Standards in the Age of AI: Why Interoperability Still Matters When Machines Can Read Anything
by Iker Echave, Head of Standards at DCSA
A common question is emerging in supply chains and logistics: if artificial intelligence (AI) can interpret documents, emails, APIs, screenshots, PDFs, and unstructured messages, do we still need standards?At first glance, the question is reasonable. Generative AI has made it dramatically easier to build applications that can understand messy information. A model can read a bill of lading, extract data from an email, translate between formats, infer intent from free text, and reconcile terminology across systems. What previously required brittle integrations, bespoke mappings, or manual workarounds can now be prototyped in days or even hours.This creates the powerful impression that perhaps AI makes standards less necessary. But that impression confuses capability with economics. It also confuses human tolerance for messy processes with machine requirements for reliable execution. The question is not whether AI can understand data in different shapes and forms as the evidence show that, increasingly, it can. The question is whether using scarce AI capacity for that purpose is a good use of resources, especially when the problem being solved is fixed, repetitive, and avoidable.In a scenario where more operational work in logistics shifts from humans to AI agents, another question becomes more important: can those agents do their jobs correctly without a high-quality data input?
The MVP Illusion
By now, many people have seen an impressive demo in which someone builds a prototype that translates unstructured information, runs it through a series of LLM-enabled automations, updates key data into a system of record, and then uses that data to perform an operational process. It feels like magic. This is the MVP illusion.In the prototype phase, a team can build a proof of concept that reads carrier emails, shipment updates, messages from a port or haulier, customs documents, or free-text exception reports, and then makes an intelligent recommendation to replan a shipment. The demo works. The model understands enough. The data does not need to be perfectly structured. From there, it is easy to draw the wrong conclusion: if AI can interpret messy data, perhaps standards are no longer necessary. However, production is a different story all together.A prototype operates under forgiving conditions: low volume, limited edge cases, nearby humans, unclear cost allocation, and tolerance for occasional errors. A supply-chain system running thousands of purchase orders and shipments across carriers, forwarders, ports, terminals, shippers, banks, customs authorities, insurers, and software providers needs reliability, auditability, repeatability, security, and cost control.Every inconsistent timestamp, bespoke status code, unclear event definition, missing field, proprietary API response, unstructured PDF, or conflicting source must still be interpreted, reconciled, validated, and governed. AI may perform that work better than previous technologies, but the work has not disappeared. It has merely moved into the model. And model work is with costs.
From Human Workflows to Agentic Workflows
It has become a commonly shared belief that many logistics tasks currently performed by people will increasingly be performed by machines and AI agents. These agents will monitor shipments, read exceptions, update systems of record, trigger workflows, notify customers, reconcile documentation, check compliance, and recommend or even execute operational decisions. Human operators are remarkably good at compensating for weak processes. They know which email to trust, which customer uses an unusual reference format, which port message is incomplete, which field is usually wrong, and which system needs to be updated even when the formal process says otherwise. Much of logistics still works because humans add judgment, memory, discipline, and informal workarounds on top of imperfect data flows.AI agents do not naturally inherit that organisational context. They require reliable inputs, clear semantics, consistent timing, trusted sources, and disciplined feedback into the system of record. If agents are expected to act on behalf of the organisation, they need to know not only what a message says, but what it means, whether it is authoritative, how it maps to the operational process, and which downstream system must be updated.Many corporations are weak at building reliable processes to ingest data with discipline. Data arrives through emails, portals, spreadsheets, PDFs, APIs, EDI messages, manual updates, and local exceptions. Humans patch the gaps. But when work shifts to AI agents, those gaps become operational risk.Standard APIs and shared data definitions create discipline by design. They enable good data flows without requiring every organisation, team, or user to maintain perfect process behaviour. A standard API connection can provide an agent with structured, timely, machine-readable information at the point of use, rather than asking the model to infer meaning from messy operational traces, the organisation feeds the agent standardised events, documents, references, and status updates.This is a major advantage of DCSA standards. They do not merely help systems exchange data. They create the conditions under which AI agents can operate reliably.
Standards as a Way to Avoid Waste
Peter Drucker famously distinguished efficiency from effectiveness: efficiency is doing things right; effectiveness is doing the right things. One of his most enduring observations is that there is little value in doing efficiently what should not be done at all. That insight is directly relevant to AI in supply chains.Using AI to repeatedly clean up avoidable integration friction may be technically impressive, but it may also be strategically wasteful. If the industry can agree once on the meaning of a shipment event, a location identifier, a party role, a timestamp, a document field, or an exception code, then every AI system should not have to rediscover that meaning transaction by transaction.The highest-value use of AI is not doing unnecessary translation more efficiently. It is avoiding unnecessary translation altogether, so AI capacity can be redeployed to higher-return work.Standards are therefore not merely a legacy mechanism for system-to-system integration. In the age of AI, they become a way to reduce unnecessary inference by preventing expensive intelligence from being spent on low-value interpretation tasks and compressing shared meaning before data reaches the model. In other words, standards are how we stop spending intelligence on translation and start spending it on transformation.
From Interoperability to Intelligence Allocation
The traditional argument for standards was interoperability: systems need common formats and definitions so they can exchange data. That argument still holds. But AI adds a new argument: standards are also about intelligence allocation.Today, many users experience AI under forgiving economics. AI capabilities are often bundled into subscriptions, subsidized through enterprise pilots, supported by cloud credits, funded by venture-backed providers, or absorbed inside large technology companies pursuing adoption and market share. As a result, users may not fully feel the marginal cost of each prompt, retrieval step, model call, retry, validation, or agentic workflow.That is unlikely to remain the dominant experience forever. As AI moves from experimentation to production, organizations will increasingly need to manage tokens, GPU capacity, latency, and model usage as operating resources. Cost will be allocated to products, workflows, customers, transactions, and business outcomes. Teams will not only ask, “Can the model do this?” They will ask, “Is this the best use of our token budget?”. At that point, the economics of non-standardization become more visible.If an AI agent spends tokens interpreting a recurring data field whose meaning could have been standardized, that is a low-return use of compute. If a model repeatedly reconciles event definitions that could have been aligned at the industry level, that is recurring waste. If scarce GPU capacity is consumed translating predictable integration differences, it is not being used for higher-value work such as risk prediction, exception prioritization, network optimization, fraud detection, disruption response, or customer decision support.This is especially important in global supply chains, where the same types of information are exchanged repeatedly across many parties. Shipment events, transport documents, booking data, arrival notices, release statuses, customs references, equipment movements, and exception messages are not one-off mysteries. They are recurring coordination problems.AI can absorb complexity. Standards prevent unnecessary complexity from being generated in the first place. When tokens become a managed budget, standards become a cost-control mechanism. More importantly, they become a way to reserve scarce intelligence for the work that deserves it.
Conclusion
AI does not make standards obsolete. It changes why they matter. In the pre-AI world, standards made digital integration possible. In the AI world, standards make digital integration economically sensible and operationally reliable. They reduce ambiguity, lower token consumption, improve data quality, support auditability, and free AI capacity for higher-return activities.The question for supply chains is therefore not: “Can AI understand messy data?”. The better question is: “Should we spend scarce AI capacity repeatedly interpreting problems that standards could eliminate?”. And as AI agents take on more work, another question becomes just as important: “Can those agents act correctly if the data flowing into them is inconsistent, incomplete, or undisciplined?”As AI matures, organizations will become more disciplined about where tokens, compute, and model reasoning are spent. In that world, standards are not bureaucracy. They are leverage: a way to protect scarce intelligence from avoidable work and to give AI agents the reliable inputs they need to execute correctly. Standards are how AI becomes scalable.