Build Versus Buy Is Back on the Table: What Cheaper Engineering Changes, and What It Does Not


For a decade the question of whether a bank should build its own fraud and financial crime detection capability was treated as settled. Vendors had the typology coverage, the consortium data and the regulatory track record; internal builds were a graveyard of half-finished rules engines. That consensus is now being reopened in earnest, because AI-assisted engineering has cut the cost of writing detection software to a fraction of what it was three years ago, and chief technology officers have noticed.
TrustSphere's view is that the sceptics and the enthusiasts are both half right. The economics of writing code have genuinely changed, and some of the vendor premium that institutions have been paying no longer buys anything that a competent internal team cannot produce. But the economics of operating a detection capability have barely moved, and that is where most of the money always sat. Model governance, typology research and maintenance are not engineering problems, and AI does not make them cheaper at anything like the same rate.
Our position, stated plainly: for tier 1 institutions with mature model risk infrastructure and genuinely distinctive data, the build case is stronger today than at any point since the mid 2010s, and some of them should take it. For everyone else, the new arithmetic mostly narrows the gap without closing it, and the institutions most excited by the headline build number are usually the ones least equipped to carry the costs that number leaves out.
Why the question has come back
Two forces are converging. The first is cost. Engineering teams using AI-assisted development are producing working detection pipelines, feature stores and case tooling in months rather than years, with far smaller teams. The second is a quiet dissatisfaction with the vendor market itself. The mean score across the vendors TrustSphere assesses on its RiskTech Index is 6.06 out of 10, with the strongest platforms reaching only the low sevens. That is an adequate market, not an exceptional one, and buyers can see it. When the alternative to an adequate vendor was a £6 million internal programme, adequacy won easily. When the internal programme appears to cost £2 million, the conversation changes.
There is also a control argument that resonates with boards. An internal capability can be tuned to the institution's own book, its own customer base and its own risk appetite, without waiting on a vendor roadmap. That argument is real, but it is routinely overweighted, for reasons the economics below make clear.
A worked comparison from a recent engagement
The clearest way to show where the logic holds and fails is a comparison TrustSphere worked through with a mid-tier UK bank during 2026, anonymised and rounded. The bank wanted a behavioural fraud detection capability for faster payments, replacing an ageing rules engine.
In 2023 the bank had costed an internal build at roughly £5.2 million over eighteen months, with a team of around twenty engineers and data scientists. Re-costed this year with AI-assisted development, the build estimate fell to roughly £2.1 million over ten months with a team of nine. That is a genuine, dramatic fall, and it is why the question is back on the table.
The vendor alternative was a specialist platform at about £950,000 a year in licence fees plus £600,000 of implementation. On build cost against licence cost alone, building now looked obviously cheaper: £2.1 million once, against £5.35 million over five years.
Then we added what the build number left out. Running the capability required a permanent team for model retraining, rule tuning, typology research and data engineering, which the bank's own second line costed at around £1.1 million a year once validation support was included. Over five years the build path totalled roughly £7.6 million against £5.35 million for the vendor route. The gap between build and buy had narrowed from roughly three times to around 1.4 times, which is exactly why the decision now deserves serious analysis rather than reflex. But the build path was still the more expensive one, and every pound of the difference sat in categories the original business case had not priced.
Where the build case genuinely holds
The logic holds where three conditions are met at once. The institution already carries the fixed costs of a mature model risk function, so the governance burden is marginal rather than new. Its data is genuinely distinctive, meaning a vendor's consortium view adds relatively little to what the institution can see in its own book. And its scale lets it amortise a permanent detection engineering team across enough volume that the per-transaction cost beats a licence.
That describes a handful of tier 1 banks and the largest payment providers, and almost nobody else. It also describes a class of narrower builds that TrustSphere actively encourages at all tiers: alert triage tooling, enrichment layers, investigator copilots and reporting pipelines built around a vendor core. Cheap engineering makes the edges of the stack an excellent place to build. It does not make the core one.
Where the logic fails
The first failure point is model governance. In the UK, the PRA's supervisory statement SS1/23 on model risk management sets out principles covering model identification, governance, development, independent validation and risk mitigants, and it explicitly extends to artificial intelligence and machine learning techniques. Those expectations apply to a detection model whether a vendor wrote it or an internal team did. The difference is that a vendor amortises its validation evidence, documentation and monitoring tooling across its whole client base, while an internal build carries the full burden alone. In our engagement data, governance and validation costs are the single most underestimated line in build business cases, typically by a factor of two or more.
The second is typology coverage. A bank sees only the fraud that hits its own book. A vendor with several hundred institutions on a shared platform sees new scam patterns, mule networks and document fraud techniques across the market, and ships that intelligence to everyone. AI-assisted engineering does nothing to close this gap, because it is a data position, not a software one. Internal teams can build a detection engine in months; they cannot build five years of cross-industry fraud outcomes.
The third is maintenance debt, and it is the quiet killer. AI assistance accelerates writing code far more than it accelerates understanding, auditing and safely changing code under regulatory scrutiny. Fraud typologies move quarterly; the internal team that built the engine gets reassigned to the next strategic programme; and three years later the institution owns a bespoke system nobody fully understands, which is precisely the position the buy consensus was formed to escape. We have already seen early versions of this in 2026 engagements: AI-generated codebases that were cheap to create and are proving expensive to own.
What this means for each stakeholder
For tier 1 banks, the honest conclusion is selective insourcing. The strongest candidates are institutions that already run SS1/23-grade model governance at scale and can treat detection as a permanent engineering discipline rather than a project. Even there, the sensible pattern is to build differentiating layers while retaining vendor or consortium data feeds for typology coverage.
For challengers and fintechs, the headline build number is a trap. The engineering saving is real, but these are precisely the institutions without spare validation capacity, without deep typology research and with teams that turn over quickly. Their advantage lies in being demanding buyers: the narrowed cost gap is leverage in licence negotiations, not a reason to build.
For payment service providers and acquirers, transaction volumes make the amortisation maths more attractive, but the typology problem is more acute, because payments fraud intelligence is overwhelmingly a network effect. A PSP building alone is choosing to fight a networked adversary with a single-node view.
For technology vendors, the message is blunt. The part of the price that paid for software assembly is collapsing, and buyers know it. Vendors whose value is genuinely in consortium data, typology research, validation support and managed model updates should price and sell on exactly that, because it is the part an internal team cannot replicate. Vendors selling configurable rules engines at platform prices will lose this argument within a few years.
For regulators and supervisors, a wave of internal builds means a wave of new in-scope models with thinner documentation and less tested governance. Supervisory attention on model inventories and validation independence will land hardest on institutions that built quickly and governed slowly.
The decision that actually matters
Build versus buy was never really a technology question, and cheaper technology has not made it one. It is a question about which fixed costs an institution is willing to carry forever: governance, research and maintenance, or licence fees and vendor dependence. AI-assisted engineering has moved the boundary, and for a small number of institutions it has moved it decisively. For most, it has simply made the wrong answer cheaper to start and no cheaper to live with. The institutions that get this right in 2026 will be the ones that cost the ownership, not the build.
Suggested Next Steps
Map every proposed internal detection model against SS1/23 expectations before approving the business case, and have the head of model risk countersign the governance cost estimate.
Use the narrowed cost gap as negotiating leverage at vendor renewal: ask vendors to justify their price in terms of consortium data, typology updates and validation support rather than software features.
Pilot internal AI-assisted builds on the edges of the stack first, such as triage and enrichment tooling, where governance exposure is low and the engineering saving is immediate.
Sources: Bank of England Prudential Regulation Authority, SS1/23 Model risk management principles for banks; TrustSphere proprietary engagement data; TrustSphere Risk Index, April 2026.
Companion vendor assessment: today's TrustSphere Risk Index post assesses Experian against this problem. Read it at www.trustsphere.ai
TrustSphere helps financial institutions design and deploy intelligent fraud and financial crime detection solutions. Visit www.trustsphere.ai



Comments