AI agents

AI agents that do real work in your business, and know when to hand over

We build and run AI agents that triage and answer support tickets, read contracts, watch markets and suppliers, and move work through your systems, acting when they're confident and handing over to a person when they're not. Here is what an agent is, when you need one, and how we build one you can trust.

Talk to us about an agentSee an agent in production
The short answer

An agent decides its own next step

Most business AI today suggests or executes: a copilot drafts and a person decides; an assistant carries out a task it is given. An agent is different in one way that matters: it plans and acts within boundaries you set, choosing which tools to use and what to do next until the job is done or it hands over to a person. Agentic AI is the name for systems built that way, and an agentic system is several specialised agents working together under policy controls to deliver one governed outcome.

That autonomy is the value and the risk. It is why an agent can resolve a support ticket end to end, and why it needs boundaries, measurement and a way to hand over that a copilot never did.

The four levels of AI

From copilot to agentic system

The scale we use in every programme, from CONVERGE. Most businesses get the most value from the first two levels first; agents earn their place where the work is repetitive, rule-bound and costly in people's time.

  • Copilot

    AI suggests, a person decides. Drafting, summarising, searching.

  • Assistant

    AI executes a defined task on instruction. Classify this, extract that, reply to this.

  • Agent

    AI plans and executes within set boundaries, using tools, and hands over when it should.

  • Agentic system

    Specialised agents, orchestrated under policy controls, deliver a governed outcome.

The four levels of AI in full

Workflow or agent?

Workflows before agents

Not every AI system should be an agent. In a workflow, the model and its tools are orchestrated through paths you define in code: predictable, testable and cheaper to run. In an agent, the model directs its own process. There are five well-established workflow patterns: prompt chaining, routing, parallelisation, orchestrator and workers, and evaluator and optimiser. Between them they cover most business processes.

So we start with the simplest design that does the job, and add autonomy only where the task genuinely needs it. A routing workflow that sends each invoice to the right approver is not less impressive than an agent; it is more likely to still be working next year.

What we build

Agents for the work that costs your people time

Each is built on frameworks we have built and run before, tailored to your systems.

How we build one

Scope, build, prove, then run and improve

  • Scope

    One to two weeks: the process, the systems, the escalation rules and the evaluation set, agreed before any code.

  • Build

    Three to eight weeks on our frameworks, connected to your data and tuned against the evaluation set.

  • Prove

    A bounded production pilot on real traffic, with people reviewing every escalation and the numbers reported weekly.

  • Run and improve

    Monitoring, evaluation on every change and improvement from the cases it hands over, as part of the service.

Trusting an agent in production

What makes an agent safe to run

  • Confidence-based escalation

    The agent acts when it is confident, hands the case to a person when it isn't, and records both the decision and the reason.

  • Evaluation before scale

    A set of real cases with known-good answers, drawn from real failures, run on every change before it ships.

  • Least privilege

    Access to only the systems and actions the job needs, with every action logged, so a mistake is contained and visible.

  • Governance from the start

    Risk category, data protection, human oversight and transparency decided for every use case before it is built.

An agent we built and run

Around 100 tickets a week. First response in minutes.

The support agent at Agency51 reads each ticket, works out whether it can answer from the client's own code and documentation, answers when it is confident and hands over to a person when it is not. It is one example of the agent-based systems we build.

  • ~100Support tickets a week
  • Hours → minutesFirst-response time
Questions

About AI agents

What is agentic AI?

AI that plans and acts towards a goal within boundaries you set, choosing its own next step and its tools, rather than only suggesting (a copilot) or carrying out a single instruction (an assistant). An agentic system combines several such agents under policy controls.

What is the difference between an AI agent and a chatbot?

A chatbot answers in a conversation. An agent does work: it looks things up in your systems, takes actions such as updating a record or drafting a reply, and decides what to do next. Many good agents never talk to a customer at all.

Does our business need an agent?

Often not first. Most of the early value is at the copilot and assistant levels, on tools you may already own. Agents make sense where the work is high-volume, follows rules you can write down, and costs people real time: support triage, document processing, routine operational decisions. A discovery programme tells you which, in your business.

How long does it take to build one?

On our frameworks, typically one to two weeks to scope and agree the evaluation set, then three to eight weeks to build, followed by a bounded production pilot with the numbers reported weekly.

Which models do agents run on?

We build on Claude, OpenAI and open-weight models and choose on the evidence for each use case. The design is model-agnostic, so changing model is a configuration change rather than a rebuild.

Have a process an agent could own?

Tell us what the work looks like today. We'll tell you whether it needs an agent, a workflow or neither, and what it would take to build and run.

Talk to us about an agentAI development