Enterprise AI in action

How DHL delivers responsible AI

The world’s largest logistics firm shows that making artificial intelligence work at scale requires clear rules, calibrated risk and constant feedback

Enterprise AI in action Image collages Pillar 3 DHL hero

Every year DHL, the planet’s largest logistics firm, moves roughly 1.8bn parcels across 220 countries and territories. Each one follows a route planned almost to the minute, from the hub where it is sorted to the van that delivers it. Coordinating this has long relied on operations research, a branch of applied mathematics shaped in wartime and later used to run global supply chains. Now DHL is adding artificial intelligence (AI) to its operations. 

The technology can improve many aspects of delivery, but unlike the explicit models and paper trails that preceded it, AI’s judgments are probabilistic, its reasoning is hard to explain and its errors cannot always be traced. This poses a question that extends far beyond logistics: how can firms deploy AI systems responsibly when the technology itself resists easy oversight?

DHL’s emerging answer is shaped by Shreya Ruikar, who leads AI initiatives across the firm. In conversation with Economist Enterprise, she shuns grand rhetoric and describes responsibility as a practical framework that helps make AI usable at scale. It is built around lightweight governance that turns anxieties into decisions, safeguards that tighten as the cost of a mistake rises and feedback loops that expose AI’s flaws before they spread. Taken together, DHL’s approach offers lessons for any organisation wrestling with the gap between the technology’s promise and the messy reality of its use.

Trust the process

Abstract principles and lofty value statements, from “AI transparency” to “human-led AI”, rarely survive contact with a firm’s day-to-day operations. What endures is process. At DHL, any new AI project needs to pass through a formal governance framework that determines whether it may go forward or requires further scrutiny. Part of why DHL formalised this process was to make sure its AI use complies with the European Union’s AI Act, which entered into force in August 2024 and will take effect in stages over the following years. It requires firms to identify and classify AI risks, document how they build and use models, and put human oversight in place for sensitive applications.

But complying with regulations was only part of the story. There was an organisational reason too. “Once people saw there is a structured process, with a legal team and technical analysis behind it, trust really jumped up,” Ms Ruikar recalls. By spelling out the expectations of AI systems in advance, DHL’s process reduced second-guessing and defensive behaviour. “The conversation went from ‘Are we doing something for governance?’ to ‘Are we doing enough?’” Ms Ruikar recalls hearing from teams and leadership. And once that trust was established, Ms Ruikar says, the leadership conversation moved on from fretting about hallucinations to weighing value, returns and which AI use cases were worth backing.

A collage illustration of graphics and computers

Research suggests this is not unique to DHL. A global study by MIT Sloan Management Review and Boston Consulting Group, published in November 2024 and based on a survey of over 3,400 respondents across 21 industries, finds that the hardest part of scaling AI is managing uncertainty and building trust rather than improving models. Firms that cope best do so by formalising decision-making early, from assigning ownership and standardising review to making risk a matter of process. In DHL’s case, that discipline cleared the ground for harder, more productive arguments—about which uses of AI are worth backing.

“Once people saw there is a structured process, with a legal team and technical analysis behind it, trust really jumped up.”

Shreya Ruikar

Head of AI strategy for people experience, DHL

Head of AI strategy for people experienceDHL

Account for mistakes by design

Despite a common review process, responsibility for AI at DHL is not one-size-fits-all. Ms Ruikar says DHL calibrates oversight to the potential impact of each AI use case. That calculation goes beyond financial exposure to include safety, social effects and other less tangible risks. Where the stakes of using a certain AI system are high, the firm advances the project in iterative steps.

This logic shapes everything from rollout speed to human oversight. One example is DHL’s AI-powered internal search tool, used by 600,000 employees to find policies and forms across countries. Because the errors it is most likely to make—such as confusing one country’s policy with another’s—are low-cost, DHL rolled out the tool quickly.

A collage illustration of graphics

By contrast, where AI’s mistakes can carry heavier consequences, DHL moves more deliberately. An AI system to sift applications for frontline roles, for example, was piloted in one country, limited to a small number of jobs and used only for straightforward checks such as whether an applicant holds a forklift licence or can work night shifts. “Just because you have an AI solution, you don’t need to delegate the entire hiring decision to it,” says Ms Ruikar. This stepwise expansion also helps firms avoid the all-or-nothing thinking that so often delays valuable AI projects or rushes dangerous ones into the real world.

The European Union's risk management guidance for AI systems, published in November 2025, aligns with DHL’s approach to AI. It recommends that organisations tailor controls to risk levels and use phased deployment, verification and continuous re-evaluation as core safety mechanisms.

“Just because you have an AI assistant, you don’t need to commit to doing 200 things with it. You can start with 10 things, learn from feedback, and expand to 20, then 50.”

Shreya Ruikar

Head of AI strategy for people experience, DHL

Head of AI strategy for people experienceDHL

Turn learning into a safety net

As is the case with humans, perfect AI models do not exist. DHL has therefore built feedback into the design. In its employee-facing search tool, users can flag weak or incorrect answers, and that feedback is tracked and used to refine the system. At the same time, the tool nudges staff on how to use it properly through short pop-up prompts. It works so long as both learn. 

DHL makes that expectation explicit. It reminds staff that AI models can err—and that flagging mistakes is part of improving them. This feedback-driven approach, says Ms Ruikar, has surfaced problems no pre-launch testing would have caught. One model, though instructed not to perform tasks outside its scope, was found converting kilometres to miles per hour—a random behaviour flagged only because a user marked the answer in miles as unhelpful. “There was no other way to know,” says Ms Ruikar.  Another example is of an internal tool instructing staff in 2025 to wear masks in the canteen—a lapse due to a mislabelled policy from 2021. The instance prompted a content-governance review, fixing the root cause rather than just the policy’s labelling.

A collage illustration of graphics and man at packstation locker

Research on safety-critical AI systems increasingly emphasises this iterative approach. A study from 2024 by researchers at Carnegie Mellon University and Princeton University argues that the effects of generative AI cannot be evaluated in isolation, because they unfold through feedback loops between AI outputs and human behaviour over time. Meaningful assurance, they write, requires “reasoning about how the feedback loop formed by AI outputs and human behaviour may drive the interaction towards different outcomes”. In this view, risk emerges through use, not just design. DHL’s approach reflects that logic: it treats AI deployment as an evolving system, where oversight, monitoring and user interaction continuously shape whether tools remain safe once they are in operation.

Engineer the boundaries you cannot control

Like most firms, DHL uses the large language models of external providers—and that means accepting their limits. “You cannot control the model itself,” Ms Ruikar says plainly, arguing that responsibility needs to shift to what she calls “boundary engineering”, the process where firms layer guardrails on AI by both demanding stricter protections from vendors and designing systems to fail safely.

A collage illustration of graphics and conveyor belt

On the technical side, this means steering models with the right type of prompts (known as “prompt engineering”) to make them say “I don't know” rather than fabricate answers. On the operational side, it means pushing model creators to tighten the space for mistakes. “We layer guardrails on top of those set by model creators, but we can never be fully sure that everything is covered,” Ms Ruikar explains.

DHL also treats transparency and user education as part of what AI specialists call “control surface”, the set of levers, technical and human, through which AI’s behaviour can be shaped in practice. For customer-facing applications, the EU AI Act mandates disclosure that a system is AI-powered, a rule DHL views as helpful, not burdensome. “When it comes to transparency, it is clear what we need to do,” says Ms Ruikar.

Responsible AI in action

DHL has turned the careful management of AI into a tool to accelerate its business. It achieves this by hard-wiring safety checks into its software, acting swiftly on feedback and strictly policing its suppliers. This precise setup will not suit every company—few operate across 220 countries with such vast exposure to mistakes. Yet the underlying principle that strict rules spur innovation applies to almost any industry. This idea anchors the third instalment of our Enterprise AI in action series. Another article from the instalment describes how E.on sets firm boundaries for its AI systems. We also explore how firms can turn responsible AI use into a daily routine, in an article drawing on conversations with experts from the UN and Cambridge University.

Sources

  1. EU Artificial Intelligence Act (artificialintelligenceact.eu). Implementation Timeline. Last updated 1 August 2024. Available at: https://artificialintelligenceact.eu/implementation-timeline/

  2. MIT Sloan Management Review & Boston Consulting Group (BCG). Learning to Manage Uncertainty, With AI. November 2024 (published online November 11, 2024). Available at: https://sloanreview.mit.edu/projects/learning-to-manage-uncertainty-with-ai/

  3. European Data Protection Supervisor (EDPS). Guidance for Risk Management of Artificial Intelligence systems. November 2025 (dated 11 Nov 2025). Available at: https://www.edps.europa.eu/system/files/2025-11/2025-11-11_ai_risks_management_guidance_en.pdf

  4. Bajcsy, Andrea; Fisac, Jaime F. Human-AI Safety: A Descendant of Generative AI and Control Systems Safety. May 2024 (posted 16 May 2024). Available at: https://arxiv.org/abs/2405.09794

More from the series