the efficiency side, copilots and constrained agent loops inside the servicing centre: document extraction on loan files, first-line query handling, exception triage. On the growth side, an agentic layer over the customer data that scores signals and proposes a next action. Same programme, same two buyers, and a worse asymmetry than before.

The efficiency half instruments itself by construction. An agent harness logs every step it takes, every tool it calls, every hand-back to a human, so handling time and containment rate fall out of the trace without anybody designing a measurement. The growth half still needs an evaluation set, a holdout and an agreed baseline, and those remain somebody's problem for next quarter.

Graded autonomy then pulls the two apart on timing. Exception triage inside an operations centre can run at higher autonomy sooner, because its errors are cheap, visible and reversible. A model that shapes a credit decision or a customer offer cannot.

The regulatory load is not evenly distributed either. With the EU AI Act in force, an internal productivity agent and a system feeding creditworthiness assessment do not sit in the same tier, and the second acquires documentation, human oversight, logging and monitoring obligations that the first largely does not. Model risk management functions already understood this; programme sponsors frequently do not. The growth half is not only harder to measure, it is slower to ship, and the two effects compound.

Which means the drift that this old proposal only risked is now close to the default. Left alone, an agentic programme sold as growth and efficiency becomes an efficiency programme with a growth appendix, and the appendix is quietly re-planned every quarter until it disappears.

Nobody decides that growth loses. Growth loses because it was asked to wait, and the thing it was waiting behind arrived quantified.

What to insist on before either half starts

Five things, none of which needs a data scientist, all of which have to land before one is worth hiring.

One document covering both halves, with one baseline date, so the two benefit cases are computed against the same starting picture. One owner accountable for both numbers, senior enough to be asked why one moved and the other did not. The growth measurement designed and instrumented before the efficiency half ships, because once agents are running in the servicing centre the population is contaminated and the counterfactual is gone. A holdout the efficiency programme is contractually not allowed to touch. And a standing rule that neither half goes to a steering committee alone, because a half presented alone is a half that has stopped being compared.

That last one costs nothing and does most of the work. It is also the first thing dropped when a calendar gets tight.

We instrument Agentic OS this way for exactly that reason: the trace proving an agent saved handling time and the evaluation set proving a recommendation changed a customer decision are the same infrastructure problem. Solve it once and you can afford both mandates. Solve it only where it is easy and you have already chosen, and you will find out which half you chose about three quarters later.

2

Workstreams named in the executive summary, growth and cost takeout

1

Halves actually present in the pack, the other deferred to a separate document

50/50

Proposed investment split per proof case, return split left as an open question

4

Further opportunity areas named and explicitly unscoped

Running growth and cost reduction together is the correct instinct, and it is rarer than it should be. The failure is not in the pairing. It is in what happens once one half is allowed to arrive with arithmetic and the other with intent, and the imbalance is then read as evidence about which was the better idea.

The structure here, and the one figure quoted, are as recorded in a data-monetisation proposal put to a large European retail banking group: its executive summary, its commercial page, and its appendix of opportunity areas. That document produced recommendations rather than delivered results, and the copy I worked from is an unfinished draft. The deferral is not an artefact of that draft: the executive summary states twice that the cost half will be a separate proposal, so it is a scoping decision the document made rather than a page nobody got to. Reading that decision as a structural asymmetry rather than an administrative one is mine.