Banking · Running a build-buy-partner decision for core AI capability
A mid-size US bank (roughly $30 billion in assets) is staring at a deadline. Its AML (anti-money laundering, the regulatory regime requiring banks to detect and report suspicious financial activity) transaction-monitoring system is nine years old. Compliance staff describe its false positive rate as "unmanageable," and a recent regulatory exam flagged model documentation gaps. The CIO has three options on the table: build a new system in-house using the bank's data science team, buy a vendor platform, or co-develop with a fintech partner. The board wants a recommendation in six weeks. This lesson equips you to run that decision.
Why this decision is harder than it looks
AML transaction monitoring uses AI/ML (machine learning) to score transactions for suspicious activity. It replaces or augments static rules ("flag any wire over $10,000") with pattern detection that adapts to new laundering techniques.
The stakes cut both ways. Under-detect and you risk regulatory penalties under the Bank Secrecy Act (BSA, the core US AML law) and enforcement from FinCEN (Financial Crimes Enforcement Network). In Europe, national regulators apply the EU's AML directives, and the new EU Anti-Money Laundering Authority (AMLA), based in Frankfurt, is moving toward direct supervision of the largest cross-border institutions as the single AML Regulation phases in from 2027. Over-detect and you drown investigators in false positives. Industry estimates commonly cite false positive rates above 90% for legacy rules-based systems. Treat that figure as directional, not exact.
This isn't a generic "AI adoption" choice. The system sits inside a heavily audited, high-liability compliance function. That changes the calculus compared with, say, adopting an AI tool for marketing copy.
The three paths, honestly assessed
Build in-house
You retain full control over the model, the data pipelinedata pipelineETL (Extract, Transform, Load) is a data integration process that pulls data from sources, reshapes it into a consistent format, and writes it into a target system.View full definition →, and the logic. This matters when regulators ask "why did the model flag this transaction?" (a requirement often called model explainability).
Costs are real and recurring: data scientists, ML engineers, MLOpsMLOpsMachine Learning Operations: combining ML and DevOps practices to industrialise, deploy, monitor, and retrain models reliably in production.View full definition → infrastructure, and the compliance and validation staff needed to document the model under supervisory guidance like the Federal Reserve's SR 11-7 on model risk management. Most mid-size banks underestimate this second cost bucket.
Build makes sense when you have a differentiated data asset (a genuinely unusual customer base or transaction pattern) or when off-the-shelf models perform poorly on your segment. It rarely makes sense as a first move for a mid-size bank with a lean data science bench.
Buy a platform
Vendors like NICE Actimize, Oracle Financial Services, SAS, and Feedzai sell AML platforms with pre-trained detection models, case management workflows, and pre-built regulatory reporting.
The advantage is speed to compliance and a shared model validationmodel validationThe process of measuring whether an AI model performs well enough, and safely enough, for the business decision it is meant to support.View full definition → burden, since these vendors have already been through multiple bank examinations. The disadvantage is fit: a vendor model trained across hundreds of institutions may not catch patterns specific to your customer base. A bank heavy in trade finance or correspondent banking has different risk patterns than a retail-heavy bank.
Buy also means vendor lock-in risk and less transparency into model internals, which can complicate explaining decisions to examiners. You are compliant, but dependent.
Partner or co-develop with a fintech
In this hybrid path, the bank supplies data and compliance expertise, and a fintech supplies ML engineering and a lighter-weight platform. It is increasingly common for capabilities like alert triage and transaction scoring, where firms such as Hawk AI or ThetaRay operate.
Partnership can combine domain fit with faster iteration than build, and more customization than buy. The catch is governance. Who owns the model documentation when the regulator asks? Who is liable if the model misses a sanctions match? These need to be contractual, not assumed.
A repeatable scoring framework
Score each option 1 to 5 on these dimensions, weighted by what matters most to your institution:
| Dimension | What to ask |
|---|---|
| Regulatory defensibility | Can you produce model documentation an examiner will accept? |
| Data fit | Does the model perform on your actual customer/transaction mix, not a vendor's benchmark? |
| Time to compliance | Can this be validated and live before your next exam cycle? |
| Total cost of ownership | Licensing/build cost plus the compliance, validation, and maintenance tail |
| Talent dependency | Do you have (or can you retain) staff to run this, whichever path you choose? |
| Explainability | Can you trace a specific alert back to the features that triggered it? |
In the AML case, the exam finding centered on documentation gaps. That means regulatory defensibility and explainability should carry the heaviest weight, often pointing toward buy or partner, where vendors bring examination-tested documentation. Build leaves the bank to create that documentation from scratch.
A worked mini-calculation
Say the bank's current system generates 100,000 alerts per year, with a 90% false positive rate (an industry-cited estimate, treat as approximate). That is about 90,000 false positives and 10,000 true positives. Each alert takes an analyst 20 minutes to review.
- Current analyst hours: 100,000 × 20 minutes = 33,333 hours/year
- At a fully loaded compliance analyst cost of roughly $60/hour (a reasonable US estimate for 2026, not vendor-specific), that's about $2 million/year in review labor alone.
Now suppose a new AI-enhanced system (build, buy, or partner) cuts false-positive alerts by a third, from 90,000 to 60,000, while catching the same 10,000 true positives. Total alerts drop to about 70,000. That saves 30,000 reviews, or roughly 10,000 analyst hours, worth about $600,000/year. The false positive rate is still around 86% (60,000 of 70,000). Even a large reduction in false-positive volume moves the headline rate less than people expect, so check which number a vendor is quoting.
Weigh that saving against the cost of each path over a 3-year horizon. Vendor licensing fees typically run in the high six to low seven figures annually for mid-size banks. Build costs concentrate upfront in talent and infrastructure. Partner costs are usually a blend. The ROI (return on investment) case only works if you model the compliance and validation tail cost, not just the license or dev cost. Many banks make this mistake when the business case looks good on a vendor's slide deck but ignores internal implementation work.
Knowledge check
1. Why does the AML transaction-monitoring context make the build-buy-partner decision fundamentally different from adopting an AI tool for something like marketing copy?
2. A bank's legacy AML system has a very high false positive rate. What is the core tradeoff this illustrates in transaction monitoring?
3. Why does model explainability matter specifically as an argument for building an AML system in-house rather than buying a vendor platform?
4. Select ALL correct answers about why this bank's AML decision is considered high-stakes and time-pressured.
Select all the correct answers.
5. Select ALL correct answers about factors that make AML transaction monitoring a good candidate for AI/ML rather than static, rules-based flagging alone.
Select all the correct answers.
What examiners actually care about
Whichever path you pick, US and EU regulators are converging on similar expectations: demonstrable model governance, not a specific vendor or build choice.
The Fed's SR 11-7 framework is still the reference standard as of 2026. It requires independent validation, ongoing monitoring, and clear documentation of model limitations. The EU AI Act imposes additional obligations on "high-risk" AI systems. AML transaction monitoring is not explicitly listed as high-risk, but it may draw scrutiny because of its impact on individuals' access to financial services. Banks should track guidance from the European Banking Authority and AMLA rather than assume a classification.
In practice, a "buy" decision does not outsource regulatory responsibility. The bank remains accountable even if the vendor built the model. This is the most common misunderstanding in build-buy-partner decisions across banking AI, and it applies well beyond AML.
🎬 [VIDEO: "How Banks Use AI to Fight Money Laundering" - youtube.com/results?search_query=how+banks+use+ai+to+fight+money+laundering - search for recent explainer content from banking technology channels covering AML transaction monitoring and AI, useful for a visual walkthrough of alert scoring and case management workflows]
Applying it to the bank's decision
The exam finding was about documentation, not detection accuracy. So the strongest initial move is likely buy or partner, not build. A vendor platform brings examination-tested documentation out of the box. A fintech partnership can be layered in later for segment-specific tuning once the base compliance gap is closed.
Build remains the long-term option. After two or three years running a vendor or partner solution, the bank may identify a genuinely differentiated detection need the market doesn't serve. At that point it has both the data and the internal experience to build credibly.
Key Takeaways
- Build-buy-partner decisions in banking AI must weight regulatory defensibility and explainability as heavily as cost or speed, especially for compliance-critical systems like AML monitoring.
- Buying a vendor platform does not transfer regulatory accountability. The bank remains liable for outcomes under frameworks like SR 11-7.
- Always model the full cost tail (validation, documentation, maintenance), not just license or build price, when calculating ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition →. Be precise about whether a vendor quotes a cut in false-positive volume or in the false positive rate.
- Partnership works best when governance and liability are defined contractually upfront.
- A weighted scoring framework (regulatory fit, data fit, time to compliance, TCO, talent, explainability) keeps the decision structured rather than driven by vendor pitches.