When the Cost Programme Runs Out of Road

There is a kind of meeting I have learned to recognise before anyone speaks. Finance has the numbers, operations has the process maps, and everyone already agrees the obvious work is finished. Nobody argues about consolidating a back office, because it was consolidated years ago. The argument is about what comes next, and nobody in the room has a good answer.

I keep a proposal in my files from a decade back that records that moment better than anything I could reconstruct now. It went to a large European retail banking group. I was on the team that wrote it, carrying quality assurance rather than delivery, and one page of it was mine and was still an empty placeholder in the version I kept. So this is not a story about what we achieved there. It is about what the client had already achieved before we arrived, and the question that achievement left on the table.

The number belongs to the bank

The client-context page carries more figures about the client than any other page in the document, and every one of them describes the group's own position rather than anything we did. One line carries the argument: cost/income down from ninety percent to fifty-three, across all business units and functions, given as the reason the low-risk strategy was working. Thirty-seven points. The document credits nobody outside the bank, and neither will I.

The same page records the route, which matters because most large institutions took it. Non-core mid-office and back-office activity was shed. The asset management arm went to a buyer whose business is running asset management. Insurance was partnered rather than built. Physical reach ran through a distribution partner, where the bank met customers at counters it did not own.

Read as a cost story, that page is close to a model answer. Read as anything else, it is a series of doors closing quietly behind a good result. Divestment, partnering and rented distribution are one-shot levers. You pull each once, book the improvement, and arrive at a good cost line holding fewer instruments than you started with. The page is headed as strategy and targets, yet every figure printed on it is historic, and the forward-looking half is stated in words rather than numbers.

I should be exact about what the document says and what it does not. It nowhere states that the cost programme had run out of road, and nowhere connects the cost history to the reason we were in the room. That connection is mine, and should be weighed as an argument, not a finding. What the document supplies instead is an address line. It was written for the group's chief financial officer and its chief operating officer, and for no technology executive at all. A data programme sold to the two people who own cost and capital, framed as monetization rather than as infrastructure, was written by people who knew where the pressure was.

Every cost decision was also a data decision

The very next page argues that heavy prior investment in core banking now lets the bank capture all the data it historically did not, and that it should mine and monetize it. Put that page beside the cost page and they pull against each other.

The cost page is a record of giving away the surfaces where customers transact. The counter belongs to the distribution partner. The fund belongs to the buyer of the asset management arm. The policy belongs to the insurance partner. Behaviour generated at those surfaces sits inside somebody else's operating system, under somebody else's contract, and it does not arrive in your warehouse because you want it to. Both claims cannot be fully true at once.

I do not think anyone was being dishonest: two pages written by different people about different questions is how the mistake gets made. Every efficiency decision of that period was also a data decision, and nobody costed the data half, because nobody yet had a use for it that survived a business case. That is no reason to regret the outsourcing. It is a reason to write data terms into the next contract that touches a customer surface, which almost nobody does at renewal.

What the proposal actually proposed

Honesty about tense matters. This document is a pitch: a recommendation and a method, not a result, and the file I kept was a working draft. One page is a headline over a placeholder repeated down the body. The commercial page holds two proof cases, each marked as an even split of the investment between firm and bank, and where the pay-off should sit there is a question the firm is asking itself in front of the client about whether it takes fees or a share of the cash, and at what ratio.

I find that page more useful than the polished ones. Outcome-based pricing was already the ambition, and it stalled where it still stalls: neither side can agree what would have happened anyway. Splitting the cost of a first proof case is arithmetic. Splitting the benefit needs a measurement baseline both parties sign before the work starts, not after the result arrives.

The method page held up better. Five lettered steps over three phases, with two gate markers whose placement the file no longer preserves, and the lettering is the part to keep: a business question first, data preparation only second, then analysis, then validation, then a named final step for tracking whether the benefit landed. Question before data, and benefit tracking inside the method rather than in a business case nobody revisits.

90% to 53%

Cost/income ratio, the client's own prior achievement

5

Steps in the proposed method

2

Gate markers on the method diagram

50/50

Investment split per proof case, return split unresolved

The next unit of margin lives inside decisions

A decade-old document is worth reading now because the position it describes is where most large institutions sit today. Offshoring happened. Consolidation happened. Process automation took the deterministic work. What is left is a long tail of decisions too variable for a rule and too numerous for a specialist, and that tail is where the remaining margin sits.

The opportunity pages of that proposal, written for four executives, point at that tail without having the words for it. Measure the distribution of lead times through an operation rather than the average, because the average says nothing about the tail you are paying for. Score every customer signal, online and offline, and route it. Model redemption behaviour rather than the contractual schedule, and let treasury fund against the behaviour. Reframe collections as helping a customer who intends to pay find a way to, which is a segmentation problem before it is an automation one.

Each is now within reach of an agentic system, and each is a reason for care. An agent loop that reads a case file, retrieves the policy governing it, proposes a decision and acts on it changes the cost of being wrong. A badly scored lead used to cost a wasted call; now it can cost an unsupervised customer interaction. That argues for graded autonomy rather than against automation: propose where the consequence is reversible, act where the evaluation set says it holds, keep a person in the loop where it is not. It also argues for treating agent harnesses the way a bank already treats a scorecard, under model risk management, with a named owner.

The regime has caught up too. That proposal cheerfully suggested feeding retailer and social signals into acceptance criteria. Under the EU AI Act, creditworthiness assessment sits in the high-risk tier, and the same idea now arrives carrying lawful basis, contestability and a documented case for why a proxy is not doing work the model should not. The technique survived. The burden of proof around it did not.

Divestment, partnering and rented distribution are one-shot levers. You pull each once, book the improvement, and arrive at a good cost line holding fewer instruments than you started with.

Count the proof, not the promise

One last thing that document taught me, and it applies to how AI is sold now. The opportunity set is spread beautifully: two ideas each for four executives, eight in all, across operations, commercial, finance and risk. The track record is not spread the same way. Thirteen delivery cases sit in the appendix, and most of them are process and IT analytics: root causes behind repeat failures, usage logs for a system nobody had measured, transaction capacity ahead of a payments deadline, bottlenecks in a digital journey. Solid work, and several of them carry hard savings figures. Only two of the thirteen sit on the revenue side the proposal was actually selling, both on the same lending product at an unnamed retail bank, and neither of those two carries a number.

So the promise was topline and the proof was cost. That is the standard shape of a proposal, and the standard shape of an AI vendor deck today: the use cases fan out across the org chart, the evidence clusters wherever the firm has already been paid to work, and nobody is expected to check whether the two overlap. Ask the seller which of the ideas pitched to you has run in production on the side of the business you are being sold, with a number attached.

An organisation that has taken its cost ratio a long way down has earned a harder question than where to cut next. The move worth making finds margin inside decisions the business has never been able to make consistently, and getting there is measurement, ownership and graded autonomy rather than a platform purchase. That is what RealAI's Platform team is for once the direction is set, and it starts by naming the decision you want to move.

Figures are as recorded in a data analytics proposal submitted to a large European retail banking group: its client-context page, its problem statement, its method and effort pages, and its case appendix. The cost/income improvement is stated there as the client's own achievement, not a consulting outcome, and the document produced recommendations rather than delivered results. Reading the cost history as the reason the proposal existed is mine.