Chatbot0.0s
What is agentic AI?
Agentic AI is software that carries out multi-step work toward a stated outcome, choosing which tools to call and in what order, rather than only answering a question. Agentic AI reads the state of a job, picks the next action, runs it, checks the result, then either continues or stops and asks a person. The outcome is the instruction. The route is decided at run time.
That single difference is what separates agentic AI from the automation most businesses already own. A scheduled script does what it was told, in the order it was told, and fails when reality does not match the diagram. An agent handles the case where the number of steps is not known in advance, because a supplier replied in a different format, or a record was missing, or the answer required checking three systems instead of one. We build agentic AI systems for South African businesses from Cape Town, for 35+ companies over 3+ years.
What is the difference between a chatbot, a workflow and an agent?
A chatbot answers. It takes a question, produces a response, and the work of acting on that response stays with the person reading it. A workflow executes a path somebody drew in advance, the same way every time, which is exactly what you want when the path never changes. An agent decides the path at run time from the goal, the tools available, and whatever it finds along the way.
The practical test is whether the steps are known before the job starts. If they are, workflow automation is cheaper to build, faster to run and far easier to debug at 02:00. If the number and order of steps genuinely depend on what the system discovers mid-job, that is where AI agents earn their keep. Most real systems we build are mostly workflow with a small agentic portion, not the other way around.
When is agentic AI the wrong choice?
Agentic AI is the wrong choice more often than the market admits, and saying so early saves everyone a bad quarter. It is wrong when the process is already deterministic, because a rules path handles a known sequence with less cost and no ambiguity. It is wrong when the action is irreversible and nobody reviews it, since a payment, a deletion or an outbound commitment deserves a human hand on the release.
It is wrong when latency matters more than flexibility, because a reasoning loop that calls several tools will never beat a direct lookup on speed. It is wrong when nobody can state what a good outcome looks like, since an agent with a vague goal produces confident activity rather than finished work. We test all four before proposing anything. If the honest answer is that a fixed workflow solves it, that is what we build and what we say.
How does an agentic AI system actually run?
An agentic AI system runs a loop. Read the goal and the current state, choose a tool, call it, read what came back, decide whether the outcome has been reached. Tools are ordinary integrations with defined inputs and outputs: a CRM lookup, a document fetch, a database write, a message send, a calendar check. The agent may only call the tools it has been handed, which is what keeps the loop inside a fence.
State is the part most demonstrations skip. A run needs to remember what it already tried, so a failed call is retried differently rather than repeated forever, and so a person picking up an escalation can see the trail. We assemble these systems with n8n, Make.com and custom services, with reasoning handled by OpenAI, Anthropic Claude or Google Gemini. Where one goal splits across specialists, a multi-agent team divides the tools between them.
What guardrails does agentic AI need?
Agentic AI needs three things before it touches production: a boundary on what it can reach, a record of what it did, and a condition that makes it stop. Tool access is scoped so an agent can only call the systems its job requires, and write access is kept narrower than read access. Actions that commit the business to something wait for human approval before they execute.
Step limits and timeouts end runaway loops rather than letting them bill through the night. Every call is logged with its inputs, outputs and timing, so a run can be reconstructed afterwards instead of guessed at. Consent, retention and access rules follow POPIA, the same way we treat any personal data we handle. Secrets stay outside the model context. None of this is exotic, it is the operational discipline any system with write permissions has always needed.
How does a business start with agentic AI?
Start with one process that costs real hours and has a finish line somebody can describe in a sentence. We map how it runs today, then mark each step as fixed or judgement. The fixed steps get automated first, because they are the cheap, reliable win and they shrink the problem before any agent is involved.
If the remaining decisions genuinely vary from job to job, we add an agent for that portion only, with tool access scoped and approval points agreed in writing. The pilot runs on the business's own systems for two to four weeks, with logs reviewed together so the behaviour is observed rather than assumed. What works gets widened, what does not gets cut. The business owns the workflows, prompts and data. Teams comparing approaches for the local market can also read our view on agentic AI in South Africa.
Related capabilities. The same parts, your business.
Keep reading. Pages close to this one.
Tell us what runs slow. We build what fixes it.
Send one message describing the process that eats hours. We reply with an honest read on whether it needs an agent, a workflow, or neither, and what it would take to build.