Explainable to Auditable: Tamper-Proof AI for US Supply Chains

At Cincy AI Week, I had the opportunity to present “From Explainable to Auditable: Tamper-Proof AI for US Supply Chains,” a talk focused on one of the most urgent questions facing critical industries: how do we trust AI when it begins influencing the systems that move medicine, materials, energy, and essential goods across the economy?

This topic sits at the intersection of much of my career working across life sciences manufacturing, global supply chains, digital transformation, enterprise architecture, AI, and industrial technology ecosystems. It also builds directly on the ideas I explored in my book, REWIRE! Using the Digital Ecosystem Playbook to Reinvent Your Business, where I wrote about how businesses must move beyond isolated systems and think in terms of connected digital ecosystems.

For the past few years, the AI conversation has been dominated by one word: explainability.

Organizations have focused on whether a model can explain why it made a recommendation, whether a user can understand the output, and whether the business can see which factors influenced the answer. These are important questions, and they have helped build early trust in AI systems.

However, in critical supply chains, those questions are no longer sufficient. As AI moves from experimentation into operational decision-making, a more important question emerges: can the system prove what happened?

This is the shift we need to make, from explainable AI to auditable AI. Explainable AI focuses on understanding why a model produced a particular output. Auditable AI goes further. It asks whether we can prove what the system saw, what policy governed the action, who authorized it, whether the evidence was authentic, and whether anything changed after the fact.

This distinction matters because supply chains are no longer simple, linear chains. They have evolved into complex digital ecosystems composed of suppliers, manufacturers, logistics providers, hospitals, regulators, cloud platforms, APIs, documents, sensor data, forecasts, purchase orders, quality records, and partner systems.

AI is now entering that ecosystem.

It is being used to forecast demand, interpret supplier documents, flag shortages, recommend substitutions, optimize inventory, prioritize shipments, and support deviation investigations. Increasingly, AI is moving closer to the workflows that determine what gets bought, made, shipped, released, delayed, or denied.

This creates enormous opportunity, but it also introduces a new class of risk.

The next major supply-chain failure may not begin with a port closure, a factory fire, or a missing raw material. It may begin with an AI system making a decision based on manipulated data, outdated policy, compromised documents, unverified supplier records, or incomplete context.

The concern is not simply that AI could make a wrong decision. The deeper concern is that, after the decision is made, no one may be able to prove why it happened.

Life Sciences Shows Why this Matters

In life sciences, supply-chain failure carries consequences far beyond delayed inventory. It can mean that a patient does not receive chemotherapy, that an emergency department lacks a critical medication, or that a temperature-sensitive therapy loses its integrity in transit. It can also mean that a manufacturer cannot prove whether a batch, supplier, shipment, or document should have been trusted.

This makes life sciences an important lens for understanding the future of AI-enabled supply chains.

In this industry, trust is not theoretical. It is not enough to assume that systems worked correctly. Organizations must prove the batch, the source, the process, the deviation, the release, and the chain of custody.

Now AI is entering this same operating environment.

The question is not whether AI can make the supply chain faster, it clearly can. The question is whether it can do so without making the system unverifiable.

Speed without verifiability does not create resilience. It accelerates risk.

The Trust Crisis is Already Here

The pressure on supply chains is already visible. In the first quarter of 2024, the United States reached 323 active drug shortages, the highest level recorded since national tracking began. This is not an abstract issue; it directly affects patient access.

At the same time, cyber and geopolitical risks are increasingly intersecting with supply-chain operations. Threat reporting shows that state-linked actors are targeting technology and AI ecosystems, while logistics networks have seen increased targeting from China-nexus actors. Modern supply chains depend on digital foundations, software, cloud platforms, identity systems, APIs, data flows, partner networks, and AI infrastructure, that adversaries are actively attempting to compromise.

AI changes the nature of these threats.

Traditional cyberattacks may steal data or disrupt systems. In contrast, AI-era attacks can manipulate the system’s perception of reality. Adversaries can alter supplier records, modify logistics events, poison documents, compromise model inputs, inject malicious instructions, manipulate policy signals, or corrupt software dependencies. If AI trusts compromised evidence, it can make incorrect decisions while appearing to optimize outcomes.

This is a form of decision warfare. In supply chains, decisions determine what gets sourced, routed, released, substituted, expedited, quarantined, or withheld. When those decisions are compromised, the consequences can be significant.

Explainability is not auditability

A model explanation does not provide a chain of custody. A dashboard does not constitute proof. A log file does not serve as a complete decision record. Even human approval is insufficient unless it can be demonstrated what the human saw, what they approved, what policy applied, and whether the underlying evidence changed afterward.

This is the gap many organizations have yet to address.

While companies are building AI experiences, pilots, copilots, and agents, many are not investing in the evidence infrastructure required to defend those systems when something goes wrong.

In regulated and critical supply chains, the standard must be higher.

Systems must be able to answer fundamental questions about what the AI saw, where the data originated, whether documents were authentic, which policies applied, whether suppliers were approved, whether shipments met required conditions, which models or agents were involved, what tools were used, who authorized actions, what changed as a result, and whether decisions can be independently verified later.

If a system cannot answer these questions, it should not influence consequential supply-chain decisions.

The answer is the decision receipt

A decision receipt is a tamper-evident record of an AI-assisted recommendation, decision, or action.

In everyday life, a receipt does not guarantee that a product was perfect or that a service was satisfactory. It simply proves that a transaction occurred, documenting when and where it happened, what was involved, and who participated. A decision receipt serves a similar purpose for AI systems.

It does not claim that the AI was correct. Instead, it provides a verifiable record of the evidence and context behind the decision.

A comprehensive decision receipt captures the business action or recommendation, the data and documents used, source systems and timestamps, cryptographic hashes to verify integrity, the models or agents involved, applicable policies, approval paths, actions taken, outcomes produced, and the audit trail necessary for future verification.

This is how AI becomes defensible, not just intelligent or explainable, but capable of standing up to scrutiny.

Simplified Architecture of Auditable AI

Auditable AI requires more than advanced models; it requires a robust operating architecture.

The first layer is the evidence layer, where all critical inputs, such as supplier documents, purchase orders, batch records, logistics events, temperature readings, certificates of analysis, inventory signals, forecasts, quality records, and policy documents, are captured as verifiable evidence. The key question is not simply whether data exists, but whether it can be proven authentic, current, authorized, and unchanged.

The second layer is the provenance layer, which establishes lineage. It tracks where data originated, who interacted with it, how it was transformed, which version was used, and whether it changed after a decision was made. Provenance transforms fragmented data into a defensible chain of custody.

The third layer is the policy layer. AI systems must operate against explicit, machine-enforceable policies rather than vague business intent. These policies include supplier approval rules, cold-chain thresholds, regulatory constraints, quality requirements, financial limits, and escalation protocols. The more consequential the action, the stronger the policy enforcement must be.

The fourth layer is the agent and action layer. Each AI agent must have a defined identity, permission boundaries, and a clear scope of authority. Systems must distinguish between AI that summarizes, recommends, drafts, escalates, or executes actions, as each carries a different level of risk.

The fifth layer is the receipt layer, where every significant AI-assisted action generates a decision receipt. This creates a compact, verifiable evidence package that can be reviewed by executives, auditors, regulators, customers, insurers, or investigators. It enables organizations to move from reactive forensics to continuous assurance.

Deterministic AI controls the action. Generative AI accelerates the thinking.

This distinction is particularly important in life sciences.

Deterministic AI serves as the control layer, supporting repeatable, rules-based, testable, and validated workflows such as supplier checks, release gates, temperature monitoring, quality rules, approval routing, and compliance enforcement.

Generative AI functions as the intelligence layer. It excels at summarization, investigation support, document analysis, regulatory interpretation, scenario planning, and decision support.

The risk arises when generative AI is allowed to operate as a control layer without sufficient auditability.

Generative AI should enhance human decision-making by accelerating understanding and insight. Deterministic systems should govern whether actions are permitted. Decision receipts should document what occurred when actions are taken.

Compliance is becoming evidence infrastructure

Compliance is also evolving. Traditional approaches relied on periodic audits, manual evidence collection, static procedures, and retrospective investigations. The emerging model emphasizes continuous assurance.

In this model, evidence is generated at the time of action, policies are enforced as code, decisions are linked to source records, approvals are tied to what approvers actually saw, audit trails are created automatically, and cryptographic integrity is embedded into workflows.

This is not an expansion of compliance bureaucracy. It is the development of the evidence infrastructure required for AI-enabled operations.

Regulators, boards, and customers are increasingly asking whether organizations can prove that their AI systems used the correct data, followed appropriate policies, operated within defined authority, and produced trustworthy records.

This is the central governance challenge.

Supply chains are now geopolitical attack surfaces

Supply chains have evolved beyond business networks into geopolitical attack surfaces.

AI amplifies this risk because adversaries no longer need to directly disrupt operations. Instead, they can target the AI’s perception of the supply chain.

By manipulating supplier records, logistics events, quality documents, software dependencies, model inputs, or policy signals, adversaries can influence decisions without directly interfering with physical operations.

This shifts the focus of supply-chain cybersecurity from protecting systems to protecting decision integrity.

In an AI-enabled supply chain, the greatest risk is not an AI system acting unpredictably. It is an AI system making decisions based on corrupted evidence while appearing to function correctly.

The leadership mandate

Organizations exploring AI in supply chains must ask new questions.

Rather than focusing solely on where AI can be applied, leaders must consider where AI will influence decisions that may need to be defended in the future. They must identify which decisions require evidence, which data sources must be trusted, which policies must be machine-enforceable, which actions require human oversight, which agents need defined authority boundaries, and which decisions require formal receipts.

Scaling AI without designing for auditability introduces significant risk.

A critical principle for the next era of AI-enabled supply chains is to design the receipt before scaling the agent.

The Future Belongs to Provable AI

The organizations that succeed will not be those with the most impressive AI demonstrations. They will be those that can operationalize AI safely, defensibly, and at scale.

They will be able to prove what the AI saw, which policies applied, who authorized actions, whether evidence changed, and what ultimately occurred.

This represents a progression from “trust me” to “explain me” to “prove me.”

Explainability was the first step. Auditability is what enables AI to operate at the core of critical supply chains.

In the future, the most important AI systems will not simply provide answers. They will be capable of standing up to scrutiny and proving their actions.