What is agentic AI, and what can it actually do for a business?
Agents differ from chatbots in one specific way: they take actions. Here is what that changes, where it works today, and where it still fails.
Key takeaways
- An AI agent is a system that can take actions in other software, not just generate text.
- The practical difference from a chatbot is tool use: an agent can query a database, file a ticket, or send an email.
- Agents work well on high-volume, well-bounded tasks with a clear definition of correct.
- They fail on tasks where being wrong is expensive and hard to detect, which is why evaluation matters more than the model.
The word 'agent' has been stretched to cover almost anything with a language model attached. The distinction that actually matters is narrow: an agent can take actions in other systems. A chatbot answers you. An agent goes and does something, then tells you what it did.
How an agent differs from a chatbot
A chatbot receives text and returns text. An agent receives a goal, decides which tools to call, calls them, reads the results, and repeats until it believes the goal is met. Those tools are ordinary software functions, a database query, an API call, a calendar write.
This is why agents can be genuinely useful and genuinely dangerous in the same breath. A chatbot that hallucinates gives you a wrong answer. An agent that hallucinates issues a refund.
Where agents work today
- Triage and routing: reading an incoming request, classifying it, and putting it in the right queue with the right context attached.
- Research and summarisation across your own documents, with citations back to the source.
- First-draft generation for repetitive artefacts: job descriptions, release notes, RFP responses.
- Data entry between systems that have no integration and are unlikely to get one.
The common thread is that all four are high-volume, well-bounded, and easy to check. A human reviewing an agent's classification takes seconds. A human verifying an agent's financial reconciliation takes as long as doing it themselves, which removes the point.
Where they still fail
Agents degrade badly on long task chains, because a small error at step two compounds by step nine. They struggle when the definition of success is subjective. And they are poor at recognising the boundary of their own competence, which is why confidence thresholds and human escalation paths are not optional extras.
Why evaluation matters more than model choice
Teams spend weeks choosing between models and no time at all building the harness that tells them whether quality has changed. This is backwards. Models get swapped and updated constantly, often without your involvement if you are using a hosted API. The only durable asset is a suite of test cases, reviewed by someone who knows the domain, that runs on every change.
On one clinical project we shipped a 340-case evaluation suite alongside the feature. It has since caught two quality regressions introduced by upstream model updates, neither of which would have been visible until users complained.
A sensible way to start
- 01Pick one task that happens at least fifty times a week and has an obvious right answer.
- 02Write down how you will measure success before building anything.
- 03Build the evaluation set first, from real historical examples.
- 04Ship to a small internal group with a human review step still in place.
- 05Remove the review step only when the numbers say you can.
- Agentic AI
- Automation
- LLM
Questions this raises
- What is agentic AI?
- Agentic AI describes systems that can take actions in other software rather than only generating text. An agent receives a goal, decides which tools to call, executes them, reads the results, and repeats until the goal is met. The defining capability is tool use.
- What is the difference between an AI agent and a chatbot?
- A chatbot receives text and returns text. An agent can act: querying a database, filing a ticket, or sending an email. The practical consequence is that a chatbot's mistakes produce wrong answers, while an agent's mistakes produce wrong actions.
- Is agentic AI safe to use in production?
- It is safe for tasks that are high-volume, well-bounded, and cheap to verify. It is not yet safe for tasks where an error is expensive and hard to detect. Production deployments should include confidence thresholds, human escalation paths, and an evaluation suite that runs on every change.
- How much does it cost to build an AI agent?
- A focused agent handling one well-defined workflow typically costs $30,000 to $80,000 to build, including the evaluation harness. Ongoing costs are model inference plus monitoring, which for most business workloads runs in the hundreds rather than thousands of dollars per month.