AI & Data Science10 min read

Almost every AI product is now sold as agentic, which makes the word hard to trust. The difference is real, though, and it matters to anyone deciding what to build or buy, because it changes what the software is allowed to do, what can go wrong and what it costs to run.
The short version: generative AI produces content in response to a request and then stops. Agentic AI is given a goal, plans the steps, uses tools such as search, documents and business systems to carry them out, checks the result and keeps going until the goal is met or a person needs to decide. Most useful AI agents are built on generative models. What makes them agentic is the loop around the model and the access it is given.
What generative AI does
Generative AI models learn patterns from very large amounts of text, code or images and use them to produce new content. Ask for a summary of a contract, a draft reply to a customer or an explanation of a policy and you get one. ChatGPT, Microsoft Copilot in Word and Claude are all generative tools in everyday use.
The important feature is that a person stays in the loop by default. The model suggests and a person decides whether to send the email, accept the code or file the summary. That makes generative AI comparatively low risk to adopt. The main failure is a confident wrong answer, which is why the most useful business deployments ground the model in the organisation's own documents and show where each answer came from.
What makes AI agentic
An agent wraps a model in a loop that lets it work towards a goal with less supervision. Four things turn a generative model into an agent:
- A goal rather than a single prompt, such as resolving a support ticket or reconciling this week's invoices
- Tools it can call, such as search, a database query, an email account or an API into a finance system
- Memory of what it has already done and found, so it can plan the next step
- Permission to act, within limits, rather than only to suggest
Give a model those four things and it can read an incoming request, look up the customer, check the order in the ERP, draft the response, raise the refund and record what it did. That is valuable, and it is also why agents need more care than a chat window. An agent that can act can act wrongly, and the cost of a mistake is no longer a bad paragraph but a wrong record, a wrong payment or a message sent to the wrong person.
| Generative AI | Agentic AI | |
|---|---|---|
| What you give it | A prompt or a document | A goal and a set of tools |
| What it returns | Content: text, code, images, summaries | Completed work: records updated, tasks done |
| Who acts on the output | A person, every time | The agent, with a person approving key steps |
| Access to your systems | Usually none, or read only | Read and write, scoped to what it needs |
| Main risk | A plausible but wrong answer | A wrong action taken at speed |
| Controls it needs | Grounding in your documents, source citations | Boundaries, approvals, logging, testing on real cases |
| Good first uses | Drafting, summarising, answering staff questions | Triage, data entry, reconciliation, routine follow up |
Agentic AI and AI agents are not quite the same thing
The terms are used interchangeably, but there is a useful distinction. An AI agent is a piece of software: a support agent, a scheduling agent, an invoice agent. Agentic AI is the approach, meaning systems designed to pursue goals with some autonomy, often by coordinating several agents. A larger agentic system might have one agent that reads documents, another that checks them against policy and a third that prepares the record for a person to approve.
For a buyer, the practical question is not which label a vendor uses but what the software is allowed to do on its own, what it can reach, and how you will know what it did.
Where each one earns its keep
Generative AI is the right tool when the work ends with a person reading something. Drafting proposals and replies, summarising long documents, answering staff questions from policies and manuals, and turning meeting notes into actions all fit. These projects are quick to deliver and easy to judge, which makes them a good first step.
Agentic AI earns its keep where work follows a known process across several systems and is done often enough that the minutes add up. Classifying and coding expenses, triaging inbound requests, checking documents against rules, and preparing records for approval are typical. The agents we have built for clients follow this pattern. The customer service agent for the Young Endeavour Youth Scheme answers applicants from authoritative documents with human controls in place, and the financial data governance platform for Beach Energy classifies expenses and flags anomalies with a person reviewing the results.
Why so many agent projects stall
Interest is running well ahead of results. Deloitte's 2026 State of AI in the Enterprise survey found that 69 percent of Australian organisations say they are using agentic AI, but only 28 percent have moved 40 percent or more of their AI pilots into production. Gartner expects over 40 percent of agentic AI projects to be cancelled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls.
- Using agentic AI69%
- 40% or more of pilots in production28%
Source: Deloitte, State of AI in the Enterprise 2026, survey of more than 3,000 director to C suite leaders.
In our experience the pilots that stall share a pattern: the agent was given a broad goal instead of a narrow process, nobody agreed how accuracy would be measured, and the controls were left until the end. All three are cheaper to settle before the build than after it.
The controls an agent needs before it acts
Every agent we build has the same five controls, because they are what make autonomy safe enough to use. They line up with the six essential practices in the National AI Centre's Guidance for AI adoption, which ends with maintaining human control. If a vendor cannot describe each one for their product, the product is not ready to touch your systems.
- A boundary: the agent reaches only the systems and data it needs, ideally in Australian hosting or your own environment
- Approval before consequential actions: anything that affects customers, staff, money or compliance waits for a person
- A log of every step: what the agent read, what it decided and what it changed, with the source
- Testing on real cases: the agent is measured against your own documents and agreed accuracy targets before it works unsupervised
- Privacy by design: from 10 December 2026, organisations covered by the Privacy Act must explain in their privacy policy when automated decisions significantly affect people, and the agent's log is the record you need
What we would do
Start generative. Pick one team and one body of documents, put a grounded assistant in front of it, and measure whether answers are right and time is saved. That project teaches you what your data is like and builds the trust you will need later.
Then make it agentic one step at a time. Give the assistant a single tool, such as raising a ticket or updating one field, with a person approving every action. Widen what it can do on its own only as the log shows it getting things right. That path is slower than a big launch, but it is the one that ends with an agent your team relies on. The agents we have built for government and business all started this way.


