What Are AI Agents?
Not chatbots. Not assistants. AI agents reason, access your systems, and take actions. This guide explains what that means in practice, and when an agent is worth building vs when something simpler will do.
The Spectrum: Chatbot → Assistant → Agent
Rule-Based Chatbot
Follows a predefined script. If user says X, respond with Y.
Can: Answer FAQs, provide opening hours, share return policy.
Cannot: Handle unexpected questions, access systems, take actions.
Cost: Cheap. Often free with your support platform.
When it’s enough: Predictable queries, static answers, basic 24/7 coverage.
AI Assistant
Uses a language model (Claude, GPT) to understand natural language. Draws on a knowledge base to answer varied questions.
Can: Handle unexpected questions, paraphrase, explain complex topics, search documentation.
Cannot: Take actions. Can tell your return policy, but cannot process the return.
Cost: £2,000–£8,000 to build.
When it’s enough: Customers need help understanding products or docs, but the interaction ends with information, not action.
AI Agent
Reasons, uses tools, and takes actions within your systems.
Can: Look up orders, check tracking, process returns, update CRM, send confirmation. All within a single conversation.
Cannot: Be built cheaply or quickly. Requires proper tool integration and testing.
Cost: £8,000–£25,000 to build.
When it’s worth building: High-volume interactions requiring system access, decisions, and multi-step processes.
The Three Capabilities That Define an Agent
Tool Use
The agent interacts with your systems through APIs. It can query databases, update records, send emails, create calendar events, process payments, or trigger any API-connected action.
Example: A support agent with access to order lookup, tracking check, refund processing, CRM update, calendar booking, and email sending, choosing which tools to use based on the conversation.
Memory
Context that persists across interactions. The agent remembers previous conversations with the same customer, what actions it took, and where a process was left. It also knows your business processes: return approval before refund, manager sign-off above certain amounts.
Example: Customer calls back about a partially resolved billing issue from last week. Agent picks up where the last interaction left off.
Reasoning
The agent assesses situations and decides what to do. Not a fixed decision tree, but genuine contextual judgement powered by large language models, connected to actions.
Example: A lead qualification agent reads an enquiry from a 30-person recruitment agency asking about candidate screening automation. Despite the company being “small” by rigid scoring criteria, the agent recognises high intent from the specificity of the request and routes to a senior salesperson with a note explaining why.
What an Agent Interaction Actually Looks Like
Customer says: “I ordered a blue jacket last week but received a red one.”
Agent looks up the order by customer email
Checks what was ordered vs what was shipped. Identifies fulfilment error.
Checks stock for the correct item
Offers customer a choice: replacement or refund
Processes whichever the customer prefers (Stripe refund or new shipment)
Updates the order record
Sends confirmation email
Each step is a tool call. The sequence adapts based on what the agent discovers. A chatbot cannot do any of this.
When You Don’t Need an Agent
Fixed rules, no exceptions
If the logic is always “when X, do Y” then a standard n8n or Make workflow is cheaper and more reliable.
Low volume
An agent handling 5 interactions/day is expensive per interaction. Agents make sense at 50+/day or 1,000+/month.
Stakes too high for autonomy
Large financial transactions or safety-critical decisions may need human oversight that negates the agent’s benefit.
Information only
If customers just need access to information (not actions), a well-built AI assistant costs a fraction of an agent.
What Agents Cost
Focused Agent
£8,000–£15,000
3–5 tools, single workflow (e.g. customer support for order queries).
Multi-Workflow Agent
£15,000–£25,000
8–12 tools, multiple workflows (support + returns + billing + account management).
How We Build Agents
Define
Scope the agent’s role, tools, and boundaries. What can it do? What must it escalate?
Integrate
Build and test each tool connection to your systems. Shopify, Stripe, CRM, support platform.
Train
Configure the agent’s reasoning for your specific processes, edge cases, and brand voice.
Supervised Launch
Deploy with human oversight for 2–4 weeks. Every interaction reviewed.
Autonomous Operation
Transition to exception-only escalation. Agent handles the volume; humans handle the edge cases.
Key point: We use Claude or GPT as the reasoning engine (whichever performs best for your use case), n8n for orchestration, and your existing systems. You own everything. No proprietary platform lock-in.
Frequently Asked Questions
Find Out If an AI Agent Is Right for Your Business
Book a free discovery call. We’ll tell you honestly whether an agent is the right solution, or whether something simpler will serve you better.