AI Agents for Business: What They Cost and Where They Fail

AI agents for businesses are a type of technology that has gained attention recently. These agents are capable of performing tasks autonomously, such as triaging incoming inquiries, extracting data from documents, reconciling records, and handling first-line support. Unlike chatbots, which generate text for human decision-making, agents take actions by calling systems, writing records, and sending messages. This distinction makes agents riskier, as their actions have real-world consequences.

The cost of building an AI agent is primarily in tool integration, evaluation harnesses, guardrails, and the human handoff path. A simple internal agent costs between £5,000 and £12,000, while a retrieval-backed agent ranges from £12,000 to £30,000. A multi-step agent with real system access can cost around £30,000 to £75,000, depending on the complexity of the system access and monitoring/rollback measures.

AI agents work best in scenarios that are repetitive, bounded, and can tolerate a review step. Examples include triaging inbound enquiries and routing them with a draft response attached, extracting structured data from unstructured documents with human confirmation, and reconciling records between two systems while flagging exceptions.

However, agents struggle when a mistake is expensive or hard to detect, when the task requires knowledge only in a human's head, or when the process varies each time. Deployments that work in production typically follow the pattern of the agent handling volume, while a human handles judgment. Deployments that attempt to remove the human entirely from processes requiring judgment are more likely to fail.

There are two primary failure modes when deploying AI agents. The first is arithmetic, where the reliability of individual steps in an agent's chain significantly impacts the overall success rate. For instance, if a single step is reliable 95% of the time, a five-step task succeeds about 77% of the time, and a ten-step task succeeds around 60% of the time.

This risk is often overlooked in demos, which run the happy path. The second failure mode is the lack of rollback in business processes. Agents that send emails or update customer records cannot undo their actions, leading to potential data inconsistencies and other issues.

Evaluation is the most significant engineering cost associated with AI agents. Unlike conventional software, agents produce different outputs each time they run, making it challenging to test them against expected output. To address this, teams need to create a graded set of realistic cases with defined criteria for acceptable behavior, which should be automatically run on every change. Building this evaluation harness can take as long as the agent itself.

Another critical factor in AI agent deployment is cost variance. Token spend scales with retries and chain length, meaning that a failing agent will cost more than a working one. An agent that loops will cost significantly more. Budgets set on the assumption that average behavior is typical behavior tend to be inaccurate, which can lead to overspending.

Recently, the European Union AI Act took effect on August 2, 2026, requiring any person interacting with an AI system to disclose that they are interacting with an AI system, label AI-generated synthetic audio, image, video, and text, and disclose deepfakes. These transparency obligations apply regardless of whether the system is deemed high-risk or not. The UK has not passed a dedicated AI Act, but businesses serving EU customers or used in the EU must comply with the Act's transparency duties now.