As companies hand more operational decisions to AI, an old enterprise weakness is becoming newly dangerous: data that is technically correct, but whose meaning depends on a human knowing the context.
Somewhere in your ERP system, there is probably a field called “order delivery date”.
It sounds reassuringly precise. A date is a date, after all.
Except it may not be.
For someone in a warehouse, “delivery date” might mean the day a truck leaves the gate. For finance, it may mean the point at which the customer has received the goods and revenue can be recognised. Both values can be technically correct. Both can be perfectly formatted. Yet they describe different moments in the same transaction.
For years, organisations could live with ambiguities like this because people supplied the missing context. Planners knew which field to trust. Finance teams understood the exceptions. Experienced employees recognised that two systems were using the same words differently.
Much of that judgement was never formally documented. It did not need to be. The people doing the work knew.
Now companies are beginning to give AI systems authority not merely to analyse information, but to act on it.
That turns an old data-management problem into something considerably more consequential.
A misleading field on a dashboard can cause a bad decision. An autonomous system can interpret that same ambiguity as fact, act on it and repeat the decision thousands of times before anyone realises what has happened.
The next phase of enterprise AI therefore has a less glamorous prerequisite than better models. Companies need to confront the semantic debt buried inside their data.
The hidden middleware was human judgement
This was one of the most striking ideas in my recent conversation with Sagar Paul, Chief Revenue and Growth Officer at The Modern Data Company.
His argument is that much of enterprise information architecture was built for human consumption. Reports, dashboards and databases could tolerate inconsistencies because people acted as an informal correction layer. They recognised implausible numbers, knew which definition applied in a particular situation, and sometimes exercised restraint over information they were technically entitled to access.
An autonomous system has none of that institutional memory unless we make it explicit.
And autonomy is advancing quickly.
McKinsey reported in June 2026 that just 7% of companies had fully scaled AI across their organisations. More than two-thirds of high-performing companies identified data as the primary obstacle to enabling AI. Its conclusion is important: making data searchable is not enough. AI systems need information that is governed, traceable, reusable and understood in context.
Deloitte exposes the other side of the problem. Its 2026 State of AI in the Enterprise research found that 74% of respondents expect their companies to be using AI agents at least moderately by 2027. Yet only 21% say they have mature governance for autonomous agents.
In other words, decision-making autonomy is advancing faster than the organisational machinery underneath it.
That is not an argument against agentic AI. Supply chains are precisely the sort of environments where machines should help: enormous numbers of variables, constant change, narrow decision windows and more information than human teams can reasonably process.
But the more authority we give machines, the less room there is for undocumented human correction.
Clean data can still mean the wrong thing
Corporate data programmes have traditionally concentrated on accuracy, completeness, consistency and timeliness.
All matter.
None is sufficient.
The harder question is whether the information means what the machine thinks it means in the context of the decision being made.
Sagar described four sources of friction: inconsistent formats, variable data quality, temporal misalignment and semantic ambiguity. The last deserves far more executive attention.
Return to our “order delivery date”. A company can standardise every date into the same format, remove duplicates and ensure every field is populated.
The dataset may pass every conventional quality check.
It can still be wrong for the decision.
I think of this as semantic debt: the accumulated burden of business definitions, exceptions and contextual knowledge that organisations have left implicit because people were able to compensate.
Much like technical debt in software, semantic debt can remain almost invisible while the system continues doing roughly what is expected of it. Problems emerge when requirements change.
AI changes the requirements.
Machines force companies to turn tacit knowledge into explicit context.
McKinsey reaches a related conclusion in its research on AI data readiness. It argues that organisations should define what “good enough” means according to each use case and its risk profile, rather than pursuing some universal standard of perfect data. It also identifies semantic integrity, metadata and lineage as increasingly important as AI retrieves, transforms and recombines information.
This is a profound shift.
A number used to indicate whether demand is broadly rising or falling can tolerate uncertainty that would be unacceptable in a regulatory filing. A warehouse exception may require information seconds after it changes; a quarterly planning model may not.
An autonomous purchasing system needs something else again. It may need to know where a field originated, who owns it, how recently it changed, what permissions govern its use and what the information signifies in that particular business process.
Data quality is becoming inseparable from decision quality.
AI can automate leakage just as effectively as value
Consider procurement.
A large manufacturer may negotiate supplier agreements containing volume rebates, timing thresholds, price breaks, credits and service obligations. On paper, the savings can be compelling.
Execution is another matter.
Thousands of purchase orders move through different systems. Buyers work under pressure. Contract terms sit elsewhere. A purchase made on the 29th rather than the 2nd can alter a discount threshold. A rebate can be successfully negotiated and still never be realised.
Sagar raised exactly this type of problem in our conversation: the commercial logic can be sound while the operational environment fails to connect negotiated terms to purchasing decisions when they actually occur.
AI could be extremely valuable here.
An agent with trustworthy contract data, clear decision rules and purchasing context could identify leakage across thousands of transactions – something no procurement team could feasibly monitor manually.
But placing an agent on top of the existing process does not magically repair it.
If supplier identities are inconsistent, contract terms are poorly structured, definitions differ between systems or responsibility is unclear, automation can carry those weaknesses forward with remarkable efficiency.
Automation does not confer correctness.
It scales the logic and context it receives.
Real-time visibility can become real-time noise
A related problem sits behind our fascination with real-time data.
“Real-time visibility” has been sold for years as an almost unqualified benefit in supply chains. Often it is valuable. But information does not acquire business value simply by arriving faster.
Sagar uses telecommunications equipment as an example. A router can continuously generate data on temperature, uptime, software versions, bandwidth and operating conditions. Multiply that across a large installed base and the organisation suddenly possesses an extraordinary stream of real-time information. Much of it redundant.
If it cannot determine which signals require intervention, visibility becomes an alert factory. People eventually ignore the warnings because the organisation lacks the ability to distinguish action from noise.
Supply chains face the same problem.
Another million data points do not create another million insights. They can generate more infrastructure cost, more distraction and more opportunities for automated systems to respond to signals that do not matter.
The useful unit is not the data point.
It is the decision.
Which shipment should be expedited? Which supplier requires intervention? Which machine needs maintenance today rather than next Tuesday? Which inventory deviation threatens customer service, and which can safely be ignored?
The objective should not be real-time data everywhere.
It should be decision-ready data where timing changes the outcome.
A change of use can turn good data into inadequate data
I encountered a smaller version of this problem in my own work.
Earlier this year, researchers published a peer-reviewed study using 52 interviews from my supply-chain podcast archive as their dataset. That prompted me to look differently at hundreds of interviews accumulated over years. They were no longer simply individual pieces of content. Together, they represented a potentially valuable research corpus.
So I began turning the archive into an AI-searchable research system.
Problems appeared almost immediately.
Dates had been stored differently over the years. Metadata conventions had changed. Names and categories were inconsistent. Concepts that were perfectly intelligible within an individual episode became harder to compare across hundreds of conversations.
Nothing was fundamentally wrong with the original data.
It had done the job for which it was created.
The job changed.
That distinction matters for companies now pointing AI at information accumulated for entirely different purposes. Transaction records built for ERP workflows, maintenance notes written for technicians, supplier documentation retained for compliance and conversations captured in CRM systems may all acquire new uses once machines can retrieve, combine and act upon them.
A dataset can become inadequate without changing at all.
The context around it changed.
Context may become the competitive advantage
That leads to a strategic question.
If powerful AI models are increasingly available to every serious competitor, where does durable advantage come from?
Probably not from access to the model itself.
It may come from an organisation’s ability to provide superior context.
Deloitte’s 2026 research found that 42% of organisations regarded their AI strategy as highly prepared, while companies reported lower preparedness across infrastructure, data, risk and talent. Deloitte is explicit about the underlying problem: legacy data architectures are not sufficient for real-time autonomous AI.
Semantic integrity, provenance, lineage and governance sound like technical architecture.
Increasingly, they are business strategy.
Senior leaders need to know which decisions are being delegated to machines and what happens when those decisions are wrong. They need to ask whether important business concepts have consistent definitions across functions. They need to know whether automated systems can establish the provenance and freshness of the information they use—and whether their authority matches the potential consequence of a mistake.
Crucially, this cannot become another one-off “data cleansing” programme.
Semantic debt is not eliminated by scrubbing a database and declaring it AI-ready. Meanings change. Products change. Regulations change. Business models change. The same information can require different standards in different circumstances.
The capability companies need is continuous: preserving meaning as information moves between systems, functions and increasingly autonomous decision-makers.
Before machine speed, organisational clarity
AI will make far more routine supply-chain decisions over the next few years. In many cases, that is desirable.
Machines can monitor conditions continuously, compare far more variables than a human planner and identify patterns across contracts, inventory, logistics and operations at a scale no operations team could realistically match.
But intelligence at the top of the technology stack cannot compensate indefinitely for ambiguity underneath it.
For decades, experienced employees were the hidden middleware of the enterprise. They reconciled conflicting definitions, spotted implausible outputs and supplied context that the systems themselves did not contain.
As more of that judgement moves into software, organisations have to make the invisible explicit.
Which brings us back to that harmless-looking delivery date.
A person may know which definition applies.
A machine needs the organisation to know.
Before asking whether an AI model is capable enough to make supply-chain decisions at machine speed, leaders should therefore ask a harder question:
Is our organisation capable of explaining, precisely and consistently, what its own data means?
That is the question at the heart of my latest Resilient Supply Chain conversation with Sagar Paul of The Modern Data Company. We explore the hidden role human judgement has played in enterprise systems, why real-time information can become digital exhaust, how data productisation changes the equation, and what leaders need to establish before machines start acting on operational information.
As companies move from AI assistance to AI autonomy, that plumbing may prove considerably more consequential than the latest model release.
Leave a Reply