An AI agent decides its own next step. Most of your admin does not need that.
An AI agent is software that uses a language model to decide its own next step towards a goal — which tool to call, what to read, when to stop — rather than following a script written in advance. That separates it from a workflow, where the steps are fixed. Most small-business admin is better served by the workflow.
The definition, and the one line that matters
The cleanest published definition draws a line between two things that get sold under one name. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths.” Agents are “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks” (Anthropic, Building effective agents, 19 December 2024).
The difference is who holds the steering wheel. In a workflow, you wrote the steps and the model fills in a blank inside them. In an agent, the model reads the situation, picks a step, checks the result and picks the next one. That freedom is the point, and it is also the cost.
The same paper names five workflow patterns — prompt chaining, routing, parallelisation, orchestrator-workers and evaluator-optimiser — and advises building the simplest thing that works before reaching for an agent. That advice comes from a company that sells the models agents run on. Take it seriously.
How few Australian businesses actually use one
Around one business in eight uses any AI at all. The ABS Characteristics of Australian Business 2024–25 (released 25 June 2026) found 12% of businesses reported using AI in their workplace in 2024–25. Among small and micro businesses it was around 11%; medium businesses 22%; large businesses 35%, up from 9% in 2021–22.
Use tracks with whether a business changes anything else. Innovation-active businesses reported AI use at 20% against 6% for the rest (same release). The same survey found 21% of businesses had a cyber security incident in the year, which matters when you are about to hand a piece of software the keys to your inbox.
Those figures cover any AI, including a staff member using a chatbot to draft an email. The share running a true agent — software acting on its own inside your systems — is not measured by the ABS, and we are not going to guess at it.
Where an agent earns its keep, and where a workflow does
The test is simple: can you write the steps down in advance? If yes, build a workflow. If the steps depend on reading each case and deciding, an agent is the honest tool. Most jobs in a small business fall on the first side of that line.
- KNDR·01Workflow: quote to job to invoiceThe steps never change. Quote accepted, job created, invoice raised in Xero. There is nothing for an agent to decide, so paying agent prices for it is waste. See quote to invoice.
- KNDR·02Workflow: debtor remindersDay 7, day 14, day 21, escalate. A rule handles it. The only judgement call is which customers to exclude, and a human should make that one. See debtor chasing.
- KNDR·03Agent, cautiously: inbox triageEvery email is different. Reading one, deciding whether it is a quote request, a complaint or a supplier invoice, and routing it — that is a decision per item, which is what a model is for. See inbox triage.
- KNDR·04Agent, cautiously: intake from messy sourcesVoicemail transcripts, photos of a handwritten job sheet, a customer’s half-finished web form. Extracting the fields is judgement work. Writing them to the job system afterwards is a workflow again.
The failure rate nobody puts on the brochure
Most agent projects will not survive. Gartner predicted on 25 June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Its analysts described most current projects as early-stage experiments “mostly driven by hype.”
The same release coined the term “agent washing” for vendors relabelling chatbots and robotic process automation as agents, and estimated only around 130 vendors out of thousands sell something genuinely agentic. If a product page says “agent” and the demo shows a fixed sequence of screens, it is a workflow with a new sticker.
Gartner also expects 15% of day-to-day work decisions to be made autonomously by agents by 2028, from 0% in 2024, and 33% of enterprise applications to include agentic AI by then, from under 1%. Both numbers can be true at once: the technology is arriving, and most first attempts at it will be written off.
The rules that apply when an agent touches personal information
An agent that reads your inbox reads your customers’ personal information. The OAIC guidance on commercially available AI products (21 October 2024) recommends organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools, and that privacy policies say plainly where AI is used.
From 10 December 2026, organisations covered by the Privacy Act must disclose in their privacy policy where automated decision-making uses personal information to make decisions that significantly affect people (MinterEllison, on the Privacy and Other Legislation Amendment Act 2024). Whether the Act covers you at all depends on turnover — the Australian Privacy Principles, explained for SMBs walks through the $3 million line.
The public wants a person in the loop. In the OAIC Australian Community Attitudes to Privacy Survey (August 2023, 1,916 respondents), 96% wanted conditions such as a right to human review before AI is used in decisions affecting them. The National AI Centre’s Guidance for AI Adoption (17 October 2025) turns that into six practices, the last of which is “maintain human control.” It replaced the 2024 Voluntary AI Safety Standard and its ten guardrails.
What we have built, and what we have not
We have built automation against nine systems: Xero, MYOB, QuickBooks Online, ServiceM8, Simpro, Stripe, GoHighLevel, Zapier and n8n. Most of that work is workflow, not agent, because most of the jobs were workflow jobs. Where a model sits inside a build, it usually does one bounded task — classify, extract, draft — and hands back to fixed steps.
We have not published measured results from an agent running in a client’s business. Kindra AI was founded in 2025. When those measurements exist they will be on this site with the sample size attached, and not before.
One constraint shapes every agent we would build for you: the systems it talks to have hard limits. Xero allows 60 API calls a minute and 5,000 a day per organisation (Xero developer limits FAQ). An agent that “just checks” too often runs into that wall by lunchtime. The integration capability register lists what each system will and will not allow.
How to tell an agent from a chatbot with a new label
Ask what it can do without you. A chatbot answers when spoken to. An agent takes an action — creates the job, sends the reminder, files the document — and then decides what to do next based on the result. If every step needs a human click, it is an assistant.
Ask what it talks to. Agents work through APIs, and the developers building those APIs are only starting to design for them: 24.3% of respondents to the Postman 2025 State of the API Report (8 October 2025, over 5,700 respondents) were designing APIs with AI agents in mind, and only 10% used the Model Context Protocol regularly. An agent is only as capable as the doors it has keys to — what an API integration is explains those doors.
Ask what happens when it is wrong. It will be, sometimes. The useful questions are whether the build logs every action, whether a person can reverse it, and whether it stops when unsure rather than guessing.
The questions people actually ask.
01Is ChatGPT an AI agent?
On its own, no. A chat window that answers questions is an assistant. It becomes an agent when it is given tools — your calendar, your job system, your email — and permission to use them in sequence without asking each time. The same model can be either; the wiring around it decides.
02Will an agent replace my office admin?
Not one we would build. The tasks an agent handles well are the ones your admin does on autopilot: sorting, extracting, drafting the first version. The judgement calls — the customer who is always late but always pays, the quote that needs a phone call first — stay with a person. The honest promise is fewer hours on the boring half, not an empty desk.
03Do I need one to keep up with competitors?
The ABS says 11% of small businesses used any AI in 2024–25. You are not behind. Fix the process first, automate the fixed process, and add an agent only where a decision per item genuinely needs making. Buying an agent to feel current is how projects end up in Gartner’s 40%.
Stop bending.
Start shipping.
Bring the one flow that keeps costing you a Tuesday afternoon. We map it live and tell you what we'd build, what we wouldn't, and what it costs.