The agent hype cycle is in full swing
Research Nester forecasts the agentic AI market growing 40% annually from $8.6bn in 2025 to $263bn by 2035. Every software vendor is now either adding "agentic" to their product name or launching an "AI agent platform." The noise is extraordinary.
Underneath the noise: agentic AI is genuinely transforming some workflows for UK businesses right now. The key word is *some*. The pattern we're seeing is that the businesses getting real value from agents are solving a very specific type of problem. The businesses burning budget on agents are trying to solve a different type of problem with the wrong tool.
What agents are actually good at
An agent handles tasks that have three characteristics:
Multi-step. The task requires a sequence of decisions and actions, not a single lookup or generation. "Read this document, extract the key data, run a compliance check against our criteria, draft an approval email, and route it to the right person" — that's an agent task. "Summarise this document" — that's a prompt.
Tool-using. The task requires the AI to interact with external systems — read from a database, write to a CRM, send an email, call an API. A chatbot that only generates text isn't an agent.
Judgment-requiring. Some step in the process requires a decision that can't be reduced to a lookup or a rule. "Is this customer onboarding document complete enough to proceed?" requires judgment. "Is this field blank?" requires a check.
If a workflow has all three characteristics, an agent is likely the right tool. If it doesn't — if it's a linear sequence of lookups and transformations with no judgment required — workflow automation (n8n, Make, custom code) is simpler, cheaper, and more reliable.
The three agent patterns we're seeing work in UK businesses
Intake and triage. Agents that process incoming requests — leads, support tickets, compliance documents, procurement requests — qualify or classify them, extract structured data, and route them to the right human or next step. High volume, repetitive judgment, measurable outcomes.
Research and summarisation. Agents that monitor sources (news, competitor sites, regulatory updates, tender portals), synthesise relevant information, and surface it to the right person at the right time. The "reading pile" that no one has time for.
Coordination and dispatch. Agents that match supply to demand — driver to job, advisor to case, resource to request — based on multiple variables, proposing the match for human approval. The logistics dispatch pattern we've written about previously.
What's still theatre
Agents-as-chatbots. A chatbot that answers support questions is not an agent. It doesn't take action. It doesn't coordinate. Calling it an "AI agent" is a marketing decision.
Replacing judgment with automation. Any vendor pitching that an agent will make your business decisions is selling you something dangerous. The agents that work are the ones with human approval gates on the decisions that matter. Autonomy is a dial, not a switch.
"We'll figure out the use case later." We won't take this work. The businesses that deploy agents successfully know exactly which workflow they're targeting and can measure whether it's working. The businesses that fail spend three months "exploring" and can't point to a single concrete output.
The honest recommendation
If you have a specific, high-volume workflow that requires multi-step judgment and tool use, an agent is likely the right investment and the ROI is calculable before you build anything. Book a discovery call and we'll tell you honestly whether it makes sense. If you don't have that specific workflow identified yet, start there — not with the agent.