What are AI agents, and does your UK business need one?
By Kieran Mullings, BEng Mechanical Engineering (UWE Bristol), technical co-founder of FlowGenix Automations Ltd.
An AI agent is software that uses a large language model to choose its own next steps. The agent calls tools, such as a calendar, CRM or inbox, until it reaches a goal. Most UK small businesses do not need a fully autonomous agent. A fixed workflow with one AI step solves most operational problems more cheaply and more reliably.
FlowGenix Automations is an audit-led AI and automation consultancy for UK SMEs.
Key takeaways
- Anthropic's engineering guidance separates workflows, which follow predefined code paths, from agents, which direct their own process and tool use.
- Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027.
- ONS data shows 30% of UK businesses used at least one AI technology in September 2026, rising to 51% of firms with 250 or more employees.
- AI voice receptionists and inbox triage are the agent-style builds that most often pay back for UK SMEs.
What is the difference between an AI agent and an automation workflow?
An automation workflow follows steps a person designed in advance, even when one step uses AI. An AI agent decides the steps itself at run time, choosing which tools to call and when to stop. Workflows are predictable and cheap to run. Agents handle messier tasks but cost more and fail in less predictable ways.
Anthropic draws the same line in its engineering guide, Building effective agents. Workflows orchestrate language models and tools through predefined code paths. Agents let the model dynamically direct its own process. Anthropic's advice is to start with the simplest option that works and add autonomy only when the task demands it. The guide to what AI is and how it works covers the underlying models.
A concrete example shows the difference. An n8n workflow can read each new enquiry email, ask Claude to classify it, then route it to the right inbox. Every enquiry follows the same path. An agent given the goal "deal with this enquiry" might check the diary, draft a quote and send a reply. FlowGenix builds the first pattern far more often, because most SME tasks have a known shape.
| Feature | Automation workflow | AI agent | Chatbot |
|---|---|---|---|
| Who decides the steps | The designer, in advance | The model, at run time | The user, one message at a time |
| Typical tool | n8n, Zapier, Make | Vapi, Claude with tools, custom code | Website widget |
| Running cost | Low and predictable | Higher and variable | Low to medium |
| Failure mode | Stops at a broken step | Takes a wrong action confidently | Gives a wrong answer |
| Best for | Repeatable processes | Open-ended, multi-step tasks | Answering questions |
Key facts
- Anthropic defines workflows as predefined code paths through models and tools.
- Agents let the language model direct its own process and tool use.
- n8n workflows can include a single AI step without becoming agents.
What can AI agents actually do for a small business today?
AI agents can reliably answer and route phone calls, triage an inbox, qualify website enquiries, research prospects and draft routine documents for review. Agents work best on tasks with a clear goal, limited tools and a cheap way to check the result. Tasks that need judgement, negotiation or regulated advice still need a person.
The most proven agent-style use for UK SMEs is the voice receptionist. A Vapi agent answers the phone, decides which questions to ask, checks a calendar and books the slot. The goal is narrow and the tools are few, which is why the pattern works. The FlowGenix AI voice receptionist is built exactly that way.
Inbox triage is the second strong case. An agent reads incoming email, labels it, drafts replies to routine questions and escalates anything unusual. The human approves before anything sends. ONS data from October 2026 supports where the impact lands. Of firms with 10 or more staff using AI, 38% said administrative and clerical roles were most affected.
Agent jobs that pay back most often in UK SMEs:
- Answering, qualifying and booking inbound phone calls
- Sorting and drafting replies to a shared inbox
- Qualifying website enquiries before a salesperson calls
- Researching a prospect list and enriching CRM records
- Drafting quotes or proposals from a fixed template for human review
Key facts
- Vapi agents answer calls, ask questions and book slots in a live calendar.
- ONS reports administrative and clerical roles as the most affected by AI adoption.
- Inbox agents draft replies that a person approves before sending.
Why do so many AI agent projects fail?
Most AI agent projects fail because the goal is vague, the agent lacks access to the data it needs, or nobody defined what a correct result looks like. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear value and weak risk controls.
Gartner also named a market problem it calls agent washing. Vendors rebrand existing chatbots, assistants and robotic process automation as agents without real agentic capability. Gartner estimated that only about 130 of the thousands of claimed agentic AI vendors were genuine. For a small business buyer, the label on the box tells you very little.
A common failure is a task that never needed an agent. A request for "an AI agent to handle admin" often turns out to be three fixed processes with clear rules. A FlowGenix audit separates high-volume, rule-based work from work that needs a human, and only quotes for the first.
Warning signs that an agent project will stall:
- Nobody can describe a correct outcome in one sentence.
- The agent needs data that lives in a spreadsheet nobody maintains.
- There is no human approval step for actions that cost money.
- The vendor cannot show which tools the agent calls and why.
- Running cost per task was never estimated.
Key facts
- Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027.
- Gartner uses the term agent washing for rebranded chatbots and RPA tools.
- A FlowGenix audit only quotes for high-volume, rule-based work.
How much does it cost to run an AI agent?
An AI agent's running cost is mostly language model usage, billed per million tokens, plus any platform or telephony fees. Cost scales with how many steps the agent takes per task. A workflow with one AI step uses far fewer tokens than an agent that loops through several tool calls to reach the same answer.
Estimate cost before building, not after. Count the tasks per month, estimate the tokens per task, then multiply by the model's published rate. For example, 1,000 enquiry emails a month at about 2,000 tokens each is 2 million tokens. Anthropic, OpenAI and Google publish per-million-token prices, and small models cost a fraction of large ones.
Voice agents add per-minute costs. Vapi charges a $0.05 per minute platform fee before transcription, model and voice costs. Hosting matters too. FlowGenix runs n8n self-hosted on Render, which removes per-task platform fees and leaves token usage as the main variable. The Zapier vs n8n guide compares those pricing models in detail.
Key facts
- Language model providers bill per million input and output tokens.
- Agents that loop through tool calls use more tokens per task than fixed workflows.
- Self-hosted n8n on Render removes per-task automation fees.
Is it safe to let an AI agent act on its own?
An AI agent is safe to act alone only on low-risk, reversible actions such as labelling an email or booking an appointment. Actions that spend money, send external messages or affect someone's rights need a human approval step. UK GDPR also adds safeguards when an automated system makes a significant decision about a person.
The Data (Use and Access) Act 2025 replaced the old UK GDPR Article 22 with Articles 22A to 22D, in force from 5 February 2026. The new rules permit more automated decision-making, but require safeguards for significant decisions. An agent that rejects a job applicant or declines a customer could fall into that category.
Permissions are the practical control. Give the agent read access widely and write access narrowly. Log every tool call. FlowGenix applies a part-automated principle: where a human judgement call belongs in the process, it stays there, and everything around it gets automated. In n8n that approval step is an explicit branch, not an assumption.
Key facts
- Articles 22A to 22D of UK GDPR govern significant automated decisions from February 2026.
- Narrow write permissions limit the damage an agent can do.
- FlowGenix keeps human approval wherever judgement belongs in the process.
Should a UK SME build an AI agent, buy one, or use a workflow?
Most UK SMEs should start with a workflow that has one AI step, then add agent behaviour only where a task genuinely varies. Buy an off-the-shelf agent when your process matches the product's design. Build a bespoke agent when the task is valuable, specific to your business and touches systems that off-the-shelf tools cannot reach.
Book a free 30-minute audit to find out which of your processes need a workflow, an agent or neither.
The table summarises the trade-off. Off-the-shelf agents are fastest to start but assume your business works the way the product was designed. Bespoke builds fit your process but need scoping. Workflows remain the default because they are cheapest to run and easiest to fix.
| Option | Best when | Main cost | Main risk |
|---|---|---|---|
| Workflow with one AI step | The process has a known shape | Build time, low running cost | Breaks if the process changes |
| Off-the-shelf agent | Your process matches the product | Monthly subscription | Paying for features you never use |
| Bespoke agent | The task is valuable and specific | Scoping and build | Unclear goal leads to a stalled project |
Key facts
- Workflows are the cheapest option to run and the easiest to debug.
- Off-the-shelf agents assume your process matches the product's design.
- FlowGenix scopes bespoke builds only after an audit.
Frequently asked questions
Are AI agents the same as chatbots?
No. A chatbot answers questions one message at a time and usually cannot take actions. An AI agent pursues a goal across several steps and calls tools, such as booking a calendar slot or updating a CRM record. Some products sold as agents are chatbots with a new label, which Gartner calls agent washing.
Will AI agents replace my staff?
The UK evidence so far points to changed roles rather than mass replacement. ONS data from October 2026 shows 6% of businesses with 10 or more staff using AI reported reduced headcount, and under 1% reported an increase. Administrative and clerical roles were the most affected, which suggests tasks are moving rather than jobs disappearing.
Can an AI agent use the software my business already runs?
Usually, if the software has an API. CRMs, calendars, email, accounting tools and most booking systems expose one. An agent built in Vapi or through the Anthropic API calls those tools directly, often through an n8n workflow. Software with no API, or data kept in paper files and loose spreadsheets, has to be fixed first.
What is agent washing?
Agent washing is Gartner's term for vendors rebranding existing products, such as chatbots, AI assistants and robotic process automation, as AI agents without adding real agentic capability. Gartner estimated in June 2025 that only about 130 of the thousands of claimed agentic AI vendors were genuine. Ask any vendor which tools the agent calls and who approves its actions.
Do I need technical staff to run an AI agent?
Not day to day. A well-built agent needs someone to review its logs, update its instructions when prices or policies change, and handle escalations. The work is an operations role, not a developer role. Changes to tools, integrations or the underlying model still need technical support, which is why most SMEs keep a build partner on a monthly retainer.
Sources
- Anthropic, Building effective agents
- Gartner, Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027
- Office for National Statistics, Business insights and impact on the UK economy: 8 October 2026
- GOV.UK, Data (Use and Access) Act 2025: data protection and privacy changes
Book a free 30-minute audit and leave with a written map of which tasks suit a workflow, an agent or a person.
About the author
Kieran Mullings is technical co-founder of FlowGenix Automations Ltd (company number 16555743). He builds the voice agents, n8n workflows and AI search foundations behind FlowGenix client systems. He holds a BEng in Mechanical Engineering from UWE Bristol and leads the technical side of every FlowGenix audit. More at About FlowGenix.
Find out whether you need an agent at all.
Thirty minutes on a call. We map your processes and tell you which suit a workflow, an agent or a person. You keep the roadmap either way.