AI in the Chain

I have spent the past few years exploring how artificial intelligence could support supply-chain work.

I remain convinced that the opportunity is significant. I have seen people automate repetitive tasks, accelerate analysis and find better ways to work with information that previously required hours of manual effort.

But I have also learned something less comfortable:

If an AI-enabled process stops when its creator goes on holiday, the organization has not adopted AI. One individual has.

This is not a criticism of the people doing the work. Quite the opposite.

Advanced users are often the ones proving what is possible. They experiment before formal programmes are ready, connect data that was previously fragmented and create practical solutions for problems others have accepted for years.

The difficulty begins when these personal successes are mistaken for organizational transformation.

The holiday test

The holiday test is simple.

When the most advanced AI user is away, does the work continue in the same way?

Can someone else run the workflow? Are the prompts, rules, data sources and assumptions documented? Is there a backup owner? Can the team identify when the output is wrong? Is the automation connected to a standard process, or does it live inside one person’s files, account or memory?

In many organizations, the answer is uncomfortable.

Part of the functionality disappears. Reports take longer. Manual work returns. Colleagues wait for the expert to come back.

The individual has become more productive, but the process has also become dependent on that individual.

This is AI-enabled key-person risk.

It resembles the spreadsheet problem companies have lived with for decades: a critical workbook is created by a talented employee, gradually becomes part of the operation and is understood by almost no one else.

AI can reproduce this problem at much greater speed and complexity.

We should celebrate personal innovation. But we should not confuse it with organizational capability.

What this looks like in practice

Imagine a logistics specialist who creates an AI-enabled tracker that monitors customs correspondence and highlights shipments at risk of delay.

The solution works well—until the specialist goes on holiday. The tracker depends on a personal mailbox, undocumented prompts and access that no backup user has. A critical customs request remains buried in an inbox, and the shipment is delayed.

The technology did not fail. The operating model did.

Now consider a demand planner who uses a personal AI assistant to consolidate regional forecast files. One region changes the structure of its spreadsheet, but the workflow continues running and silently matches the wrong columns.

The output looks complete, so the error is not immediately visible. Unless someone verifies the result against defined controls, an overstated requirement could flow into the supply plan and eventually produce unnecessary purchasing or inventory.

A similar risk can appear in order management. An experienced employee creates an AI workflow to compare customer requirements with product and order data. The solution identifies missing fields and obsolete product references—but only when that employee runs it.

During an absence, the team returns to the manual process. A retired part number is not identified until late in the order cycle, delaying confirmation and creating avoidable rework.

These are not arguments against AI.

They demonstrate the difference between a useful personal solution and a dependable organizational capability.

To become part of the operation, each workflow needs shared access, documented rules, validation controls, a backup owner and a clear process for handling changes or unexpected outputs.

The dangerous AI failure is not always an obvious error. Sometimes it is a plausible output, an undocumented dependency or a process that quietly returns to manual work when its creator is absent.

Two years of exploring can become a strategy for avoiding change

Another pattern is equally common.

The organization is exploring AI. It creates communities, attends demonstrations, runs workshops, identifies use cases and launches pilots.

Then, a year later, it is still exploring.

And sometimes, two years later, it is still exploring.

Exploration is necessary at the beginning. It gives people permission to learn, reduces fear and helps the organization discover where the technology is genuinely useful.

The problem is that exploration can become a safe, permanent state.

Nothing is fully rejected, so the organization appears open to change. Nothing is fully adopted, so the operating model does not need to change.

The demonstrations continue. The use-case lists become longer. A few enthusiasts become increasingly capable. Most employees remain uncertain about what they are expected to do differently.

The organization accumulates pilots but not capability.

This pattern is not unique to one company or one industry.

Recent enterprise studies continue to describe a gap between expanding AI access and scaled operational value. Organizations are reporting productivity improvements, but far fewer are redesigning the underlying work.

Legacy-system integration and organizational resistance remain major barriers. The move from pilot to production is still unfinished in many companies.

The current AI landscape therefore contains a strange contradiction: the technology is advancing rapidly, while organizational adoption often moves slowly.

Legacy is more than old technology

When leaders discuss legacy, they usually mean old systems.

Those systems matter. AI agents cannot reliably perform operational work if information is locked across disconnected applications, poorly defined fields and inconsistent master data.

But technology is only one form of legacy.

There are also legacy processes: approval chains designed for a different level of speed and risk.

There are legacy metrics: functions rewarded for optimizing their own performance even when the end-to-end outcome becomes worse.

There are legacy roles: work organized around collecting, reconciling and forwarding information rather than making decisions.

There is legacy knowledge: critical rules stored in the memories of experienced employees.

And there is legacy behaviour: waiting for a perfect enterprise solution instead of redesigning one important workflow.

AI adoption is difficult because it touches all these layers at once.

Installing a tool is comparatively easy. Changing how work is owned, performed, checked and improved is harder.

This is why AI cannot be treated only as a technology deployment. It is an operating-model change.

Personal productivity is not process transformation

The first wave of enterprise AI has understandably focused on individual productivity.

Employees summarize documents, draft communications, prepare presentations, analyse spreadsheets and search for information faster.

These benefits are real. They build confidence and AI fluency. They can return meaningful time to employees.

But personal productivity has a natural ceiling.

If ten people each use AI independently, the company may have ten faster individuals while the process between them remains unchanged.

Information still waits in inboxes. Decisions still cross functional boundaries slowly. Outputs still require reconciliation. The same exception may be investigated several times by different people.

The next stage is not simply giving more people access to the same tool. It is moving from isolated assistance to shared workflows.

That means defining how a task begins, which data is used, what the AI produces, how the result is evaluated, who approves the next step and how the process continues when someone is absent.

The unit of transformation must move from the individual to the workflow.

The champions’ paradox

AI champions are essential.

They create energy, make the technology less intimidating and translate broad possibilities into practical examples.

But champions can also become victims of their own success.

Because they are capable, more experimental work is directed to them. Because they understand the solutions, they become the support model. Because others depend on them, they have less time to document, teach and industrialize what they built.

The organization then has visible AI activity but limited resilience.

The role of a champion should not be to own every AI-enabled task indefinitely.

It should be to help the organization move a useful experiment through a deliberate transition:

Personal experiment → Team practice → Governed workflow → Scalable capability

Each transition requires different work.

An experiment needs curiosity. A team practice needs documentation and learning. A governed workflow needs ownership, controls and evaluation. A scalable capability needs integration, support and lifecycle management.

Many organizations are good at the first step and weak at the transitions.

What I would do next

My own conclusion is that another broad list of AI ideas will not solve the problem.

The next step should be smaller, more disciplined and more operational.

1. Find the hidden dependencies

Identify AI-enabled activities that already matter to the team.

Ask what changes when their creators are absent. This reveals which personal automations are becoming operational dependencies.

2. Select a few workflows, not dozens of use cases

Choose two or three recurring processes with measurable pain: slow exception handling, repeated data validation, missing information, manual report preparation or delayed decision-making.

Prioritize usefulness and repeatability over novelty.

3. Make every solution pass the holiday test

Document the purpose, owner, backup, approved data sources, instructions, controls and recovery process.

Another trained person should be able to run and verify it without the original creator.

4. Measure the workflow before and after

Track cycle time, error rate, rework, exception closure and human overrides.

Measuring only hours saved can hide whether the end-to-end process actually improved.

5. Decide whether to stop, standardize or scale

Every exploration should reach a decision.

Stop experiments that do not produce sufficient value. Standardize useful team practices. Invest in integration and governance only where the evidence supports scaling.

This creates movement without pretending that every experiment deserves production investment.

What experts consistently recommend

Although frameworks differ, current research converges on several practical lessons.

First, start with business workflows rather than the technology. AI creates more value when it is connected to core work and measurable outcomes.

Second, redesign the work instead of layering AI over every existing step. Automating a slow or unnecessary activity does not make the end-to-end process intelligent.

Third, combine business and technology ownership.

Operational experts understand the exceptions and trade-offs. Technical teams understand architecture, integration, security and model behaviour. Neither side can scale alone.

Fourth, build evaluation and human oversight from the beginning.

A demonstration proves that something can work. A test set, defined thresholds and ongoing monitoring help establish whether it works reliably.

Fifth, treat adoption as a capability-building effort.

Access to a tool is not the same as knowing when to use it, how to verify it and how to incorporate it into a shared process.

Finally, leaders must make choices.

Scaling requires focus, resources and changes to responsibilities. A permanent exploration phase avoids these decisions but also avoids the value.

The next divide will not be between companies with and without AI

Most organizations now have access to artificial intelligence in some form.

The more meaningful divide will be between those that accumulate isolated AI usage and those that convert learning into organizational capability.

The first group may have impressive demonstrations and highly productive individuals.

The second will have workflows that continue during holidays, role changes and periods of high operational pressure.

That resilience is not a small administrative detail. It is evidence that AI has become part of how the organization works.

I am still optimistic.

Legacy organizations contain something AI projects need: experienced people who understand customers, exceptions, risks and the consequences of decisions.

The goal should not be to bypass that knowledge. It should be to capture it, challenge it where necessary and make it easier for the whole organization to use.

But optimism should not become endless exploration.

At some point, we must stop asking only, “What can AI do?” and begin asking harder questions:

• Which workflows have actually changed?
• Which results can we measure?
• Which capabilities survive when the expert is away?
• What have we deliberately stopped doing?
• Where has the organization, rather than one employee, become better?

I would be very interested to hear from other supply-chain and operations leaders:

Are you seeing the same pattern? Has AI become an organizational capability in your company, or does much of the progress still depend on a few advanced users? And what happens when they go on holiday?

References

The State of Organizations 2026 — McKinsey & Company
https://www.mckinsey.com/~/media/mckinsey/business%20functions/people%20and%20organizational%20performance/our%20insights/the%20state%20of%20organizations/2026/the-state-of-organizations-2026.pdf

The State of AI in the Enterprise 2026 — Deloitte
https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html

The State of AI: Agents, Innovation, and Transformation — McKinsey & Company
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

The Agentic Reality Check: Preparing for a Silicon-Based Workforce — Deloitte
https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/agentic-ai-strategy.html

CEOs Are Starting to See Value from AI. Now Comes Execution — BCG
https://www.bcg.com/publications/2026/how-ceos-scale-ai-value

AI Promised a Revolution. Companies Are Still Waiting — Reuters
https://www.reuters.com/business/business-leaders-agree-ai-is-future-they-just-wish-it-worked-right-now-2025-12-16/