For more than a couple of decades, financial institutions have worked under a broadly accepted assumption: “building software is expensive, risky, and unsustainable”, and the safer long-term strategy is to “buy” (COTS – Commercial Off The Shelf) where you can share costs and leverage investment in new features resulting from the pooling effect of having multiple clients on the same codebase.

Celent has long been an advocate of this approach as well - driven by past stories of spiralling costs and complexity in custom development. We have written extensively on legacy modernization across Financial Services, including how GenAI opens up new opportunities for accelerating the reduction of technical debt.

However, the ground is shifting, and faster than most business executives (even some CIOs) appreciate.

AI, especially agentic AI, is redefining the economics of software creation. The long-held industry belief that “build” is the last resort is being challenged by software development solutions that automate not just coding, but testing, integration, and even continuous remediation.

Furthermore, as any student leaving University having studied Software Engineering will tell you today, “vibe coding” (see Celent’s report - IT, AI and "Vibe-Coding" ) is fast becoming an expected step to deliver software faster and more efficiently.

Linked to this, any executive looking to make an investment in a new business solution, now needs to weigh up whether the premium paid for some marketed packaged solutions whose origins date back to the 2000s is best decision for them and their business, when faced with teams using new AI-fuelled IDEs that can deliver tailored outcomes at lower cost with solutions designed to include self-healing capabilities – ensuring that the usual regular maintenance updates are processed (whether security patches, bug fixes, new drivers, tooling changes etc). Although still at a nascent stage, the potential to fundamentally change the economics of software build (vs buy) feels significant.

A shifting cost curve

Across early adopters, productivity gains are starting to be claimed that would have seemed unrealistic two years ago:

• 26–56% reduction in engineering productivity (according to a budget model impact assessment by Penn Wharton).

• Code generation, refactoring, and documentation tasks automated at scale. In Celent’s latest annual survey of North American insurers (GenAI-oneers in Insurance) 55% of the respondents are using gen/agentic AI for coding copilots, 42% for code generation, and 35% for testing. These numbers are indicative of other markets Celent surveyed from around the world.

• Agentic systems capable of planning and repairing entire modules with minimal human oversight. The Celent survey also indicated 22% of insurers are either in production with an agentic solution or will be by year end 2026.

When the marginal cost of change collapses, the logic behind buying software because it is “cheaper and safer” becomes increasingly tenuous.

The question then becomes: “If AI makes building fast, cheap, and governed, why continue to buy systems built for a pre-AI world?”

Meanwhile, some enlightened COTS solution providers are responding to the same signals

However, most existing COTS solutions rely on architectures, design patterns, and SDLCs that predate AI by a decade or more. Although some vendors are working on layering AI “features” onto their solutions, many are still built upon fundamentally legacy foundations. Although AI-agentic solutions will be launched rapidly, the fundamental foundation still needs to be addressed for many.

Reflecting back on a Celent blog written at the start of the Insurtech wave, we are once again in a “fluid” stage of development where the old SLDC paradigm is shifting towards something new – however we are not yet 100% clear on what that new world looks like and need to hedge.

The reality today is that financial institutions upgrading their solutions face an uncomfortable truth, i.e. they are at risk of buying yesterday’s software at a price that cannot be supported tomorrow.

This creates a tricky strategic situation to navigate. Large, multi-year vendor commitments made now could harden architectural choices and lock in costs just as the market enters a dramatic period of reinvention where the economics shift fundamentally.

The New Logic of Build vs. Buy: Less Comfortable, More Pragmatic

If we combine these two trends, being a falling build/support costs and a vendor landscape in flux itself, the emerging reality presents a new set of assumptions to manage against, i.e.:

1. Build may become strategically viable once more Not everywhere. Possibly not for core engines (at least not initially). For workflows, integration layers, data pipelines, and domain-specific logic however? Very likely. AI-empowered teams are likely to be able to deliver these components faster than many vendors can configure their own platforms (or even contract for the work through their procurement teams).

Plus, AI-fuelled IDEs (such as AWS’s Kiro, Qodo AI, Devin AI and others) make the task of build much simpler and more efficient, helping to revive confidence in build once more. Furthermore, although early, the space for enterprise-ready build environments (where security, independent infrastructure management and industry-aligned design patterns are included to enable accelerated starts) are just beginning to emerge. The Celent team are already researching and speaking with early-stage firms looking to enter this fascinating new space with vertical propositions – so watch this space.

2. Buy is narrowing to areas of genuine industry IP

Unless a vendor brings regulatory expertise, network effect advantages, or deep domain models, the cost premium may become harder to defend in an AI-enabled world. Solutions may be slimmer as a result (i.e. just the differentiated IP – and not the fat solution).

3. A “commitment gap” is emerging (that will last up until we enter a new “specific phase” for AI-native solutions)

Financial institutions are increasingly unwilling to buy legacy platforms that won’t survive the AI-native transition—yet are also unable to buy AI-native alternatives as the market is still forming.

This implies a pragmatic middle ground must be taken – such as modernising foundations where essential, building selectively, and possibly defering large commitments until the AI-native landscape stabilises or confidence in native build is high.

4. Vendor risk is increasing … in places

Perhaps arguably, the first casualties of AI-native solutions are unlikely to be financial institutions, but instead a cohort of COTS solution providers whose value propositions relied on patterns from the 2000s and were late to invest in reinvention – raising the importance of effective vendor risk management and due diligence.

A landscape in transition

In our opinion, generative AI, Agentic AI and AI-enabled engineering won’t eliminate the need for vertical solutions across financial services. However, they will change the choices around what, where and with whom to invest - and flip the decision between “build vs buy” once more.

We are entering a transition period where financial institutions need to:

• Reassess long-standing assumptions about vendor dependence

• Make clear their investments in the building blocks that enabled the “new recipe” (advocated by Celent at the end of the last decade) in order be better placed to take advantage of new AI-native solutions

• Challenge every major “buy” decision against the emerging economics of build

• Where needed, adopt a waiting posture where the vendor landscape is immature

• Prepare for a wave of AI-native entrants that will reshape the competitive landscape

In summary, the industry needs to recalibrate, and quickly. Those who cling to pre-AI decision frameworks risk locking in technical and commercial debt just as software economics undergo a profound change.

My colleague, Jamie MacGregor's (who cheekily fueled this blog!) godson is just about to complete his software engineering degree. For his classes, he’s encouraged actively to both code without and with AI help, on the assumption that the commercial world of work he enters will already expect an element of software LLM mastery. How ready are financial institutions today to welcome / attract the next generation of AI-savvy engineers?