The Cost of a Thorough Investigation

Opening Scene

A full forensic investigation, done properly, takes real time, real expertise, and real budget. Not every case gets the same level of scrutiny — a detective bureau allocates its most thorough investigative resources to the cases that genuinely warrant them, and uses lighter-weight methods elsewhere. The same tradeoff applies directly to explainability: some of the techniques this series has covered are computationally expensive at scale, and deploying them everywhere, indiscriminately, is rarely the right call.

In Plain English

Explainability cost refers to the real computational and engineering expense of generating explanations — particularly SHAP values for complex models, which can be significantly more expensive to compute than the original prediction itself. At scale, generating a faithful explanation for every single prediction a production system makes can meaningfully add to infrastructure cost and latency, connecting directly to this content library’s dedicated series on data platform cost and FinOps.

The Old Way

Before this had a specific machine learning framing, the same cost-versus-thoroughness tradeoff existed in any investigative or diagnostic process:

A hospital triaging which patients get the most extensive diagnostic workup, reserving expensive tests for cases that genuinely warrant them.
An auditor sampling transactions for detailed review rather than examining every single one, balancing thoroughness against real cost.
A quality control process inspecting a representative sample of units rather than every single one off the line.

In each case, thoroughness and cost were explicitly, deliberately balanced rather than one being maximized without regard for the other.

What’s Changing (and Why AI Is the Reason)

As explainability techniques have become standard practice, particularly SHAP for complex ensemble and deep learning models, their real computational cost at production scale has become a genuine, visible line item worth managing deliberately.
Efficient approximation methods for SHAP and similar techniques have matured specifically to reduce this cost, trading a small amount of precision for significantly better speed, without abandoning faithfulness entirely.
Organizations increasingly apply a tiered approach — full, expensive explanations for the most consequential decisions, and cheaper, lighter methods for lower-stakes ones — mirroring the human review prioritization from Article 12.

The Metaphor, Fully Extended

The Investigation	Explainability Cost Concept
A full forensic investigation, reserved for cases that warrant it	Full SHAP computation, reserved for the most consequential predictions
A lighter-weight investigative method for routine cases	Cheaper approximation methods for lower-stakes predictions
A detective bureau’s limited investigative budget	An organization’s real infrastructure and latency budget
Matching investigative depth to what a case actually warrants	Matching explanation depth to a decision’s actual stakes

For Beginners: What to Actually Do

Time how long SHAP computation takes on a realistically sized model and dataset, to build real intuition for its actual cost.
Learn at least one faster approximation method as an alternative to full SHAP computation for lower-stakes use cases.
Practice explicitly asking “does this decision warrant a full explanation, or would a cheaper method be genuinely sufficient?”

For Practitioners and Leaders: The Deeper Layer

Build a tiered explainability strategy explicitly, matching explanation depth and cost to the actual consequence of each type of decision.
Track explainability compute cost as its own visible line item, connecting to the FinOps practices covered elsewhere in this content library.
Avoid the trap of computing full explanations for every prediction by default; deliberate tiering usually serves both cost and genuine stakeholder need better.

Quick Recap

Explainability techniques, especially full SHAP computation, carry a real, sometimes significant computational cost at production scale.
Efficient approximation methods trade a small amount of precision for meaningfully better speed and cost.
Organizations increasingly tier explanation depth to match a decision’s actual stakes, not apply the same method everywhere.
Explainability cost deserves the same deliberate tracking as any other infrastructure expense.

Where This Fits in the Series

Article 17 covered the real cost of thorough investigation. Article 18 covers a different risk entirely: what happens when explanations themselves become a target for manipulation.