The starting point
What is agentic AI?
A plain-language definition, written for people who have to decide something. No product names, no forecasts, and no claim that you should be doing this.
On this page
The definition
Agentic AI is software that pursues a goal over several steps, decides for itself what to do at each step, and uses tools or systems to do it. An AI agent is one instance of that, scoped to a job.
The word that matters is decides. In conventional software you write the sequence and the machine follows it. In an agent you write the objective and the boundaries, and the machine picks the sequence. That single change is what makes these systems useful on messy work, and what makes them harder to govern than anything that came before them.
What it is not
- Not a chatbot
- A chatbot returns text and a person acts on it. An agent acts. The moment software issues the refund rather than describing it, your questions move from accuracy to authority, oversight and reversibility.
- Not automation or RPA
- Rule-based automation follows the route you specified and breaks loudly when reality differs. An agent picks a route at run time: it copes with variation, and it can fail quietly by picking a plausible wrong one. You trade brittleness for a supervision requirement.
- Not a replacement for judgment
- These systems are good at volume, consistency and recall. They have no stake in the outcome and no sense of what would be embarrassing. Anything where being wrong is expensive stays with a person.
- Not artificial general intelligence
- Nothing on the market reasons across domains the way a competent colleague does. Treat any claim in that direction as marketing.
Anatomy of an agent
Six parts. When an agent misbehaves, the cause is almost always in one of them, and rarely the one people blame.
- Objective
- What the agent is asked to achieve, and how it knows it has finished. Vague objectives are the most common cause of an agent doing something odd.
- Model
- The reasoning engine. It plans steps and interprets results. It is also the part that will state something false with complete confidence.
- Tools
- The systems it can call: a CRM, a ticketing system, a payment API, a search index. Tools are where an agent stops being a conversation and starts having consequences.
- Memory and context
- What it can see: the customer record, the policy document, previous steps. Most bad answers are missing-context problems, not model problems.
- Guardrails
- The limits on what it may do: value caps, allowed actions, blocked topics, mandatory escalation. Guardrails are the difference between a deployment and an experiment.
- Oversight and logging
- Who checks the work, how often, and what record exists afterwards. If you cannot reconstruct why the agent did something, you cannot answer for it.
Levels of autonomy
Autonomy is a dial, not a switch. Most working deployments sit in the middle two positions, and most disappointment comes from buying one level and operating another.
| Level | Who does the work | What it means |
|---|---|---|
| Assistive | The person does the work | The agent drafts, summarises, or suggests. Nothing leaves the building without a human hand on it. Low. The failure mode is wasted time, not a wrong action. |
| Supervised | The agent proposes, a person approves | The agent prepares a complete action — a reply, a refund, a record change — and a person releases it. Moderate. Approval fatigue is real: review quality falls as volume rises. |
| Semi-autonomous | The agent acts inside limits | The agent completes the whole task within a defined boundary — value caps, customer tiers, action types — and escalates anything outside it. The boundary is the control. If the limits are vague, this is autonomy with extra paperwork. |
| Autonomous | The agent acts, people audit | The agent runs the task end to end. Humans review samples and investigate exceptions after the fact. Highest. Only defensible where actions are reversible, logged, and cheaply corrected. |
Where it works, and where it does not
A useful rule: agents do well on work that is high in volume, low in variation, reversible when wrong, and testable for correctness. They do badly on the opposite. Nothing about the vendor changes that.
Suited
- Repetitive queries with a documented right answer
- Drafting a first version a person will edit
- Pulling scattered records into one place
- Routing, tagging, and chasing
Not suited
- Decisions that are hard to reverse
- Anything requiring a stake in the relationship
- Novel situations with no precedent to learn from
- Work where nobody can say what “correct” is
What that looks like in each domain is covered in the six field guides — marketing, customer service, inside sales, analytics, content creation and admin and operations. Each one carries its evidence and its limits.
Seven questions before you deploy one
- 01
What exactly is this agent allowed to do without a person, and where is that limit enforced — in the product, or in a prompt?
- 02
How do we know when it is wrong? Name the signal, not the intention.
- 03
Is every action it can take reversible, and how long does reversal take?
- 04
What does it see? Which records, documents and systems, and where does that data go?
- 05
Who is accountable for an action the agent takes at three in the morning?
- 06
What happens on the day it is unavailable — does the work stop, or fall back to people who no longer do it daily?
- 07
What is the total cost including review time, not just the licence?
Common questions
- What is agentic AI?
- Agentic AI describes software that pursues a goal over several steps, chooses its own actions along the way, and uses tools or systems to carry them out. The distinguishing feature is not the model. It is that the software decides the sequence rather than following one you wrote in advance.
- What is an AI agent?
- An AI agent is a single instance of that idea, scoped to a job: it takes an objective, breaks it into steps, calls tools or systems, checks its own progress against the objective, and stops when it is done or when it hits a boundary you set.
- How is an AI agent different from a chatbot?
- A chatbot answers. An agent acts. A chatbot returns text and the person does the work; an agent issues a refund, updates a record, or sends the message. The moment software takes an action with consequences, the questions you must ask change from accuracy to authority, oversight and reversibility.
- How is an AI agent different from automation or RPA?
- Traditional automation follows a route you specified; it fails loudly when reality does not match the route. An agent chooses a route at run time, so it copes with variation but can fail quietly by choosing a plausible wrong one. You trade brittleness for a supervision requirement.
- Are AI agents reliable enough to use at work?
- For bounded, high-volume, reversible tasks with a clear correctness test, the evidence is reasonably good. For open-ended judgment, novel situations, or irreversible actions, it is not. Reliability is a property of the task you give the agent, not of the agent.
- What are the levels of autonomy?
- In practice there are four: assistive, where a person does the work and the agent drafts; supervised, where the agent acts and a person approves each action; semi-autonomous, where the agent acts within limits and escalates exceptions; and autonomous, where the agent acts and a person reviews samples after the fact. Most working deployments sit in the middle two.