Budgeting for AI Explainability Infrastructure Under CFPB Circular 2022-03

What AI Explainability Infrastructure Actually Costs at Community Bank Scale

A community bank CFO sizing the bank's AI compliance budget has heard the term explainability used three different ways in three different vendor pitches. The vendors use the word to mean different things, with different infrastructure implications and different price tags. The CFO has to translate this into a sized budget line item that makes sense at community bank scale.

The Circular's requirement is operationally specific. From the CFPB:

"Creditors who use complex algorithms, including artificial intelligence or machine learning, in any aspect of their credit decisions must still provide a notice that discloses the specific principal reasons for taking an adverse action."

The bank's adverse action notice generation has to produce reasons that are specific, accurate, and traceable to what the model actually weighted. The infrastructure that produces those notices, at scale and consistently, is what the budget funds.

What follows is the plain-English version of what the infrastructure actually does, what it costs at community bank scale, and how to evaluate the options.

What Explainability Infrastructure Actually Does

Explainability infrastructure produces, for each credit decision that involves an AI or ML model, a documented set of factors that drove the decision. The factors must be specific (not generic categories), accurate (matching what the model actually weighted), and translatable to consumer-facing language for the adverse action notice.

The technical core of explainability infrastructure depends on the model type.

• For tree-based models (gradient boosted, random forest), explainability uses techniques like SHAP values that decompose each prediction into the contribution of each input feature. The output for a specific applicant tells the bank why this applicant was scored the way they were.

• For neural network models, explainability uses techniques like LIME, integrated gradients, or attention analysis. The output is similar in form: which inputs drove this specific prediction.

• For vendor-supplied scores where the bank does not own the model, explainability comes through the vendor's API. The vendor produces the factor decomposition for each scored application. The bank consumes the output and translates it into the adverse action notice.

In all cases, the explainability output then runs through a translation layer that converts model-specific feature names into consumer-friendly language. The translation layer is the second part of the infrastructure and is often the more operationally fragile piece.

What Building Explainability Infrastructure Internally Costs

A community bank with internal AI/ML models that decides to build its own explainability infrastructure faces five cost components.

Initial implementation. Building or integrating an explainability library (SHAP, LIME, or similar), wiring it to the bank's model outputs, building the translation layer, and integrating with the bank's adverse action notice generation. For a community bank with two to four models, implementation runs in the mid-five figures to low-six figures depending on existing engineering capacity.

Translation layer maintenance. The translation from model feature names to consumer language is the piece that ages. Every time the model retrains and feature names change, the translation layer needs updating. Annual maintenance runs in the mid five figures.

Testing infrastructure. The bank needs a sample-based test program that confirms the explainability output is actually accurate against the model's behavior. Testing runs in the low five figures annually if integrated with broader model risk testing.

Documentation and audit trail. Explainability infrastructure must produce evidence trails examiners can review. The bank stores the explainability output for each adverse action and can produce a sample on demand. This is more storage and process discipline than a major capital cost, but it is a real ongoing line.

Personnel. Someone owns the explainability infrastructure, ensures it produces correct output, handles examiner requests, and updates it as models change. At community bank scale this is often a fraction of a model risk role.

The all-in annual run-rate after implementation is in the mid-five figures to low-six figures for a community bank with a moderate AI footprint. Implementation adds a one-time cost.

What Vendor-Supplied Explainability Costs

The alternative path engages the AI model vendors and the bank's adverse action notice generation vendor (which may be the loan origination system vendor) to provide explainability as part of their services.

This path has different cost dynamics.

Vendor explainability fees. Some AI model vendors include explainability output in their base pricing. Others charge separately for explainability API access or higher-tier service levels. At community bank scale, the incremental fee runs in the low to mid five figures annually per vendor.

Integration cost. The bank still has to wire vendor explainability output into adverse action notice generation. This is usually a one-time integration cost in the low five figures per vendor.

Reliance documentation. The bank documents its reliance on vendor explainability and tests periodically that the vendor's output meets the bank's compliance requirements. The reliance review is a recurring cost in the low five figures annually.

No internal personnel cost. The bank does not maintain the explainability infrastructure itself. The vendor does.

The all-in annual cost of vendor-supplied explainability, across the bank's vendor relationships, runs in the low to mid five figures for a community bank with two or three AI/ML vendors in the credit stack.

What the Do-Nothing Path Costs When Enforcement Arrives

Some banks have continued to operate adverse action notices on AI-assisted decisions without explainability infrastructure. The notices use generic factor lists or scoring-model template language that does not specifically describe what the model actually weighted for this applicant.

This path has been cited by the CFPB as inadequate under Circular 2022-03. The Circular's analysis is direct: a creditor cannot rely on the model's complexity to excuse non-specific notice content. Public enforcement actions have followed.

The cost of the do-nothing path, when an enforcement action materializes, includes investigation cost, settlement cost, restitution to affected consumers, ongoing monitoring fees, and corrective action plan execution. The math is consistent with other fair lending exposures discussed elsewhere in this series. Community bank settlements in this area have run into seven figures across the cycle.

The do-nothing path is the most expensive of the three options when the bank's actual AI footprint puts it in scope for Circular 2022-03.

What the CFO Controls in the Explainability Budget

The CFO controls four things in the explainability budget conversation.

The first is the bank's AI footprint in credit decisions. Banks with no AI in lending have no Circular 2022-03 exposure on credit decisions. The footprint is the starting point.

The second is the implementation approach. Internal build, vendor-supplied, or hybrid. The CFO sizes each option against the bank's footprint and infrastructure capacity.

The third is the testing posture. Whatever explainability infrastructure the bank uses, the bank tests its output for accuracy against the model's actual behavior. The testing cadence and methodology are decisions the CFO funds.

The fourth is the audit committee briefing. The CFO presents the explainability program, the costs, and the testing posture to the audit committee on a defined cadence. The committee's awareness and the documentation of the discussion become the bank's defense if a matter materializes.

Why "Our LOS Has Always Produced Factor Lists" Doesn't Satisfy the Circular

Somewhere in the budget conversation, the CFO will hear this argument. Our adverse action notices use the factor list the loan origination system has always produced. The system is vendor-supported. We are covered.

That is a false choice for the reasons CFPB Circular 2022-03 makes explicit. The legacy LOS factor list was designed for traditional credit scoring. AI-assisted credit decisions require the notice to reflect what the AI actually weighted for this specific applicant. A generic factor list that does not match the model's actual behavior is the exact pattern the Circular cites as inadequate.

The right framing is not whether the bank's notice generation has worked historically. It is whether the bank's notice generation produces specific, accurate principal reasons when an AI or ML model contributed to the decision. The first framing produces a legacy program. The second framing produces a Circular-compliant program.

Three Things to Work Through Before the Exam Asks

Three things a community bank CFO should work through on AI explainability budgeting. The first is an inventory of every AI/ML model in the credit decision path and the explainability posture for each one. The second is a sized analysis of build-vs-buy-vs-hybrid options against the bank's actual AI footprint and credit-decision volume. The third is the testing and audit committee briefing discipline that makes the explainability program defensible during the next exam.

When this discipline is in place, the budget defense at the audit committee is recognizably more grounded. The conversation moves from "what does explainability cost" to "review the program's most recent testing outputs," and the gap between those two postures is the difference between budget discussion and routine confirmation.

Build, Buy, or Hybrid — But Produce the Notices the Circular Requires

A community bank CFO sizing the AI explainability budget is sizing the cost of producing what CFPB Circular 2022-03 requires. The Circular's standard is straightforward: specific, accurate principal reasons for every adverse action involving AI. The infrastructure that produces those notices, at scale and consistently, can be built, bought, or assembled in hybrid form. What the bank cannot do is operate the AI without producing the notices the Circular requires.

If your bank has not sized its explainability infrastructure against the actual AI footprint in its credit decision path in the last twelve months, that is the conversation worth having with your Tech-Operations partner before the next exam 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.