How to Build AI Agents Without a Dev Team (Using Claude Code)

Nearly every guide on how to build AI agents is written for someone who already writes Python. I went through the pages currently ranking for it, and by step three most of them have you in a terminal with an API key, choosing a framework. That is a perfectly good article for an engineer. If you run a business and you just want something to stop chasing overdue invoices for you, it is the wrong map entirely.

So here is the version I give owners across a table. It starts by trying to talk you out of building an AI agent at all, because for most of the jobs people bring me, an agent is the expensive answer to a question a simple automation already solves. If your job survives that test, the rest of this walks through building the first one with tools you drive in plain English.

I will give you real numbers in pounds, and I will be specific about the parts a UK business should not hand to an agent unsupervised.

What is an AI agent, and how is it different from an automation?

An AI agent is a system where the model decides its own steps. You hand it a goal and a set of tools, and it works out which tool to use and in what order. An automation is the opposite. The steps are fixed in advance, and it runs the same path every time.

Anthropic draws that line in its own engineering guidance: workflows are orchestrated through predefined code paths, while agents dynamically direct their own processes and keep control over how the task gets done. That distinction sounds academic until you are paying for it. An agent costs more per run, takes longer, and fails in ways you did not plan for.

Here is the test I actually use. Write the steps down on paper. If you are confident they will be identical next Tuesday, you want an automation. If step four depends on what step three found, and you cannot list the possibilities in advance, you want an agent.

Do you actually need an AI agent, or a workflow?

Probably a workflow, and the industry data backs that up rather than my opinion. Gartner predicts that over 40 percent of agentic AI projects will be cancelled by the end of 2027, blaming escalating costs, unclear business value and weak risk controls. The same research firm reckons only around 130 of the thousands of vendors claiming agentic capability are the real thing. It has a name for the rest: agent washing.

The UK picture is quieter still. The government’s own AI Adoption Research, published in February 2026 off the back of a survey of 3,500 businesses, found that 16 percent of UK businesses use AI at all. Among those that do, just 7 percent use agentic AI, making it the least adopted category in the whole study. Nearly a third of businesses reported substantial barriers to it, against 18 percent for simple text generation.

Read that as reassurance, not as a warning. Almost nobody has this working yet, which means a small firm with one good agent running is not behind. It is unusually far ahead.

The failure pattern is consistent, and it is not technical. People build the impressive agent instead of the boring one. An agent that handles your whole sales process is a project. An agent that reads yesterday’s enquiries and drafts replies into your drafts folder is a Tuesday afternoon, and it is the one still running in six months.

Which job should your first AI agent do?

Pick the job that is repetitive, needs a little judgement, is cheap to get wrong, and easy to check afterwards. All four matter. Drop the last one and you will never know whether the thing is working, which is how most first agents quietly die.

That rules out anything customer facing on build number one. It rules out anything touching money without a human in the middle. What it leaves is the unglamorous middle of your week, which is exactly where the hours are hiding.

Five that work well for a small UK business:

1. Inbox triage. Claude Cowork with the Gmail connector, or n8n with a Gmail trigger and an AI Agent node. Mail arrives, the agent classifies it as new enquiry, existing client, supplier or noise, writes a one paragraph summary, and drafts a reply into your drafts folder. Nothing gets sent. The judgement is in the classification, which is why it is a genuine agent rather than a rule.
2. Invoice chaser. An n8n agent on a Monday morning schedule, connected to Xero or QuickBooks and Gmail, logging every decision to a Google Sheet. It pulls overdue invoices, checks whether that client already replied last week, picks a tone based on how late and how large the invoice is, and hands you a batch to approve in one sitting. This is the one that pays for itself first.
3. Quote drafter. Claude Code with a Skill file holding your pricing rules, pointed at a folder of past quotes. An enquiry comes in, it finds the two closest jobs you have quoted before, drafts in your house format at your normal margins, and flags anything it is unsure about at the top.
4. Weekly reporting agent. Friday at four, pull the week’s numbers from Analytics, Search Console and Stripe, compare them against the previous month, write what actually changed, and drop it in a Google Doc.
5. Lead qualifier. Triggered by your website form, enriches the company, scores it against your written definition of a good client, and explains the score in one sentence. Hot leads to Slack, everything else parked with a tag. The score advises. It never emails the lead.

How do you build an AI agent step by step?

Six steps, and the first two happen before you touch any software.

1. Write the job down as if briefing a temp. What arrives, what you do with it, what the exceptions are, what finished looks like. Vague briefs produce vague agents, and this document becomes the agent’s instructions almost word for word.
2. Decide what it may touch and what it may only draft. Reading your inbox is safe. Sending from it is not, on day one. Write the boundary down before you build, because you will be tempted to loosen it once the thing impresses you.
3. Pick where it runs. Claude Cowork if you never want to see a terminal. Claude Code if you want it working on your real files and running on a schedule. n8n if you would rather see the workflow drawn on a canvas and connect it to a webhook. All three are reasonable, and your first choice is not permanent.
4. Give it tools, which is where MCP comes in. The Model Context Protocol is an open standard for connecting AI applications to other systems, and the useful way to think about it is a USB C port for AI. It is why your agent can read Gmail or query Xero without anyone writing a custom integration first. In Claude Code you add an MCP server; in n8n you use the MCP Client node.
5. Run it on last week’s real work. Not a sample, not a test case you invented. Feed it a fortnight of actual enquiries and read every output next to what you would have done. This is the step everyone skips and the only one that tells you the truth.
6. Put it on a schedule, and keep the log. Every run writes what it did and why to a sheet or a document. When it eventually gets something wrong, and it will, that log is the difference between a ten minute fix and starting again.

Expect the second week to be the useful one. The build gets you a rough agent by dinner time. What takes longer is discovering the exceptions your process has always had and nobody ever wrote down, which surface only when the agent trips over them.

What does it cost to build an AI agent?

Less than almost every guide implies, because none of them will name a figure. For the setup described here, you are looking at roughly 30 to 35 pounds a month.

The number that catches people out is not the subscription. It is the agent that loops. An agent decides its own steps, so a badly framed goal can have it reading the same forty emails eleven times before it settles. Cap the run, watch the first week’s usage, and the bill stays boring.

What should you never let an AI agent do on its own?

Anything that reaches a customer or makes a decision about a person without someone reading it first. That is not caution for its own sake. Under Article 22 of the UK GDPR, a decision made solely by automated means that has a legal or similarly significant effect on someone needs a route to human review, and an agent scoring job applicants or refusing credit lands squarely in that territory.

The ICO has been working on this all year. It published early thinking on agentic AI in January 2026, ran a consultation that closed in May, and has signalled that fuller guidance is coming. Its stated worry is worth understanding even before the guidance lands: when an agent acts across several systems, it gets genuinely hard to say who is the controller, and the person whose data is being processed can no longer see what happened to it.

Practically, that means two habits. Keep a log the agent writes itself, so you can reconstruct any decision months later. And keep the send button human on anything that leaves the building.

There is a commercial reason too. An agent that drafts and waits gets corrected. An agent that sends gets apologised for.

When is it worth paying someone to build it?

When the job is already scoped and still hard. A first inbox agent is genuinely yours to build, and I would rather you did, because the skill compounds. What earns a fee is the awkward stuff: two systems that were never designed to speak to each other, or an agent running unattended that would cost you a client if it failed silently at 2am.

The version I would avoid is the retainer that leaves you unable to change your own agent. If someone builds it for you, insist the accounts stay in your name and the instructions are written in plain English you can edit. That is the whole reason I work the way I do.

Start with the boring agent

If you take one thing from this, take the paper test. Most of what gets sold as an AI agent is a workflow with better marketing, and knowing the difference saves you money before you have spent any. When you do find a job that genuinely needs judgement at every run, build the unglamorous version of it first.

My honest advice is to spend an afternoon on inbox triage this week. Not because triage is exciting, but because you will learn more from one agent reading your real mail than from another month of reading about agents. Then let the second one chase your invoices, because that is where the money actually is.

If you want to know which job in your business is worth automating before you build anything, my free AI Opportunity Scorecard gives you a two minute estimate of the hours and money slipping away each month. And if you would rather build the first one alongside someone who has done it before, book an AI Opportunity Audit and we will find your best candidate, build something real in Claude Code while you watch, and leave you with a roadmap you can carry on yourself.