AI agents in the enterprise: what breaks before it works

A logistics company in Rotterdam spent eight months building an AI agent to handle supplier communication and purchase order exceptions. The agent worked beautifully in testing. In production, it collapsed within two weeks, not because the model was wrong, but because the company's ERP system returned inconsistent status codes depending on which regional database was queried. The agent had no way to distinguish a genuine exception from a data formatting artifact, and nobody had mapped that edge case before go-live.

That story is not unusual in 2026. It is, in many ways, the defining story of enterprise AI agent deployment right now.

What is actually happening with AI agentsAI agentsAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.View full definition →

The term "AI agent" has been stretched past usefulness in vendor marketing. For the purposes of this article, an agent is a system that takes an initial goal, plans a sequence of steps to achieve it, calls external tools or APIs, evaluates intermediate results, and adjusts course accordingly, with minimal human involvement in each cycle. The critical word is "adjusts." Agents are not glorified chatbots with APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → access. They are systems that make conditional decisions across multiple steps, which is precisely what makes them both powerful and fragile.

As of mid-2026, the most consequential deployments span a narrow set of use cases: customer service escalation triage, financial reconciliation, IT incident response, and sales pipelinesales pipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.View full definition → qualification. Salesforce reports broad adoption of its Agentforce platform across enterprise clients (Salesforce is a vendor with commercial interest in these figures, so treat headline numbers with proportionate skepticism), but independent observation from firms like Gartner suggests that fewer than 20% of enterprise agent projects reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → stable production status within their initial timeline.

The core architectures have converged around a few patterns. ReAct (Reasoning plus Acting), first described in a 2022 paper from Google Brain and Princeton, remains a dominant framework. Multi-agent orchestration, where specialized sub-agents handle distinct tasks under a coordinating agent, is gaining ground in complex workflows. OpenAI's Assistants API, Anthropic's Claude tool-use capabilities, and Google's Gemini function-calling interface all support these patterns with varying tradeoffs in context windowcontext windowThe context window is the maximum amount of text (measured in tokens) a language model can process at once, including both the input prompt and the generated output.View full definition → length, latency, and cost.

What is changing fast is the tooling layer. LangChain and LlamaIndex, both dominant in 2023 and 2024, have lost ground to more opinionated enterprise frameworks. Microsoft's AutoGen, now at version 0.4, and CrewAI have attracted significant enterprise adoption, partly because they impose enough structure to make debugging tractable. This matters operationally: agents that fail silently in a loosely coupled system cost far more to diagnose than those that fail loudly in a structured one.

What this means for the AI user

The first thing to internalize is that agent failures are mostly not model failures. When an agent produces a wrong outcome, the instinct is to blame the LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition →'s reasoning. In practice, the failure is usually in one of four places: the tool definitions are ambiguous, the input data is inconsistent, the task decomposition was too coarse, or there is no reliable way for the agent to know when it is out of its depth.

This reframes the skill requirement. Building effective agents is closer to systems design than to prompt engineeringprompt engineeringPrompt engineering is the practice of designing and refining text inputs to guide large language models toward accurate, relevant, and reliable outputs.View full definition →. You need to think about failure modes before you think about capability. What happens when the CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → API times out at step three of a five-step workflow? What happens when a retrieved document returns in a language the prompt wasn't designed for? These are not hypotheticals; they are the questions that determine whether your agent survives contact with real data.

The second implication is about human oversight architecture. The companies seeing the most consistent value from agents are not the ones running fully autonomous pipelines. They are the ones that have mapped out exactly where a human needs to stay in the loop and built that into the system deliberately. A useful mental model: treat every agent action that writes to a system of record, sends an external communication, or commits a financial transaction as requiring a confirmation gate, at least until you have six months of production data showing the agent's error rate in that specific action type.

Third, cost control is becoming a genuine concern. A single complex agent task can chain dozens of LLM calls. At current API pricing, a poorly scoped agent workflow in a high-volume context can generate costs that would embarrass anyone presenting the ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → case six months earlier. Anthropic's Claude 3.5 Sonnet and OpenAI's GPT-4o mini offer meaningfully different cost profiles for different steps in an agent chain. A pattern worth adopting: use a smaller, faster model for tool selection and intermediate reasoning, reserve the more capable model for final synthesis or high-stakes decisions. This alone can reduce inference costs by 40-60% on typical workflows without meaningfully degrading output quality.

Getting to production without the eight-month detour

• Start with a workflow that already has a clear definition of done. "Reduce manual handling of invoice exceptions" is too vague. "Flag and draft a response for any invoice where the line-item total differs from the PO by more than 2% and the vendor is on the approved list" is actionable. Agents perform in proportion to the specificity of the task boundary.
• Instrument everything from day one. Log every tool call, every intermediate reasoning step if the model supports it, and every output. You cannot debug what you cannot see, and the first month of production data is the most valuable dataset you will ever have for improving the system.
• Run a "chaos audit" before launch. Deliberately feed the agent malformed inputs, missing fields, and API errors. Document how it behaves. If it fails silently or produces confident wrong answers, fix the failure handling before worrying about accuracy on clean inputs.
• Pilot on low-stakes, high-volume tasks where the cost of an error is recoverable. Invoice flagging, meeting summary routing, and support ticket categorization are good candidates. Customer-facing commitments and anything touching financial closes are not.
• When evaluating vendor platforms, ask specifically about observability tooling and error recovery. A platform that cannot show you a step-by-step trace of what the agent did and why is a platform that will make your debugging life miserable.

The Rotterdam logistics team eventually got their agent to production by spending three weeks mapping every variant in their ERP status code outputs before touching the agent configuration again. That three weeks was the most valuable work of the entire project. Most of the hard problems in agent deployment are not AI problems at all.