An AI agent is software that carries out a defined task on its own: it receives a trigger, works out what needs to happen, takes action inside your systems, and knows when to hand off to a person. Unlike a chatbot, which answers questions, an agent does work. Unlike traditional automation, which follows fixed rules, an agent can handle situations that need judgment, such as reading an unstructured email and deciding how to route it. In business, the agents that actually deliver are narrow, bounded and supervised, not general-purpose AI let loose on your operations.
What is an AI agent?
An AI agent is software that acts on your behalf to complete a specific task. It’s triggered by something happening, an email arriving, a form being submitted, a deal being won, then it takes in the relevant information, reasons about what should happen next, and carries out actions in your systems. When something needs judgment beyond its remit, it routes to a person.
The useful mental model is a capable new team member with a very specific job description. They can handle their task without you watching every step, they use the same tools everyone else uses, and they escalate anything unusual. What they don’t do is invent their own priorities or work outside their brief.
How is an AI agent different from a chatbot?
A chatbot responds to questions in a conversation. You ask, it answers, and the value stops at the reply. An AI agent completes work in your systems, and often you never speak to it at all. It runs in the background: capturing an enquiry, updating a record, chasing an approval, assembling a report.
The practical difference is where the value lands. A chatbot saves you looking something up. An agent removes the task entirely.
How is an AI agent different from automation?
Traditional automation follows fixed rules: when this happens, do that. It’s fast, cheap and extremely reliable, and for predictable work it’s still the right tool.
An AI agent adds reasoning. It can deal with inputs that don’t come in a fixed shape, an email written in someone’s own words, a document laid out differently every time, an enquiry that could be one of several things. Where a rule would break, an agent can interpret and decide.
The important part, and where most projects go wrong: AI is not the upgrade to automation. It’s a different tool for a different job. Pointing an agent at work that a simple rule handles perfectly adds cost, latency and unpredictability for no benefit. We use deterministic automation wherever the work is predictable, and reserve agents for the steps that genuinely need judgment. Most real workflows are a mix of both. For fuller definitions of the wider picture, see What is workflow automation? and What is operational system design?.
What does an AI agent actually consist of?
Every agent that works in production has six parts:
A trigger. The specific event that starts it. Vague triggers produce vague agents.
Inputs and context. The information it needs, from your systems, documents or the trigger itself.
A bounded task. One clearly defined job. “Triage incoming enquiries and route them” is a task. “Improve our sales process” is not.
Permitted actions. The specific things it’s allowed to do, and the systems it can do them in. Everything else is off limits by design.
Human checkpoints. The points where a person reviews, approves or takes over. High-stakes steps always route to a human.
A log. A record of everything it did and why, so you can audit it and improve it.
If any of these is missing, the agent will disappoint you.
What can AI agents do in a business?
The agents that deliver most tend to sit on high-volume, coordination-heavy work:
- Intake and triage: capturing enquiries and requests, working out what they are, routing them to the right person
- Document processing: reading documents, extracting the details that matter, pushing them into your systems
- Follow-up and chasing: pursuing outstanding information, approvals and confirmations
- Reporting: gathering data across systems and surfacing the exceptions that need attention
- Drafting: producing first-pass documents, updates and communications for a person to review
- Knowledge capture: turning meetings and activity into structured records
Notice what’s not on that list: decisions with real consequences. Those stay with people.
What does this look like in practice?
Two agents we’ve built, and why they worked.
Lead triage and prioritisation. A business was receiving a steady flow of new leads into its CRM workflow, and senior leaders were spending time assessing each one to decide which were worth pursuing. We connected an agent to that workflow that draws on the company’s own data, lead types, client types, won opportunities and closed-out leads, to assess each new lead against what has historically been a good fit and prioritise accordingly.
The effect was threefold. Senior people stopped triaging every inbound lead, the team focused effort on the leads most likely to convert, which improved conversion, and the company’s own historical data started being used to sharpen the process rather than sitting unused in a system. Note what makes this work: the judgment is genuinely uncertain, which is why an agent belongs here rather than a rule, and the agent reasons from the business’s real data rather than generic assumptions.
Scoping call to draft proposal. In this workflow, scoping calls are recorded. Once a recording completes, a deterministic workflow triggers via a webhook and passes the meeting through for AI assessment. The agent uses what was actually discussed to produce a discovery quote and presentation from a predefined template, which a person then reviews and edits. Once a human has verified it, the deterministic workflow takes over again, finalises the document and prepares a draft email ready to send to the prospect.
This one is worth studying, because it shows the whole principle in a single workflow. The sequence, the trigger, the document generation, the email preparation, is all deterministic, because those steps are predictable and deterministic automation is more reliable and cheaper for them. The agent handles exactly one thing: interpreting an unstructured conversation and turning it into a tailored draft, which no rule could do. And a person verifies the output before anything reaches a client. Deterministic where it can be, AI where it must be, human where it matters.
Where should judgment stay with a person?
Anywhere the cost of being wrong is high. Financial approvals and payments. Legal, clinical and professional advice. Anything client-facing where tone and accuracy matter. Hiring and people decisions. Anything with regulatory weight.
A well-built agent doesn’t just avoid these, it recognises when it’s reached one and routes it. Agents should also flag uncertainty rather than pushing through it. An agent that says “I’m not sure about this one” is working correctly.
Why do AI agent projects fail?
In our experience, four reasons, and none of them is the technology:
The scope was vague. An agent asked to “handle customer service” has no defined job. Narrow beats ambitious every time.
The underlying process was broken. An agent running a bad process just runs it faster. Structure first, then automate, then add AI.
AI was used where automation belonged. Expensive, slower and less reliable than a rule would have been.
There was no human oversight designed in. Either it was trusted too far and made a costly mistake, or it was trusted too little and everyone quietly stopped using it.
How do you start with AI agents?
Start with one workflow, not a strategy. Pick something high-volume and low-risk where people currently spend real time on repetitive coordination, intake, chasing, or document handling. Map how the work actually flows today. Automate what’s predictable, put an agent only on the steps needing judgment, and keep a person on anything consequential. Pilot it on one real workflow, measure it, then extend.
The businesses that succeed with agents almost never start with “we need AI.” They start with “this specific work eats too much of our week.”