AI Agent Development
An AI agent is software that uses a language model to decide on and take actions towards a goal — reading email, updating systems, drafting replies — under rules and approvals you control. We build agents for UK businesses that want more than a chatbot: software that does the work, not just talks about it, with a person kept firmly in the loop where it counts.
This page is the definitional home for our agent work. If you are weighing whether agents are ready for a business like yours, read our candid view in Are AI agents overhyped? — a sales page that also tells you where the technology falls short is one you can trust.
What is an AI agent?
An AI agent is software that perceives information, decides what to do, and takes action towards a goal — without a human triggering every step. Unlike a chatbot following a script, it reasons through variables, executes multi-step tasks, and escalates to a person when it hits a boundary it should not cross.
In plain terms: a chatbot answers; an agent acts. An agent can read a support ticket, look up the customer, draft a resolution and queue it for approval — a chain of steps, not a single reply.
What is agentic AI?
Agentic AI is the broader category: AI that pursues goals autonomously — planning, deciding and acting across tools and data — rather than only responding to prompts. An AI agent is a single instance of agentic AI; “agentic” describes the behaviour, “agent” describes the thing.
The distinction from ordinary generative AI is autonomy. Generative AI produces an answer when you ask; agentic AI decides what to do next and does it, iterating until the goal is met or a guardrail stops it. For a fuller treatment, see What is agentic AI and what it means for your business.
What can AI agents do for a business today (and what they can’t yet)
Today, agents reliably handle bounded, well-documented tasks with clean inputs: triaging and drafting email, qualifying leads against your criteria, chasing invoices, moving data between systems, and answering questions from your own documents. These are the wins worth having, and they are available now.
What agents cannot yet be trusted with is unbounded judgement, regulated decisions, or irreversible actions taken without a check. The failure mode is not usually dramatic — it is an agent confidently doing the wrong thing at scale. That is why every agent we build has a human checkpoint before anything that reaches a customer or changes a record that matters.
AI agent vs chatbot vs automation
| Chatbot | Automation | AI agent | |
|---|---|---|---|
| How it works | Scripted responses | Fixed rule-based steps | Reasons, plans, acts |
| Handles the unexpected? | No — stops at script edge | No — breaks on new input | Yes — adapts within limits |
| Takes action? | Replies only | Executes set workflow | Multi-step across systems |
| Best for | FAQs | Predictable, stable tasks | Variable, multi-step work |
How we build agents: approval gates, audit trails, graduated autonomy
Three design choices make an agent safe to run in a real business, and we build all three in from the start.
- Approval gates — the agent classifies, drafts and prepares; a person approves anything with consequences. We gate the send, not the thinking.
- Audit trails — every decision and action is logged, so you can see exactly what the agent did and why. This is both good practice and a compliance requirement.
- Graduated autonomy — the agent starts narrow and supervised, and earns wider scope only after it performs consistently. Autonomy is a destination, not a starting point.
Single agents vs multi-agent systems
Most first projects need one well-scoped agent. Multi-agent systems — several specialised agents coordinating on a larger task — are powerful but add complexity and cost, and are rarely the right place to start. We recommend a single agent that proves the value, then expand only when the workload genuinely justifies it. See multi-agent systems: when a single agent is not enough.
What does an AI agent cost to build and run?
Build cost depends on how many systems the agent touches and how much data preparation it needs. Running costs — model usage and hosting — are typically modest for small-business agents. We size both during a fixed-fee consultancy so you see the full picture before committing. For worked examples, see our guide on agent architecture choices.
Frequently asked questions
What is agentic AI vs generative AI?
Generative AI produces content when prompted; agentic AI decides what to do next and takes action towards a goal, iterating on its own. Agentic AI usually uses a generative model as its reasoning engine, but adds planning, tool use and action on top.
How does agentic AI work?
An agent runs a loop: perceive inputs, reason about options against its goal, act using the tools it has, then check the outcome and adjust. Rules and approval gates constrain what it is allowed to do at each step.
Are AI agents safe for customer contact?
They can be, with the right guardrails: a clear escalation path to a human, approval on anything sensitive, and a narrow, well-tested scope. We do not recommend letting an agent take irreversible action with a customer unsupervised.
What are examples of agentic AI in small businesses?
Common ones: an agent that qualifies inbound leads and books them in, one that drafts quotes from your price rules, and one that answers routine customer questions from your documents and hands off the rest to staff.
Written by Austen Jones, Managing Director at York Apps. Published 6 May 2026 · Last updated 8 July 2026.