A Model-Selection Doctrine for Commercial AI
In brief
- Choose the decision: Define the output, operating constraints, and cost of errors first.
- Design human review: Specify when to automate, abstain, correct, or escalate.
- Make decisions auditable: Connect each result to evidence, versions, reviewer actions, and operational signals.
A company asks whether it should use a foundation model, the more useful question is: what decision must the system support, and how will that decision be verified?
Predicting vessel movements, inspecting equipment, and extracting invoice fields demand different capabilities. A larger model may improve predictions, but accuracy alone does not make a system commercially usable, its outputs must also be reviewable, traceable, and monitored in operation. The following is a practical selection doctrine, informed by NIST’s AI Risk Management Framework (NIST, 2023).
1. Define the Decision Before Choosing the Model
Record the required output, action, latency, available inputs, error consequences, and acceptance criteria, identify who acts on the result and who is accountable when it is wrong.
For example, an AIS forecast may inform a port operator; it need not autonomously change a vessel’s route. Specify whether the forecast concerns future received AIS messages or estimated physical positions these are different targets.
Apply this: Write a one page decision brief before testing any architecture.
2. Protect the Evaluation and Establish a Baseline
Separate information available at prediction time from training only labels and information observed afterwards. Otherwise, evaluation may reward data leakage rather than useful capability.
Reserve test cases that reflect deployment: familiar conditions, new vessels or sites, and prospective operation, compare against the current manual process and the simplest credible model rules, statistical methods, or compact learned models. More general models must earn their additional complexity.
Apply this: Report performance on unseen operating conditions, not just a random test split, include review effort alongside predictive accuracy.
3. Adapt to the Failure, Not the Trend
Use the observed limitation to choose the intervention, missing knowledge may call for governed retrieval; unsuitable output categories for a trained head; repeated domain errors for finetuning; and excessive latency for a smaller model or optimisation. Sometimes better sensors, labels, or workflow design outperform another model.
Apply this: Maintain a failure log and test one justified intervention against the baseline at a time.
4. Engineer HITL, Traceability, and Observability Together
Human-in-the-loop (HITL) is not a last-minute fallback, define thresholds for automatic acceptance, human review, abstention, and escalation. Give reviewers the evidence needed to accept, correct, or override outputs; measure their workload and disagreement.
For every consequential decision, preserve an authorised trail:
source and time → transformations → model/prompt version → output → reviewer action.
Apply access and retention controls rather than storing sensitive data indiscriminately.
Monitor more than model accuracy, connect data freshness, drift, latency, and error rates with review backlog, overrides, and unresolved alerts. An alert without an owner or response action is not operational observability.
Apply this: Sample ten production decisions. Can you reconstruct each one, identify who reviewed it, and explain what happens if its input quality degrades?
5. Select on Verified Value and Release Controls
Compare total cost per independently verified case, including infrastructure, integration, operations, human review, and remediation. Report coverage and deferred cases too; a system that handles only easy examples should not look artificially efficient.
Release with versioned evaluations, monitoring owners, escalation routes, rollback criteria, and reevaluation triggers. Changes to sensors, customers, data sources, or review processes can invalidate a previously successful deployment.
Apply this: Pilot the entire workflow, not only model inference, before scaling.
References and Sources
- NIST (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1.
- NIST (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1.
- Grinsztajn, L., Oyallon, E., and Varoquaux, G. (2022). Why do tree-based models still outperform deep learning on typical tabular data?. NeurIPS, 35.