According to McKinsey’s 2025 Global Survey on AI, more than 78% of companies are now using generative artificial intelligence in at least one business function. Yet for most, top- and bottom-line impacts remain minimal. Because most GenAI initiatives get restricted to prototypes and pilots with limited scope, impact rarely scales. However, with the latest developments in the AI world, a fast-maturing trend is emerging: agentic AI.

AI agents are defined as autonomous applications that observe, plan and act using available tools that operate independently, without the need for explicit human instructions.

These applications can rapidly scale impact and allow businesses to start realizing the financial benefits of investment in AI initiatives, both through helping to reduce costs and expand product offerings. 

Last year, industry leaders were looking ahead to 2025 as “the year of AI agents.” In fact, the technology is already a part of consumers’ daily lives, in the form of such models as ChatGPT, Gemini and Claude, equipped with capabilities for web search, code interpretation, deep search and image generation, and far more capable than simple large language model (LLM)-powered chatbots. Enterprise adoption, especially for supply chain and logistics operations, is the next logical step in this progression.

Current and emerging applications of AI agents for supply chain companies include:

Process automation agents that can ingest real-time context changes in trading policies for regulatory compliance; Virtual field-ops assistants that can support dispatchers and drivers by providing roadside assistance, behavior-focused coaching, and insightful operational information; Multimodal extraction agents that are able to parse text, audio and video to flag safety risks, verify inventory, and other applications. andCustomer-experience communication agents that proactively message order and estimated time of arrival changes, resolve stop exceptions, orchestrate returns, and perform dispatching across channels. 

While the potential of AI agents in supply chains is undeniable, turning that promise into operational reality requires overcoming several critical risks. The enterprise world comes with higher expectations in terms of reliability, scrutiny and compliance. In this context, deploying GenAI agents in supply chain and logistics enterprises contemplates three key challenges that need to be understood and mitigated for successful outcomes:

Consistency and reliability at scale. Due to LLM’s frequent non-deterministic outputs, ensuring consistency is a challenge. In the case of AI agents, where there might be little or no human supervision involved, this can become an even more acute problem, disrupting workflows and halting entire sequential processes. Solutions such as making multiple passes through LLMs, setting temperature = 0, validating data and creating deterministic fallbacks are some of the current best practices to help reduce this type of risk.
Evaluation. Traditional software testing methods are insufficient for maintaining a full AI agent evaluation pipeline. With multi-step workflows, different strategies need to be set in place to provide proper trace and logging of the various states along the way. A common approach used today is the LLM-as-a-judge method, whereby inputs and outputs across the workflow are passed to a separate LLM to evaluate if they are correct. Although each case must be looked at individually to define the proper methods, understanding the relevance of having a strong evaluation infrastructure is fundamental to the success of enterprise AI agents.
Context engineering, data governance and privacy. An AI agent is only as good as the context and tools it can access. One can’t expect it to accurately extract data from a source if it lacks the context of its underlying schema, taxonomy, or access permissions. Equally important is the issue of data governance. Companies need to set up systematic guard-rails, enforce least-privilege access, and provide properly managed permissions and other measures to prevent catastrophic failures such as the database  deletion by Replit’s coding agent in July of this year. Data privacy and security best practices must be enforced and continuously monitored.

As supply chain and logistics companies face growing industry volatility, the rise of agentic AI offers a valuable opportunity to shift from reactive problem-solving to proactive autonomous action at scale. Organizations experimenting now, with a clear strategy and strong safeguards, will be best positioned to translate early adoption into lasting competitive advantage.

Filipe Santos is senior product manager at Descartes.