Is Your Business Ready for an AI Agent? What to Check First

An Australian business is ready for an AI agent when five things are already true: one repeated task where a person reads something messy and makes a call, the data that task depends on sitting in a system the agent can reach, a named person who will own the agent’s output, a baseline figure for what the task costs today, and a decision about which information the agent may see and where it may be processed. Until all five hold, choosing an agent is premature. The gap is nearly always cheaper to close than the agent is to buy, and closing it first is what makes the agent work.

I get asked how to choose an AI agent far more often than whether the business is ready for one, and the second question decides the first. The Australian Bureau of Statistics found that almost one in eight businesses, 12 per cent, used AI in 2024–25, up from 1 per cent in its earlier survey (ABS). Among small businesses the rate was 19 per cent for those that were innovation-active and 4 per cent for those that did no innovation at all. I read that gap as a readiness gap: the businesses that adopted already had a process worth pointing a tool at, and data clean enough to feed it. The rest were looking at demos for jobs they had not yet written down.

Here is the check I run before I quote anything. Five rows, and every one of them is something you can test this week without buying software.

Check	Ready looks like	Not ready looks like	How to test it in a day
The task	One repeated job where a person reads messy input and decides	”We want AI across the business”	Write the job as a single sentence with a verb, an input and a weekly volume
The data	The records live in Xero, MYOB, HubSpot or a database with an API	The records live in inboxes, PDFs and a shared drive nobody tidies	Ask where the agent would read its first record from, and open that place
The owner	A named person reviews the output and can switch the agent off	”IT will look after it”	Name the person and the hour a week they will spend on it
The baseline	Hours per week and the error rate are written down	”It takes ages”	Time the task for five working days
The boundary	A decision exists on what the agent may see and where it is processed	Nobody has asked	Answer the OAIC’s due-diligence questions in writing

What does “ready for an AI agent” actually mean?

Ready means the agent has a job, a source of truth, a supervisor, a scoreboard and a boundary, and all five exist before a supplier is chosen. An AI agent is software that reads an input, decides what to do and acts through your systems, so every readiness question is about the input, the systems and the supervision. The model it runs on is the part that matters least, because it is the part that is easiest to swap later.

Anthropic’s engineering guidance for people building agents on its models puts the cost side plainly: “The autonomous nature of agents means higher costs, and the potential for compounding errors. We recommend extensive testing in sandboxed environments, along with the appropriate guardrails.” (Anthropic). A sandbox and guardrails are things the buyer provides, not the vendor. The sandbox is a copy of your data the agent can be wrong against without consequence; the guardrails are the rules about what it may touch. If neither exists, the first live run is the test, and the people who find the errors are your clients.

Microsoft’s 2025 Work Trend Index found that 81 per cent of leaders expect agents to be moderately or extensively integrated into their company’s AI strategy within the next 12 to 18 months (Microsoft). That is an intention figure, and I take it as one. The five checks are what turn an intention into a deployment that is still running a year later, and I would not quote a build without them.

Is your data somewhere an agent can reach it?

The data is reachable when the records the task runs on live in a system with an interface a program can read and write: an accounting platform, a CRM, a practice-management system or a database. Xero, MYOB and HubSpot all publish developer platforms for exactly this (Xero, MYOB, HubSpot), and the Model Context Protocol is, in my view, becoming the standard plug between an agent and those systems. Its own documentation says to “Think of MCP like a USB-C port for AI applications.” (Model Context Protocol). Xero now lists an MCP server in its AI toolkit, which tells you where the accounting platforms think this is going.

The unready version is just as recognisable. The records live in an inbox, in PDFs attached to emails, in a spreadsheet with three people’s versions of it and in someone’s head. An agent can read unstructured material, and reading messy input is the one thing it does that a plain rule cannot, but it still needs one place to write the answer back to and one record it can trust when two disagree. IBM’s guidance on data quality puts it in the oldest words the field has, that “garbage in, garbage out” holds for machine learning as it did for everything before it (IBM).

So, my test: ask where the agent would read its first record from, then open that place and look at it. If it is a system with an API, the data check passes even if the data is imperfect. If it is a folder, the first project is a data project, and I will say so rather than quote an agent.

Who will own the agent once it is running?

Someone on your side must own the agent’s output: a named person who reads what it produced, catches what it got wrong and has the authority to switch it off. Without that person the agent is nobody’s job, and a pilot that is nobody’s job stops the first time it breaks. MIT’s 2025 study of enterprise generative AI, reported by Fortune, found that purchasing from specialised vendors and building partnerships succeeded about 67 per cent of the time while internal builds succeeded only one-third as often (Fortune), and the same report located the largest returns in back-office automation. My reading of the internal-build figure is that those systems did not fail for lack of talent. They failed because, once the first version shipped, nobody owned them.

The research on agent reliability makes the owner’s role concrete. In TheAgentCompany benchmark, which simulates a small software firm with real tasks across its tools, the most competitive agent completed 30 per cent of tasks autonomously (arXiv). The τ-bench authors, testing agents on realistic customer-service tasks with rules to follow, concluded that their “findings point to the need for methods that can improve the ability of agents to act consistently and follow rules reliably” (arXiv). Agents are improving quickly and these figures will date. The point survives: an agent that completes most of a task and is inconsistent on the rest needs a human who checks the rest, and the time that person spends is part of the cost.

The owner needs an hour a week, not a job title. In a practice of five to fifty people it is usually the person who does the task today, because they already know what a wrong answer looks like. Name them before the build, write their hour into the plan, and give them the switch.

Do you know what the task costs you today?

A baseline is the written figure for what the task costs now: the hours per week, the error rate and the delay it causes, measured over at least five working days before any agent is involved. It is the only way to know afterwards whether the agent paid for itself, and it is the check most buyers skip because the answer feels obvious. “It takes ages” is a feeling. Eleven hours a week at a bookkeeper’s rate is a baseline.

Take debtor follow-up, a job agents are often sold for. Xero’s Small Business Insights put the average time an Australian small business waited to be paid after issuing an invoice at 22.9 days in the June quarter of 2026 (Xero). If your figure is 35 days, an agent that chases politely and consistently has a measurable target. If your figure is 20 days, the problem you think you have may not be the one you have, and the agent is solving something else. The baseline decides which.

The baseline also sets the price you should pay. We have written separately about what a first agent engagement costs and what to ask a supplier; the short version is that a build should be sized against the hours it removes, and you cannot size against a number you have not measured. Where the task turns out to be a fixed rule, a workflow in n8n, Zapier or Microsoft Copilot Studio will cover it for a subscription, and the three ways to automate a process sets out when each one wins. Measuring first is what tells you which tier you are shopping in.

Where may the agent’s data go, and who has decided?

The boundary is a written decision on what information the agent may read, where it is processed and who can see what it produces, made by the business before the vendor is chosen. For an Australian business the regulator has already set out the questions. The OAIC says organisations adopting a commercially available AI product should conduct due diligence covering whether it has been tested for the intended use, how human oversight is embedded, the privacy and security risks, and who will have access to the personal information the product takes in or generates (OAIC). Then it adds the sentence I agree with most:

“Due diligence for AI products should not amount to a ‘set and forget’ approach. Regular reviews of the performance of the AI product itself, training of staff and monitoring should be conducted throughout the entire AI product lifecycle to ensure a product remains fit for purpose”

Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products (OAIC)

Two facts frame the decision. The Privacy Act binds organisations with an annual turnover above $3 million, and it covers private-sector health service providers at any size (OAIC), so a two-doctor clinic is inside it while a trades business under the turnover threshold is not. And the ABS found that just over one in five Australian businesses, 21 per cent, experienced a cyber security incident in 2024–25 (ABS). An agent with read access to your customer records is another door, and the boundary decision is where you decide how wide it opens.

The decision also has to account for what staff are already doing. Roy Morgan’s research for the March quarter of 2026 found 13.6 million Australians, 58 per cent of people aged 14 and over, used AI tools in an average four weeks (Roy Morgan); we track that indicator in our index of Australian small-business AI data. Your team is already pasting things into something. A sanctioned agent with a defined boundary is usually safer than the unsanctioned habit it replaces, and the four levels of AI data security sets out the options from a public chatbot to a model that never leaves the building. Where the material genuinely cannot leave the premises, what it takes to run a private AI agent covers that path.

How do you run the readiness check in a week?

The check takes a week of attention and no money: a sentence on Monday, a data lookup on Tuesday, a name on Wednesday, a stopwatch for five working days and a written boundary by Friday. Each step produces a document a supplier can be held to, which is the reason to do it before the first meeting rather than after.

Monday, write the task as one sentence with a verb, an input and a weekly volume. “Read every supplier invoice that arrives by email, match it to a purchase order in Xero and flag the ones that do not match, about 60 a week” passes. “Use AI in accounts” does not. If you cannot write the sentence, you are not choosing an agent yet, and what an AI agent can actually do for a business is the better starting point.

Tuesday, open the place the agent would read its first record from. Wednesday, name the owner and write down their hour. From Thursday, time the task and count the errors for five working days. Friday, answer the OAIC’s four due-diligence questions in writing and decide what the agent may see.

By the end you will know whether you are ready, and if not, which of the five is missing. The one I see missing most often is the baseline, and that is fixed with a stopwatch. This groundwork is where our AI builds start, and if the check shows the job is a rule rather than a judgement, I will say so and point you at the cheaper tool. If you would like a second pair of eyes on the five checks, start a conversation or see how we scope a first engagement.

The agent itself is the last thing you choose. Get the five in place and the choice becomes a small one, because most of what decides whether it works is already done.

Frequently asked questions

How do I know if my business is ready for an AI agent?

A business is ready when five things exist: one repeated task where a person reads messy input and decides, the records for that task in a system with an API, a named person who owns the agent’s output, a measured baseline of what the task costs today, and a written decision on what the agent may see and where it is processed. If any one is missing, fix the gap before choosing an agent; it is cheaper than the agent.

What data does an AI agent need to work?

An agent needs one trusted source of truth to read from and write back to, usually an accounting platform, a CRM or a database with an API. It can read unstructured material such as emails and PDFs, but it still needs a system to record its answer in and a record it can trust when two sources disagree. If the records live only in inboxes and spreadsheets, the first project is a data project rather than an agent.

Do I need an AI agent or just an automation?

If the task can be written as a fixed rule, with the same input always producing the same action, a workflow automation will do it for a subscription and do it more predictably. An agent earns its cost only where a person currently has to read something messy and make a judgement. Measure the task first; the baseline tells you which tier you are shopping in.

Does the Privacy Act apply to my business if I use an AI agent?

The Privacy Act binds organisations with annual turnover above $3 million and covers private-sector health service providers at any size, so a small clinic is inside it. Whatever your size, the OAIC’s guidance says to conduct due diligence on an AI product: whether it was tested for your use, how human oversight is embedded, the privacy and security risks, and who can access the personal information it takes in or generates, with regular reviews rather than a set-and-forget approach.