AI agents for business are the current version of a familiar story: a demo that works beautifully in ten minutes, followed by six months of trying to make it work reliably enough to leave unattended. The gap between those two states is where almost all the budget goes, and almost none of the marketing material describes it.

An agent differs from a chatbot in one respect that matters commercially. A chatbot produces text and a human decides what to do with it. An agent takes actions, calling systems, writing records, sending messages. That difference moves the risk from embarrassment to consequence, and it is why the engineering discipline required is closer to building a payment system than to building a content tool.

Where the money actually goes: the model is the cheapest part. The cost is in tool integration, evaluation harnesses, guardrails and the human handoff path. A simple internal agent runs £5,000 to £12,000, a retrieval-backed one £12,000 to £30,000, and a multi-step agent with real system access starts around £30,000 and reaches £75,000 or more once monitoring and rollback are properly built.

Where AI Agents for Business Genuinely Work

It is worth being specific, because the honest list is shorter than the hype and more useful.

Agents do well on tasks that are repetitive, bounded, and tolerant of a review step. Triaging inbound enquiries and routing them with a draft response attached. Extracting structured data from unstructured documents where a human confirms before it is committed. Reconciling records between two systems and flagging the exceptions rather than resolving them. First-line support that resolves the common cases and escalates cleanly.

The pattern is that the agent handles volume and a human handles judgement. Every deployment we have seen work in production has that shape, and most of the ones that failed were attempts to remove the human entirely from a process where judgement was actually required.

They do badly where a mistake is expensive and hard to detect, where the task requires knowledge that exists only in someone’s head, and where the process is genuinely different every time. An agent given a task with no stable structure will produce plausible output indefinitely without ever being right, which is worse than failing visibly.

Why Pilots Stall Between Demo and Production

The first failure mode is arithmetic and it surprises people. If a single step in an agent’s chain is reliable ninety five percent of the time, a five step task succeeds about seventy seven percent of the time, and a ten step task about sixty percent. Nobody notices this in a demo because a demo runs the happy path. It becomes obvious in week three of a pilot, at which point the fix is architectural rather than a prompt change: shorten the chains, add verification between steps, and design each step so failure is detected rather than passed downstream.

The second is that most business processes have no rollback. An agent that sends the wrong email cannot unsend it. An agent that updates a customer record has changed something other systems have already read. Building the ability to reverse an action is often more work than the action itself, and it is the part that gets deferred and then discovered.

Evaluation Is the Real Engineering Cost

The third failure is evaluation, which is where the real engineering effort sits and where teams new to this consistently underestimate. Conventional software is tested against expected output. An agent produces different output each run and can be correct in several forms, so you need a graded set of realistic cases with defined criteria for acceptable behaviour, run automatically on every change. Without it you cannot tell whether a prompt adjustment improved things or moved the failure somewhere you were not looking. Expect the evaluation harness to take as long to build as the agent.

The fourth is cost variance. Token spend scales with retries and chain length, so a failing agent costs more than a working one, and an agent that loops costs a great deal more. Budgets set on the assumption that average behaviour is typical behaviour tend to be wrong in the wrong direction. Our guide to reducing LLM latency and cost covers the caching and routing strategies that keep this bounded.

The Disclosure Rules That Now Apply

This is the part that changed recently and that most UK businesses building customer-facing agents have not registered.

The EU AI Act’s transparency obligations under Article 50 took effect on 2 August 2026 as originally scheduled. They require disclosure when a person is interacting with an AI system, labelling of AI-generated synthetic audio, image, video and text, and disclosure of deepfakes. Critically, these apply based on what the system does rather than which risk tier it falls into, so an ordinary customer service agent is caught.

The high-risk obligations were deferred. The EU Digital Omnibus on AI, Regulation 2026/1744 , was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, six days before the original deadline. It moved compliance for standalone high-risk systems listed in Annex III from 2 August 2026 to 2 December 2027, and for AI embedded in products already covered by EU product safety law to 2 August 2028. The Omnibus also added two prohibited categories to Article 5, with a grace period for the associated technical safeguards running to 2 December 2026.

The AI Act Explorer is the quickest way to check which articles touch a given system. The UK has not passed an equivalent statute and continues with a regulator-led approach, so a purely domestic deployment is governed by existing law rather than a dedicated AI framework. That distinction matters less than it sounds. If you serve EU customers, or your product is used in the EU, the Act reaches you regardless of where you are incorporated, and the transparency duties are in force now rather than deferred. Building disclosure into a customer-facing agent is cheap at design time and awkward to retrofit.

What It Costs

The bands below track the same structure as our wider AI integration cost guide , and they describe build cost rather than running cost.

An internal agent doing a single bounded task against one or two systems, with a human reviewing output, runs £5,000 to £12,000. This is the right first project and it is where you find out whether your data and processes are in a fit state.

An agent with retrieval over your own documents, handling a genuine business process with several tools, runs £12,000 to £30,000. The retrieval layer is usually the larger part of that, because the quality ceiling is set by how well your documents are structured rather than by the model. Our comparison of fine-tuning, retrieval and prompting covers which approach fits which problem.

A multi-step agent with write access to production systems, proper monitoring, rollback and an evaluation harness starts around £30,000 and commonly reaches £75,000 or more. The functionality is not what costs that. The safety apparatus is.

Running costs are separate and more variable than vendors imply. Model calls, retrieval infrastructure, monitoring and the engineering time to maintain evaluations as the models change underneath you. Budget for the last of those explicitly, because model providers deprecate and revise on their own schedule and your agent’s behaviour changes when they do.

How to Scope a First Project

Pick a task where you can already measure current performance. If you do not know how long the manual process takes or how often it goes wrong, you will not be able to say whether the agent helped, and the project will be judged on impressions.

Keep the first chain short. Three steps that work is a better foundation than ten that mostly work, and it teaches you the same lessons about your own data quality.

Build the evaluation set before the agent. Twenty to fifty real cases with known good outcomes, written down before anyone starts prompting. This single practice separates projects that converge from projects that oscillate.

Design the handoff explicitly. What happens when the agent is unsure, what the human sees, and how the case gets back into the queue. Teams that treat this as an edge case discover it is the majority of the interesting traffic.

Instrument everything from the first day. You want the full input, the reasoning, the tools called and the outcome for every run, because the failures you need to understand are the ones you did not anticipate.

When Not To Build One

If the process you want to automate is stable and rule-based, conventional software is cheaper, faster, more reliable and easier to audit. A great deal of what is currently being built as agentic could be a scheduled job with some conditionals, and would be better for it.

If your data is scattered, inconsistent or undocumented, fix that first. An agent inherits every problem in the underlying data and amplifies it, because it will confidently act on bad input rather than stopping.

Before granting an agent write access to anything, it is worth reading the OWASP Top 10 for Large Language Model Applications , which catalogues the failure classes that matter once a model can act rather than only answer. And if the honest answer to what happens when it gets it wrong is that nobody would notice for a week, do not give it write access. Read-only agents that draft for human approval capture most of the value at a fraction of the risk, and they are the version that survives contact with a real workload.

Getting a Realistic Scope

The most common way these projects go wrong is committing to an ambitious agent before establishing whether the data and processes underneath it can support one, which is a question a short discovery answers cheaply.

Mecanik builds production agents through our AI integration services , with the evaluation harness, monitoring and rollback treated as deliverables rather than extras, and handles the surrounding application work through our software development team. If you have a pilot that demos well and will not stabilise, that is a specific and fixable problem worth talking through.

Related reading: Healthcare Software Development UK: Compliance and Cost , Self-Hosting Kimi K3: Hardware, Cost and Sovereignty , Third-Party API Integration: Costs and Failure Modes and OpenAI API Integration: Adding GPT to an Existing App .

Frequently Asked Questions

What is the difference between an AI agent and a chatbot? A chatbot produces text for a human to act on. An agent takes actions itself, calling systems, writing records and sending messages. That shift moves the risk from embarrassment to consequence, so the engineering discipline required is closer to building a payment system than a content tool.

How much do AI agents cost a UK business? An internal agent doing a single bounded task with human review runs £5,000 to £12,000. One with retrieval over your own documents and several tools runs £12,000 to £30,000. A multi-step agent with write access to production systems, monitoring and rollback starts around £30,000 and commonly exceeds £75,000.

Why do AI agent pilots fail to reach production? Reliability compounds badly across steps, so a chain of five steps at ninety five percent each succeeds only around seventy seven percent of the time. Most business processes also have no rollback, and building an evaluation harness typically takes as long as building the agent itself.

Do I have to disclose that customers are talking to an AI? Under the EU AI Act, Article 50 transparency obligations took effect on 2 August 2026 and require disclosure when a person interacts with an AI system, plus labelling of AI-generated content. They apply based on what the system does rather than its risk tier, and they reach businesses serving EU customers regardless of where those businesses are based.

Were the EU AI Act high-risk rules delayed? Yes. The EU Digital Omnibus on AI, Regulation 2026/1744, entered into force on 27 July 2026 and moved obligations for standalone high-risk systems under Annex III from 2 August 2026 to 2 December 2027, and for AI embedded in regulated products to 2 August 2028. The Article 50 transparency duties were not delayed.