A decision framework for small business owners weighing a consultant against building it themselves with Claude Code, Codex, or whatever they prefer. The real costs on both sides, including the ones that never appear on an invoice, and a scorecard that tells you which path the work in front of you actually calls for.
Published: 15 July 2026. Last reviewed and sources re-checked: 23 August 2026.
01 — The Short Answer
Do it yourself when the work touches one tool, a human checks the output before it leaves, and nothing breaks expensively if it is wrong. Bring someone in when it touches several systems, involves customer or regulated data, has a deadline attached, or has already stalled once. Most small businesses need both, in that order, and the mistake that costs the most is choosing a side before looking at the specific piece of work.
That is a less satisfying answer than the two you will usually be given. One says AI is so accessible now that paying anyone is a waste. The other says it is so complex that trying alone is reckless. Both are selling something. The first is selling a tool subscription, the second an engagement.
The useful question is not which approach is better. It is what does this particular piece of work cost me on each path, counting the hours that never get invoiced. This page walks that calculation, gives you a scorecard to run it yourself, and is honest about the large amount of AI work that you should not be paying us or anyone else to do.
It will not conclude that you should hire a consultant. Roughly half of it argues the other way, because for a great deal of what small businesses actually need, an engagement is the overspend and a forty dollar subscription is the whole answer.
We would rather be the firm you call when the work genuinely warrants it than the one that sold you something it did not.
02 — Start Here
Almost every version of this decision starts with manufactured urgency, and it is worth defusing that before spending anything. The numbers used to create the panic mostly come from companies selling AI. The official Australian figures tell a much calmer story.
Reported AI use in 2024-25, across all Australian businesses. Australian Bureau of Statistics — Business adoption of Artificial Intelligence accelerates in 2024-25, released 25 June 2026
Vendor and industry surveys over the same period have put Australian small business AI adoption anywhere from 29 to 64 per cent. The national statistical agency, asking a much larger and more representative sample, found around 11%. That is a gap of three to six times, and it exists because a survey run by a company selling AI reaches people already interested in AI.
This matters for your decision in a specific way: if you believe you are two years behind, you will rush, and rushing is what turns a sensible ten thousand dollar engagement into a poorly scoped one. You are not late. You have time to do this in the right order.
Where the urgency is genuine is what happens after adoption. Deloitte Access Economics found only 5% of small and medium businesses are fully enabled to realise the benefits, with one third not using AI at all, and more than half of small business workforces having only basic or novice familiarity against 10% with advanced skills.
In other words: plenty of businesses have a ChatGPT subscription and no idea what to do with it. The competitive gap is not between businesses that use AI and businesses that do not. It is between businesses that got past the first fortnight and businesses that did not.
The ABS found that small businesses actively doing innovation work reported AI use of 19%, almost five times the rate of those doing none.
AI adoption is not a separate initiative that arrives on its own. It follows the habit of deliberately changing how the business works. If that habit does not exist yet, that is the thing to fix, and no consultant can install it for you.
Deloitte Access Economics, published by CEDA — Smaller Australian businesses are missing out on AI, 26 November 2025
03 — The Framework
Run the specific piece of work through these four, not your business as a whole. The same business will land on different answers for different jobs, and that is the correct outcome rather than an inconsistency.
How many things have to talk to each other?
One tool used by one person is a different problem to four systems that have to stay in sync. Complexity is not about how clever the AI is. It is about how many places the work can break, and how far the breakage travels before anyone notices.
One tool, one person, output reviewed by a human before it leaves the building.
Three or more systems, shared data, or anything that writes back into a system of record.
What are the hours you would spend actually worth?
This is the cost that never appears on an invoice, which is exactly why it gets left out of the comparison. The honest version is not "my time is free". It is the work you would otherwise have done in those hours, and whether you were realistically going to do this at night after everything else was finished.
You have genuine slack, or the learning itself is worth more to you than the hours it costs.
The build hours are hours you currently sell, bill, or spend on the thing only you can do.
How long until it pays back, and can you wait that long?
A DIY build is usually cheaper and slower. A paid build is usually faster and dearer. That trade is only worth making when the delay itself has a price attached to it. Much of the time it does not, and the honest answer is to take the slower road and keep the money.
Nothing is on fire. No deadline, tender, or contract depends on this landing by a particular date.
The delay has a number on it, and that number is larger than the fee.
What breaks if it is wrong, and who wears it?
The question is not whether the AI will make mistakes. It will. The question is what sits between the mistake and your customer, and whether you would find out about it from a dashboard or from a complaint.
A person checks the output before anyone outside the business sees it.
Customer data, health or financial records, anything covered by the Privacy Act, or anything that reaches a customer unreviewed.
Across a survey of 1,009 people at US businesses with 2 to 250 employees, average time saved through AI came to 5.6 hours a week, with managers reporting 7.2 hours against 3.4 hours for individual contributors. Business.com 2026 Small Business AI Outlook — Updated 20 January 2026
Read that as the size of the prize, and then put your own hourly figure against the build time. If a workflow returns five hours a week, it is worth real money every week it does not exist. Sometimes that argues for paying to have it sooner. Just as often it shows the return is large enough that a slower free build still wins comfortably. Do the arithmetic rather than the vibe.
04 — Score It
Have one specific piece of work in mind, not your business in general. Nothing is stored, nothing is sent anywhere, and no email address is required.
Be honest rather than optimistic. The scorecard is only useful if the input is.
05 — The Honest Part
This is the part most consulting pages leave out. There is a large amount of valuable AI work that you should not pay anybody to do, and knowing where that line sits is most of what separates a good AI budget from a wasted one.
Drafting, summarising, first-pass research, cleaning up notes, rewriting a quote, preparing a proposal. One person, one tool, one output that a human reads before it goes anywhere. Nobody should be selling you an engagement for this, and plenty will.
When a human is the last step, the failure mode is a wasted ten minutes rather than a wrong invoice. That review step is worth more than most of the engineering people pay to avoid needing it.
An owner who spends a month using these tools badly develops a feel for where they break that nobody can sell them. Every later decision gets better because of it, including the decision about who to hire and what to ask them. Outsourcing the first month is the most expensive saving on the table.
Twenty to forty dollars a month per seat covers considerably more than it did a year ago. For a large share of what small businesses actually need, that is the entire answer, and an engagement would be the overspend.
Claude Code, Codex and their equivalents have moved a real category of work from impossible to plausible for a non-technical owner. An internal dashboard, a script that reconciles two spreadsheets, a tool that drafts responses from your own templates: these are now weekend projects for someone patient, where five years ago they were quotes.
Anyone telling you otherwise has an interest in you not finding out. Take the weekend. The worst realistic outcome is that you learn precisely what you would have been paying for, which makes you a much harder person to sell to afterwards.
Can a person look at the output and tell whether it is right, before it reaches anyone outside the business?
If yes, build it yourself. That review step absorbs most of what would otherwise need engineering. If no, because it runs unattended or reaches a customer directly, you have left DIY territory regardless of how straightforward the build looks.
06 — The Other Half
Not because the build does not work. Because of what happens in month three, month six, and the week you are on leave. Every one of these is a failure of ownership rather than ability.
Building it is the short part. It works in week one, drifts in month three when a tool changes something, and by month six nobody remembers how it was wired together. A build with no owner is a liability that used to be an asset.
AI coding tools let people build things they could not previously build, which also means shipping things they cannot review. If you cannot read the output, you cannot judge the risk, and the tool will not raise its hand.
The first workflow is easy. The wall arrives when it has to write into your accounting system, respect who is allowed to see what, and not fire twice when something is retried. This is where most DIY builds stall, and they usually stall at about eighty per cent.
Usually the owner. If that person is on leave, unwell, or moves on, the process leaves with them. Anything the business genuinely depends on needs a second person who can fix it, which is a handover and documentation problem more than a technical one.
The most common reason a build gets abandoned is not that it failed. It is that nobody agreed what success looked like, so it was never allowed to succeed. Write the number down before you start, whichever path you choose.
The risk you cannot see from inside the tool
Martin Fowler, The VibeSec Reckoning — 27 May 2026, collating AppSec Santa, Aikido Security, Black Duck OSSRA and Sonar research
This is not an argument against building things yourself. It is an argument for knowing which builds need reviewing. A tool that summarises your meeting notes and a tool that writes into your customer database carry the same effort and completely different consequences. The coding assistant will produce both with equal confidence and will not distinguish between them for you.
07 — Side by Side
The middle column is the one most often overlooked and most often correct. You build it, with someone scoped in only for the parts where being wrong is expensive.
| Do it yourself | Hybrid | Bring someone in | |
|---|---|---|---|
| Money out the door | Tool subscriptions only | Subscriptions plus a small scoped engagement | A scoped fee, plus the cost of change that rarely gets quoted |
| Hours out of your week | High, and open-ended | Moderate, front-loaded | Low, but never zero. You still have to answer the questions |
| Time to something usable | Weeks to months | Days to weeks | Days to weeks |
| Who fixes it when it breaks | You | You, with someone to call | Them, until the engagement ends. Ask what happens after |
| Who wears it if it is wrong | You | You | Depends entirely on what the contract says |
| Capability left behind | Highest. You learned it | High, if training was in scope | Lowest, unless you insisted on it |
| Suits you when | Low complexity, low exposure, you have slack | Medium complexity, or you want to learn but not alone | High exposure, a real deadline, or a build already stalled |
Hiring the capability rather than renting it. A mid-level AI engineer in Australia sits at roughly $130,000 to $165,000 base in 2026, and a senior at $165,000 to $200,000. AI Talent on Demand — AI Engineer Salary Australia 2026, 25 March 2026, compiled from Glassdoor, Seek and placement data
For most small businesses that settles the question quickly, and it is worth knowing precisely why: not because the salary is unaffordable in isolation, but because a single hire with no one to learn from is a fragile way to hold capability. Where it does work is when there is already enough AI-adjacent work to keep the role busy and someone experienced to point it in the right direction.
08 — If You Decide to Hire
There is no best AI consultant, in Melbourne or anywhere else. The field moves faster than any ranking survives, and the right fit depends entirely on which of the three jobs you are hiring for: visibility in AI search, workflow automation, or lifting team capability. What can be assessed, before you sign anything, is method.
Ask what the baseline is, how it gets measured, and when it gets re-measured. If the answer is a case study rather than a method, you are buying a story. A method that can be run twice is the only thing that will tell you whether the money worked.
A recommendation to replace your systems before any AI work starts is usually a sign the work has been scoped to what they like building. Most genuinely useful AI work sits on top of what you already run.
The barrier for Australian small businesses is capability, not tooling. An engagement that leaves that unchanged has fixed one workflow and none of the reason it was broken in the first place.
Deloitte Access Economics, published by CEDA — Smaller Australian businesses are missing out on AI, 26 November 2025
Anyone who cannot name three things in your business that AI is a poor fit for has not looked at your business. This is the fastest question available for separating a diagnosis from a pitch.
Accounts in your name, prompts and code in your possession, data in your systems, all of it exportable. Get it in writing before you start. This is the single most common thing owners discover too late.
People, days, deliverables, and what is explicitly out of scope. Almost no Australian firm publishes rates, so a quote is only comparable when it is itemised. Ask two firms for the same breakdown and the difference in what you are actually getting becomes obvious immediately.
The last one is the most revealing and the least expected. A consultant who answers it immediately and specifically has looked at your business. One who deflects has looked at their own service list.
You Better Ask works with Australian and New Zealand small and medium businesses on AI visibility, workflow automation, and team capability. We suit you if you want the method handed over along with the work.
We are a poor fit for enterprise programmes, for one-off tool installations, and for anyone wanting a system they never have to understand. If the scorecard above put you in the first band, we are also a poor fit for this particular job, and that is a reasonable outcome for both of us.
09 — Before You Commit to Either
The scorecard on this page works on one piece of work you already have in mind. Its limitation is that it can only assess what you already thought to consider, and the most valuable AI opportunities in a business are usually the ones nobody has named yet.
That is the gap the free AI Business Audit fills. You give us your business name and website. It produces a written report on where AI genuinely fits in your operations and marketing, what each opportunity is plausibly worth, and how a leading AI model currently describes your brand when a customer asks about your category.
It is deliberately useful whichever way you decide. If the answer is do it yourself, you have a prioritised list to work through and no obligation to us. If the answer is get help, you can hand that report to any consultant, including ones that are not us, and get comparable quotes against the same scope. That is a much stronger negotiating position than describing the problem fresh to three firms and receiving three differently shaped proposals.
No credit card, no sales call attached, and the report is yours to keep.
10 — Common Questions
Not by default. Hire one when the work touches several systems, involves customer or regulated data, has a real deadline attached, or has already stalled once. For a single tool used by one person with a human checking the output, a subscription and a few focused weeks will get you further than an engagement, and you keep the judgement you built along the way.
For plenty of things, yes, and the tools have moved a long way. The limit is not what they can produce, it is whether you can review it. Around a quarter of AI-generated code carries confirmed vulnerabilities, and roughly one in five enterprise breaches now trace back to AI-generated code. If the output touches customer data or runs unattended and you cannot read it, you are carrying a risk the tool will not surface for you.
Almost no Australian firm publishes rates, so any single figure quoted online is an estimate rather than a price. Scope drives cost far more than firm size does. The useful move is to ask two or three firms for the same scope broken down by people, days and deliverables, then compare those. As a reference point on the other path, building the capability in-house means an AI engineer salary of roughly $130,000 to $165,000 for a mid-level hire in Australia in 2026.
Probably less than you have been told. The Australian Bureau of Statistics found around 11 per cent of small businesses reported using AI in 2024-25, against vendor surveys claiming figures three to six times higher. The real gap is not adoption, it is depth. Only about 5 per cent of small and medium businesses are considered fully enabled to realise the benefits.
You build it, with someone scoped in for the parts where being wrong is expensive: the data model, the security review, the integration, and the definition of done. It costs a fraction of a full engagement and it is the right answer more often than either extreme, particularly when a build has already stalled part-way.
Score the work honestly against complexity, opportunity cost, speed to ROI and risk exposure. The scorecard on this page does that in about two minutes. If you want the same assessment against your actual business rather than a generic workflow, the free AI Business Audit produces a written report on where AI fits, what it is worth, and which parts you could reasonably do yourself.
Get an evidence baseline before you commit to either road. The audit is free, takes a few minutes, and is just as useful if you decide to build it yourself.