Building In-House AI Explainability vs Buying It: CFPB Circular 2022-03 Readiness Compared

Building In-House AI Explainability vs Buying It: CFPB Circular 2022-03 Readiness Compared
TL;DR
  • CFPB Circular 2022-03 requires that AI-assisted credit decisions produce specific principal reasons for adverse action. The bank can satisfy this requirement by building explainability infrastructure internally, by buying it through vendors, or by combining both approaches. The Circular evaluates outcomes, not architecture.

  • Build, buy, and hybrid approaches each have a recognizable cost profile and capability fit. Build favors banks with internal AI engineering capability and high-volume AI lending. Buy favors banks consuming vendor-supplied scoring models. Hybrid is the most common pattern at community bank scale.

  • The CFO question is not which approach produces stronger compliance. All three can satisfy the Circular. The question is which approach fits the bank's AI footprint, engineering capability, and credit decision volume at sustainable cost.

Build, Buy, or Hybrid for AI Explainability — What the Cost Comparison Actually Shows

A community bank CFO sizing AI explainability infrastructure has heard three vendor pitches with three different cost profiles. The internal-build pitch promises full control and lower per-decision cost at scale. The buy pitch promises faster deployment and lower implementation cost. The hybrid pitch promises optimization across both. Each pitch has internal coherence. None of them publishes apples-to-apples comparison data at community bank scale.

What follows is the plain-English version of how the three approaches compare on cost, capability, and Circular 2022-03 readiness.

 

What Building Explainability Infrastructure Internally Looks Like

Internal-build explainability engages the bank's engineering team to implement explainability libraries (SHAP, LIME, or similar) against the bank's AI/ML credit models. The bank owns the code, the integration, and the ongoing maintenance.

Implementation cost runs in the mid-five figures to low-six figures depending on existing engineering capacity and the number of models. Annual run-rate after implementation in the mid five figures, primarily personnel time for maintenance and accuracy testing.

Strengths: full control, lower per-decision cost at high volume, no vendor lock-in.

Weaknesses: requires AI/ML engineering capability community banks rarely staff internally, takes 3 to 6 months from initiation to first production output, accuracy testing burden falls on the bank.

 

What Buying Explainability Through Vendors Looks Like

Buy-supplied explainability engages the AI model vendor (or a third-party explainability vendor) to produce specific principal reasons output that flows into the bank's adverse action notice generation.

Annual cost in the low to mid five figures per model vendor for explainability output. Integration cost in the low five figures per vendor at deployment.

Strengths: no internal engineering required, deployment in weeks rather than months, vendor warrants the explainability output.

Weaknesses: per-decision cost grows with volume, vendor lock-in for the explainability function, vendor accuracy claims require independent testing.

 

What a Hybrid Explainability Approach Looks Like

Hybrid combines internal and external. The bank operates its own explainability for any internally-built AI models. For vendor-supplied AI, the bank consumes the vendor's explainability output but tests it for accuracy independently and maintains its own translation layer to consumer-friendly language.

Annual cost in the mid five figures to low six figures depending on the mix of internal and vendor AI in the credit decision stack.

Strengths: capability access without full internal engineering investment, vendor relationship maintained, bank ownership preserved through translation layer and testing.

Weaknesses: requires coordination across internal and vendor teams, two cost centers to manage.

 

How Each Approach Holds Up Under Circular 2022-03

The Circular requires that the bank produce specific, accurate principal reasons for adverse action. Each approach produces compliant output when implemented well.

  • Build readiness. Strong when internal engineering executes well. The bank fully understands what its explainability produces. Documentation is internally controlled.

  • Buy readiness. Strong when the vendor's explainability is rigorous and the bank documents its reliance. Weakens when the bank does not independently test vendor output accuracy.

  • Hybrid readiness. Strongest in many cases. The bank's internal testing layer catches vendor accuracy issues. The vendor relationship reduces internal engineering burden.

  • All three pass examination when documented. None is structurally weaker. The choice should follow the bank's AI footprint and engineering capability.

 

Where the Cost Break-Even Actually Lands

The cost comparison shifts with two variables: model count and decision volume.

For banks with one or two vendor-supplied AI models and modest credit decision volume, buy produces the lowest TCO. The vendor licensing cost is small; internal engineering investment is not justified.

For banks with four or more AI models, including internal builds, and high credit decision volume, build amortizes better. The internal cost is fixed; per-decision cost approaches zero.

For most community banks under $1B AUM with mixed internal and vendor AI, hybrid optimizes across both.

 

What the CFO Controls in the Decision

The CFO controls three decisions.

  1. The first is the engineering capability assessment. Honest evaluation of whether the bank has the in-house capability to build and maintain explainability infrastructure. Banks that overestimate their capability produce implementations that age poorly.

  2. The second is the vendor relationship management. Buy and hybrid both depend on vendor performance. Diligence and ongoing oversight are CFO-funded.

  3. The third is the testing investment. All three approaches require independent accuracy testing. The CFO funds the testing program.

 

Why "Build Is More Rigorous" Misses What Rigor Actually Means

Somewhere in the comparison, the CFO will hear this argument. Build is more rigorous because the bank owns the code.

That is a false choice. Rigor is a property of implementation and testing, not source. A well-executed buy produces more rigorous explainability than a poorly-executed build. The choice should follow capability and economics, not abstract preference.

 

Three Things to Work Through Before Committing to an Approach

A community bank CFO should work through three things on the comparison. The first is a current-state inventory of every AI/ML model in the credit decision path, with internal-versus-vendor status documented. The second is a sized TCO comparison of build, buy, and hybrid against the bank's actual AI footprint. The third is the testing and documentation discipline that makes whichever approach the bank chooses Circular-compliant.

When this discipline is in place, the AI explainability conversation moves from "which approach is best" to "review the program's most recent output," and the gap is the gap between vendor pitches and operating reality.

 

Choose the Explainability Path That Fits the Footprint, Not the Pitch

A community bank CFO comparing AI explainability approaches is comparing three paths to the same Circular 2022-03 outcome. The right path follows the bank's AI footprint and engineering capability. None is structurally superior.

If your bank has not sized build, buy, and hybrid options against the bank's current AI footprint in the last twelve months, that is the conversation worth having with your Tech-Operations partner before the next budget cycle.

Five Nines Technology Group is a Tech-Operations partner for community banks and credit unions. Translating regulatory frameworks into operating discipline at community bank scale is where our team focuses.

Frequently asked questions

Can the bank rely on the vendor's explainability without independent testing?

The Circular does not require independent testing, but bank ownership of the Regulation B compliance position means the bank cannot rely entirely on vendor claims. Sampled testing is good practice and reduces examination exposure.

What is the implementation timeline for each approach?

Build: 3 to 6 months from kickoff to production output. Buy: weeks to a few months depending on vendor and integration complexity. Hybrid: scales with the build component for internally-built models.

Does the bank's loan origination system vendor typically provide explainability?

Some do, some do not. LOS vendors with integrated AI scoring increasingly bundle explainability. LOS vendors operating as pure workflow may require third-party explainability integration.

Can the bank's internal engineering team maintain explainability infrastructure without dedicated headcount?

Difficult at community bank scale. Maintenance benefits from a named owner with relevant expertise. Banks that assign maintenance to general engineering rotation produce gaps that surface in testing.

Does the bank need separate explainability for each AI model in the credit stack?

Effectively yes. Each model produces decisions with its own logic. The explainability infrastructure handles each model's output appropriately, whether through one tool or multiple.

Related Blog Posts