An AI agent is software that uses an AI model (typically a large language model) to plan, act and use tools to achieve a goal, not just answer a single prompt. Where a chatbot responds with text, an agent reads, decides, calls APIs, writes files, sends emails, runs code and persists what it has learned across steps.
The term is contested. Some vendors call any LLM-powered chatbot an "agent". Others reserve the word for systems that demonstrate multi-step autonomy and tool use. The honest read in 2026 is that the boundary is fuzzy and the marketing has moved faster than the engineering, so the word "agent" carries more weight than the underlying systems often deserve.
Concrete examples that are genuinely agentic: AI coding assistants like Claude Code, Cursor and GitHub Copilot Workspace that read a codebase, plan a change, edit multiple files and run tests. AI deep-research tools that decompose a question, search the web, read sources and synthesise an answer. Customer-service agents that read a ticket, look up the customer record, draft a reply, and escalate when confidence is low. Internal-knowledge agents that use a search tool, a database tool and a document-generation tool to answer a complex question with citations.
When the task genuinely requires multi-step reasoning and tool use (not just one model call). When the underlying tools and APIs are well-documented and stable. When a human review step can be inserted at the end (or at key decision points) without breaking the workflow. When you want adaptive behaviour rather than rigid scripted logic. When the cost of one wrong agent call is recoverable. Fixed-quote bands for SME-scale agent builds: £6,000 to £25,000 depending on the tool set and the integration depth.
When a deterministic workflow (a Zapier flow, an Apps Script, a custom integration) would do the job more reliably. When the cost of an autonomous wrong decision is severe (regulated, safety-critical, irreversible). When the task is well-defined enough that the LLM's planning step adds latency and cost without producing better outcomes than a hard-coded process. When the vendor cannot describe what happens when the agent gets stuck or makes a wrong tool call; that opacity is a real risk.
Digital Signet uses AI agents internally as part of how we ship work, and we build agentic systems for clients where the task genuinely benefits from multi-step reasoning. We are honest about which workflows fit and which would be better served by a simpler deterministic build. The delivery shape is on the AI implementation page. For the wider context on what AI can and cannot do today, the AI tools for small business guide covers the honest landscape.
Considering an AI agent build? Tell us the task and the tools involved. We will tell you whether an agent is genuinely the right architecture, or whether a simpler automation would do.
Email oliver@digitalsignet.com