Plenty of Australian small businesses that use AI are using it to write. Emails, social posts, quotes that need tidying, a product description at 9pm. That’s useful. It isn’t an AI agent.
AI agents are a different thing. An agent takes a goal, works through the steps, uses your systems along the way and finishes the job: it books the appointment, sends the follow-up, updates the CRM, takes the deposit. A chatbot answers. An agent acts. That difference decides what you can safely hand over, what it costs, and how it goes wrong.
This piece looks at what the 2026 Australian data says small businesses are really doing with AI, where agents earn their keep, and the short list of things to write down before you switch one on.
Readiness checklist
Before you switch on your first AI agent
This is the one-page scope from the article below, item by item. Tick what is already true for your business, then look at what is left before the agent talks to a customer.
The job
Ownership and handoff
Customers and data
Testing and watching
Nothing ticked yet. Start with the job: if you cannot write it in one sentence, the rest will not hold.
Key Takeaways
- NAB’s April 2026 SME survey found 42 per cent of businesses use AI, 44 per cent don’t, and 14 per cent plan to start.
- Among NAB’s AI users, 51 per cent use it in marketing and sales and 24 per cent in customer service, including chatbots. By our arithmetic, that’s roughly one SME in ten using AI in customer service.
- The National AI Centre’s tracker puts SME adoption at 43 per cent for December 2025 to February 2026. Trust is the biggest barrier, cited by around 65 per cent of non-adopters.
- Anthropic separates fixed AI workflows from agents that direct their own steps, and warns that agents bring higher costs and compounding errors.
- Sinch’s May 2026 study of 2,527 senior decision-makers found 74 per cent of enterprises had rolled back or shut down an AI customer communications agent after deployment.
- The federal Guidance for AI Adoption sets out six essential practices. The first is deciding who is accountable.
An Agent Acts, a Chatbot Answers, and That Changes What You Can Hand Over
Three words get used as if they mean the same thing. They don’t.
A chatbot answers questions in a conversation. The older ones followed a decision tree. The newer ones run on a large language model and write fluent replies. Either way, the job ends at the answer.
A copilot sits beside a person and helps them work faster. It drafts the email; you send it. It summarises the contract; you decide. A person stays in the loop on every step.
An AI agent is given a goal and gets on with it. It decides which steps to take, calls the tools it has been given (your calendar, your CRM, your payment gateway) and checks its own progress along the way.
Anthropic, the company behind Claude, draws the line cleanly in its engineering guidance on building agents [1]. 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.”
That distinction matters more than the label on a vendor’s website. A workflow follows the path you designed. An agent chooses its own path inside the limits you set. The first is predictable. The second is flexible, and flexibility is exactly what makes it useful for messy, real-world jobs like a customer enquiry that doesn’t fit the form.
It also makes it harder to control. Anthropic is blunt about the trade: the autonomy of agents means “higher costs, and the potential for compounding errors” [1]. A wrong step early in a chain gets built on by every step after it. Its advice is to find the simplest solution that works and only add complexity when it’s needed.
For a small business, that leads to a practical rule. If the job follows the same steps every time, you probably want a workflow with AI in it, not an agent. If the job changes shape depending on who’s asking, an agent starts to make sense, as long as someone has drawn the boundaries first.
So when you’re shopping for AI agents for business, ask the vendor one plain question: what can this thing do without a human pressing a button? The answer tells you whether you’re buying a chatbot, a copilot or an agent, and it tells you how much governance you’ll need.
Four in Ten Small Businesses Use AI. Far Fewer Have Handed It a Process.
The headline adoption numbers are steady across the two biggest Australian trackers.
NAB’s SME Business Insights report, published in April 2026 from a survey of about 670 non-farm SMEs taken between 10 November and 9 December 2025, found 42 per cent of businesses use AI today, 44 per cent don’t, and 14 per cent plan to introduce it [2]. The National AI Centre’s monthly tracker, run by Fifth Quadrant with at least 400 SMEs a wave, puts adoption at 43 per cent across December 2025 to February 2026, with February alone at 44 per cent [3].
Four in ten sounds like a lot. Look at what the four in ten are doing, though, and a different picture comes through.
Among NAB’s AI users, the most common area is marketing and sales at 51 per cent, then operations and logistics at 39 per cent [2]. Customer service, which NAB defines to include chatbots and automated support, sits at 24 per cent. Finance and accounting is 22 per cent. HR and recruitment is 10 per cent.
Run the arithmetic across the whole sample and customer service AI reaches about one SME in ten (42 per cent of businesses using AI, times 24 per cent of those using it in customer service). That’s our calculation from NAB’s figures, not a number NAB publishes, but it makes the point. Most small business AI is still about producing content and analysis, not running a process that touches customers.
The National AI Centre’s data says the same thing from another angle. Among businesses using or planning to use AI, content generation and data analytics lead at 54 per cent each, with cybersecurity and threat detection at 48 per cent [3]. The Centre describes these top uses as “productivity-oriented and relatively low-risk to experiment with.”
Low risk is the key phrase. Writing a draft is easy to check. A booking taken at 11pm by software, with a deposit attached, is not.
Why the gap? The National AI Centre lists three barriers [3]. Trust is the biggest: around 65 per cent of non-adopting businesses cited distrust of AI decision-making or a strong preference for keeping human control. Relevance is second, with 54 per cent of non-adopters saying AI isn’t relevant to their business. And 19 per cent of SMEs said they simply don’t know how to use AI in their business.
Those barriers matter more for agents than for any other kind of AI, because an agent is the version that takes the human out of individual steps. If trust is the brake on using AI to write a newsletter, it’s a much bigger brake on letting AI confirm a booking.
The upside is real enough to keep the question alive. MYOB reported in April 2026 that SMEs using AI were growing 2.8 times faster than those not using it, based on anonymised transaction, bank-feed and payroll data from hundreds of thousands of SMEs [4]. The same release, reported by IT Brief, said 46 per cent of more than 1,000 SMEs in its Business Monitor survey were not using AI and did not intend to in the next 12 months. Treat the growth figure as a correlation, not proof that AI caused the growth. Businesses that adopt early may already be the faster-moving ones.
The adoption finding, then, is this. Using AI is now ordinary. Running AI inside a customer-facing process is not, and that’s the step where AI implementation either pays off or gets switched off.
Lead Response, Quoting and Follow-Up Are Where an Agent Earns Its Keep
If you run a service business, three jobs eat time every week and leak money when they slip: replying to a new enquiry, turning a request into a quote, and following up when the customer goes quiet. All three are repetitive. All three have a clear finish line. All three suffer when they wait.
NAB’s survey backs up where owners think the value sits. Asked where AI could help their business most, 35 per cent of SMEs picked automating repetitive tasks, ahead of marketing and sales optimisation at 31 per cent [2]. Another 16 per cent said AI wouldn’t help them at all.
Lead response
Speed to first reply is the clearest case for an agent. The best-known evidence is old and American, so read it as direction rather than a local benchmark. In a 2011 Harvard Business Review study of 1.25 million sales leads across 42 US companies, firms that tried to contact a lead within an hour were nearly seven times as likely to qualify it as firms that waited even an hour longer, and more than 60 times as likely as those that waited 24 hours or more [5]. The same authors audited 2,241 US companies and found 23 per cent never responded to a web lead at all.
An agent doesn’t sleep, doesn’t get stuck on a job site and doesn’t leave a form submission until Monday. It can reply in minutes, ask the two or three questions a tradie or clinic always needs answered, and either book the next step or hand a warm lead to a person.
Quoting
Quoting is where the time goes for many trades and service firms. The useful agent here doesn’t invent prices. It gathers the details, checks them against your published price rules or your quoting software, drafts the quote and queues it for a human to approve before it goes out. That approval step is the difference between a helpful agent and an expensive mistake.
Follow-up
Follow-up is the job everyone means to do and nobody does consistently. An agent can send the reminder on day two, check whether the customer opened the quote, answer a simple question about it, and pass the conversation back when the customer wants to negotiate. It’s dull work, which is exactly why it suits software.
Notice what these three jobs share. Each one has a defined start (an enquiry arrives), a defined finish (a booking, a sent quote, a closed loop) and an obvious point where a human should step in. That shape is what makes an agent safe to run. Open-ended jobs with no finish line are where agents wander.
Most of the Cost of an Agent Sits Around the Model, Not in It
Ask what an AI agent costs per month and you’ll usually get a quote for the software. That’s the smallest part of the bill to understand, and often the smallest part of the effort.
There are four cost lines in a working agent.
Model usage. The language model is billed by subscription seat or by usage. Vendors such as Anthropic publish their plan and API pricing openly, with API usage charged per million tokens of text in and out [6]. For a small business handling a modest number of conversations, this line tends to be predictable once you’ve watched a month of real traffic.
The platform and integrations. The agent has to reach your calendar, CRM, inbox, quoting tool and, if it takes money, your payment gateway. Each connection has to be built, tested and kept working when one of those systems updates.
Guardrails and testing. Rules about what the agent may say, what it must refuse, when it hands over, and how it proves it’s working. Anthropic’s guidance recommends “extensive testing in sandboxed environments, along with the appropriate guardrails” before agents run on their own [1].
Human review time. Someone has to read the logs, check the edge cases and fix the instructions. In the first month that’s real hours.
How heavy are the last two lines? Sinch’s May 2026 research gives a clue at the enterprise end. Across 2,527 senior decision-makers in 10 countries, including Australia, it found 84 per cent of AI engineering teams spend at least half their time on safety infrastructure rather than on improving the customer experience [7]. A small business won’t have an engineering team, but the work doesn’t disappear. It moves to whoever builds and runs the agent.
That’s why two quotes for “an AI agent” can be miles apart. One prices the model and a template. The other prices the integrations, the guardrails and the first months of tuning. Ask every vendor to split their price into those four lines. If they can’t, you don’t yet know what you’re buying.
A practical budgeting test: before you sign, ask what it will cost to run the agent for a month after launch, including the time someone spends checking its work. If nobody has an answer, the cost hasn’t been scoped. It’s been postponed.
The Confident Wrong Answer Is the Failure That Costs You
The worry most owners voice is the missed call or the clunky bot. The bigger risk is quieter: an agent that gives a customer a wrong answer, politely and with total confidence.
A missed enquiry can be recovered with a callback. A wrong answer about a price, a refund, a policy or a booking time has to be honoured, explained or apologised for, and the customer remembers it.
There’s already a legal line in the sand overseas. In Moffatt v. Air Canada, decided by British Columbia’s Civil Resolution Tribunal on 14 February 2024, the airline’s website chatbot told a customer he could claim a bereavement fare after travelling, which the airline’s actual policy didn’t allow [8]. Air Canada argued it couldn’t be held liable for information from the chatbot. The tribunal called that “a remarkable submission” and said it “should be obvious to Air Canada that it is responsible for all the information on its website.” The airline was ordered to pay $812.02 in damages, interest and fees.
The amount was small. The principle wasn’t. Whatever your agent says on your website, in your name, you said.
The enterprise data shows how often this goes wrong in practice. Sinch’s study found 74 per cent of enterprises had rolled back or shut down an AI customer communications agent after deployment, and the rollback rate was 81 per cent among organisations with the most mature governance frameworks [7]. Sinch reads that second number as a sign the underlying problem is infrastructure and operational capacity, not governance paperwork alone. Mature teams may also simply be the ones watching closely enough to catch problems and pull the plug.
Either way, rollback is normal. Plan for it.
Four guardrails prevent most confident wrong answers:
- One source of truth. The agent answers from your approved prices, policies and service list, not from its general knowledge of how businesses like yours usually work.
- A refusal rule. If the answer isn’t in the source, the agent says so and hands over. “I’ll get someone to confirm that” is a good answer.
- No commitments without a human. Prices outside the published list, refunds, discounts and anything that changes a contract go to a person.
- A tested handoff. When the agent passes a conversation on, the person receiving it gets the full history, so the customer never repeats themselves.
None of these is technical wizardry. They’re decisions. And decisions are what most failed deployments skipped.
Write Down Six Things Before an Agent Touches a Customer
The federal government has already done the heavy lifting on what responsible AI use looks like. The Department of Industry, Science and Resources’ Guidance for AI Adoption sets out six essential practices [9]:
- Decide who is accountable.
- Understand impacts and plan accordingly.
- Measure and manage risks.
- Share essential information.
- Test and monitor.
- Maintain human control.
The guidance is written for organisations of every size. For a 10 to 50 person business, it translates into a one-page scope you can write in an afternoon. The checklist at the top of this article is that page, item by item.
The job, in one sentence
Write what the agent is for in a single sentence, then list what it may do and what it may never do. “Replies to new enquiries, asks our three qualifying questions and books a call” is a scope. “Handles customers” is not.
A named owner
The guidance starts with accountability for a reason. Name one person who owns the agent’s behaviour, reviews its logs and has the authority to switch it off. Write down when a human picks up a handoff, including after hours.
Customers know it's AI
The Office of the Australian Information Commissioner’s guidance on commercially available AI products says public-facing AI tools, such as chatbots, should be clearly identified as such to users, and that privacy policies should explain how the business uses AI [10]. Tell people when they’re talking to an agent.
Personal information rules
The same OAIC guidance recommends, as best practice, that organisations don’t enter personal information, and particularly sensitive information, into publicly available generative AI tools [10]. Know which product your agent runs on, where its data goes and who can see it. If the agent touches payments, payments go through your payment processor, never through a chat transcript.
Tested before launch
Run the agent against a set of real past enquiries before it goes live, including the awkward ones: the angry customer, the request you don’t service, the question about price. Check every answer against your source of truth.
Watched after launch
The OAIC warns that due diligence on an AI product “should not amount to a ‘set and forget’ approach” [10]. Log every conversation. Read them often in the first month. Decide in advance what result would make you switch the agent off, and what you’d fall back to if you did.
That’s the whole page. Businesses that write it before buying tend to buy a smaller, better-defined agent, which is usually the right one.
How 21 Webs Runs Agents Inside Its Own Agency
We don’t only sell this. We run our own agency on the same AI stack we recommend to clients.
Our team includes more than 25 AI agents. Each one has a defined job and reports to a human lead, the same way any other role in the business does. Our marketing and SEO operation runs on an agentic workflow built with Claude Code, and we’ve written about what that did to our own economics in the AI-native agency.
For clients, we build agents into real workflows, including agents that connect to CRM and booking systems and agents that process payments on demand where the workflow needs it. The structure is the one this article describes: a defined job, a human owner, and a loop where we build, measure and optimise, with every change checked against what it was meant to do.
If you’re weighing up where AI fits alongside the tools you already pay for, our piece on ChatGPT for business covers what a general assistant does well, and where it stops. And if your business already runs a chatbot, the contact centre data in four in ten Australian contact centres plan to replace or upgrade their chatbots shows why so many are being reconsidered.
The First Agent a 10 to 50 Person Business Should Deploy
Put the evidence together and an order of operations falls out of it.
Start with lead response. It has the clearest finish line, the best evidence for speed, and the lowest downside if the agent hands over early. An agent that replies fast, asks the right questions and books the next step is easy to test against what you do today.
Add follow-up next. Once the agent is reliably reading your enquiries, let it chase the quiet ones on a schedule you set, with a human picking up anyone who wants to talk terms.
Bring in quoting last. Quoting touches prices and commitments, which is where a confident wrong answer costs the most. Keep human approval on every quote until the logs show the agent is right, consistently, on your real jobs.
Leave open-ended jobs alone for now. Complaints, disputes, refunds and anything with legal weight stay with people. Not because AI can’t write a reply, but because the cost of a wrong one is too high for a first deployment.
Measure one number per job. Time to first reply for lead response. Quotes followed up within your target window. Quote turnaround time. If the number doesn’t move after a month, change the setup before you add anything.
For businesses that also want to be the answer when customers ask AI tools for a recommendation, that’s a separate job. Our AI SEO work covers visibility in AI answers, and our wider SEO services cover the search side.
Important FAQs
What is an AI agent?
What is the difference between an AI agent and a chatbot?
How much does an AI agent cost to run?
Do small businesses actually need AI agents?
The Bottom Line
Using AI is ordinary now. Handing it a job is the real decision.
Four in ten Australian small businesses use AI, and the most common uses are writing and analysis. AI agents are the next step: software that acts, not just answers. They suit jobs with a clear start and finish, like lead response, follow-up and quoting, and they fail in predictable ways, mostly by giving a confident wrong answer.
Write the one-page scope first. Name an owner, tell customers it’s AI, test it on real enquiries, and watch it closely after launch. Do that and an agent becomes a member of the team rather than a liability with a chat window.
Which job in your business should an AI agent do first?
Tell us where your enquiries, quotes and follow-ups get stuck, and we’ll tell you honestly whether an agent fits. Contact us or read more about 21 Webs.
Sources
[1] Anthropic, “Building effective agents”, Anthropic Engineering, 19 December 2024. https://www.anthropic.com/engineering/building-effective-agents
[2] NAB Behavioural and Industry Economics, “NAB SME Business Insights. Embracing AI: Adoption and Key Opportunities Identified by SMEs”, April 2026. About 670 non-farm SMEs surveyed 10 November to 9 December 2025. https://news.nab.com.au/content/dam/nab-news/documents/sme-ai-adoption-april-2026.pdf
[3] National AI Centre, “AI adoption insights: December 2025 to February 2026”, first published 7 May 2026, updated 3 June 2026. SME AI Pulse by Fifth Quadrant, minimum 400 SMEs per monthly wave. https://www.ai.gov.au/news-and-insights/blog/ai-adoption-insights-december-2025-february-2026
[4] MYOB, AI and SME growth analysis and Business Monitor survey, as reported by IT Brief Australia, “MYOB says AI-using SMEs are growing 2.8 times faster”, 22 April 2026 (MYOB’s own release could not be located; trade press report used). https://itbrief.com.au/story/myob-says-ai-using-smes-are-growing-2-8-times-faster
[5] James B. Oldroyd, Kristina McElheran and David Elkington, “The Short Life of Online Sales Leads”, Harvard Business Review, March 2011. https://hbr.org/2011/03/the-short-life-of-online-sales-leads
[6] Anthropic, Claude plans and pricing, checked 3 October 2026. https://claude.com/pricing
[7] Sinch, “The AI Production Paradox”, 13 May 2026. 2,527 senior decision-makers across 10 countries including Australia. https://sinch.com/news/sinch-releases-ai-production-paradox/
[8] Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal of British Columbia, 14 February 2024. https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
[9] Department of Industry, Science and Resources, “Guidance for AI Adoption”, National AI Centre (current version published 5 May 2026). https://www.industry.gov.au/publications/guidance-for-ai-adoption
[10] Office of the Australian Information Commissioner, “Guidance on privacy and the use of commercially available AI products”, 21 October 2024. https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products
