The Future of Banking Growth in an AI World

Start with the industry picture, because it frames everything else. McKinsey's Global Banking Annual Review has tracked an industry that is generating record absolute profit while its return profile softens. Net income for the global banking industry has been running at around 1.3 trillion dollars, and yet return on tangible equity has been easing rather than climbing, with the review pointing to a decline of roughly half a percentage point between 2024 and 2025, from approximately 12.4% to 11.8%. Both figures are estimates derived from a global panel of institutions, and definitions of tangible equity vary, so treat the direction as more meaningful than the decimal.

That combination, high absolute profit and softening returns, is the most dangerous position a management team can occupy. It funds complacency. The pressures underneath are familiar to anyone running a retail or commercial bank today.

• Margin pressure as rate tailwinds fade and deposit betas normalise.
• Deposit competition from digital banks, money market funds and wallets that reprice in days rather than quarters.
• Slower balance growth in mature markets, with the strongest volume growth in segments that are hardest to serve profitably.
• Service expectations set outside the category, by companies whose entire business is the interface.
• Fintech and embedded finance moving the point of financial decision into merchant, payroll and platform journeys.
• Rising operating, resilience and compliance costs that are largely non-discretionary.

None of these is solved by a faster back office. They are revenue-quality problems, and revenue quality is a function of how well an institution understands and serves customer decisions.

The dominant operating model in retail banking growth is still the product campaign. A product owner has a target, a campaign is built, a list is pulled, a channel is booked, a response rate is measured. It is a model designed for a world in which the constraint was reach.

The constraint is no longer reach. Customers can be contacted trivially. The constraint is relevance and timing, and those are properties of the customer's decision rather than of the bank's product calendar. Reframing growth around decisions changes what the organisation builds.

A product campaign asks who is most likely to say yes. A decision model asks who is most likely to be better off, and whether the bank's intervention is what made the difference.

When boards ask where AI will actually show up in the P&L, I find it useful to reduce the answer to four pools. McKinsey's work on generative AI estimated the technology could add between 200 and 340 billion dollars in annual value to the global banking sector, equivalent to roughly 2.8% to 4.7% of industry revenues. That is a modelled forecast rather than an observed outcome, and it assumes adoption at scale, but the distribution across pools is instructive.

Stronger acquisition. Better targeting matters less than better qualification and better onboarding. The largest acquisition gains I have seen come from removing friction and from identifying which prospects will still be active in month twelve, rather than from squeezing the response rate on the top of the funnel.

Higher customer lifetime value. This is the deepest pool and the least well measured. It is built from activation, primacy, engagement depth, product fit and retention, and almost every one of those is a sequence of small decisions rather than a single sale.

Lower operating cost. Real, and largely where institutions have concentrated so far: contact deflection, document handling, code assistance, compliance drafting. The trap is treating released capacity as a cost saving without deciding what it will be redeployed towards.

Better risk decisions. Improvements in credit decisioning, fraud detection and collections often produce the fastest measurable return, because the baseline is well instrumented and the counterfactual is comparatively easy to construct.

In commercial leadership roles I have repeatedly seen capable institutions produce impressive pilots and disappointing portfolios. The failure points are consistent.

• Disconnected pilots, each proving feasibility, none changing an operating process.
• Fragmented data, where the customer view required for a decision spans four systems with three definitions of the same field.
• Technology-led ownership, which optimises for deployment rather than for a commercial outcome.
• Automating a poor journey, which delivers the wrong experience faster and at lower cost.
• Model optimisation without commercial accountability, where lift in AUC is celebrated and lift in margin is never measured.
• Weak frontline adoption, because the recommendation arrives outside the tool the banker actually uses.
• Missing incrementality measurement, so the programme cannot distinguish its own effect from the market's.
• Trust and explainability gaps that make risk and compliance functions, correctly, unwilling to approve scale.

The common thread is that the model is treated as the product. The model is not the product. The decision is the product, and a decision requires an owner, a channel, a threshold, a feedback path and a measurement design.

This is the loop I use when working through the question with banking leadership teams. It runs continuously rather than as a project, and trust sits at the centre because every stage depends on it.

Most banks do not have a technology gap here. They have an ownership gap. The decisions that matter cross business lines, and the organisation is structured by product.

What works, in my experience, is a small number of decision domains with named cross-functional owners: customer acquisition and onboarding, primacy and engagement, credit and affordability, retention and recovery. Each domain has a business owner accountable for a commercial outcome, supported by analytics, product, technology, risk, marketing and frontline representation, with a standing measurement discipline.

Two details make the difference between a working group and an operating model. First, the domain owner controls a budget and can stop things. Second, the domain has a fixed review cadence where the agenda is incremental value delivered, not activity completed. Everything else is negotiable.

Models trained on historical lending and engagement data will reproduce historical patterns, including patterns that reflect who was previously served rather than who is creditworthy. In banking this is not an abstract fairness concern, it is a regulatory and reputational exposure, and it requires active testing for disparate outcomes rather than an assurance that protected attributes were excluded from the feature set.

There is a real risk of financial exclusion at the margin. Optimisation towards profitable segments is rational for each individual decision and can, in aggregate, quietly withdraw service from customers who most need it. Institutions should measure coverage and access alongside conversion.

Over-personalisation carries its own cost. Customers who feel observed rather than served will disengage, and in financial services the trust deficit is expensive to repair. The boundary between helpful and intrusive is set by the customer's expectation, not by what the data permits.

Data quality remains the most common practical constraint. Decision quality is bounded by the reliability of the underlying signals, and many institutions are still reconciling customer identity across acquired estates.

Explainability and human accountability are non-negotiable in regulated decisions. Someone must be able to say why a customer received a recommendation, a limit or a decline, and a named person, not a model, must remain accountable for the outcome.

Finally, a fair objection to the whole argument: some of the largest banking value creation of the past decade came from disciplined cost management and balance-sheet strategy, not from customer intelligence. That is true. The argument here is not that intelligence replaces those disciplines, but that it is the remaining lever that has not been systematically pulled.