Home/Resources/Our perspective on practical AI for Australian businesses.
Practitioner perspective Practitioner perspective

Our perspective on practical AI for Australian businesses.

What we believe about AI implementation in Australian businesses — written by Adam Dong, founder of Horizon AI. The principles behind the work, the patterns that consistently break in production, and the operating philosophy that shapes every engagement.

By Adam Dong, Founder of Horizon AI · Last reviewed 2026-05-21

AI is not the work; it's an ingredient in the work. The most useful AI implementations in Australian SMBs in 2026 are unglamorous: a draft email approved by a human, a triaged inbox routed correctly, a fact-find pack assembled before a meeting. Horizon AI exists because most Australian businesses don't need an AI transformation strategy — they need one specific workflow to stop hurting, shipped inside the software the team already uses, with a person still on the approval gate. Everything below explains why we believe that, and what that belief produces in practice.

AI without a workflow is just a chatbot

A general-purpose AI chat window in the corner of someone's screen does not change how a business operates. What changes operations is an AI workflow with a specific trigger, a specific decision boundary, a specific output, and a specific person who reviews it. "Use ChatGPT more" is not an implementation strategy. "When a tenant arrears email comes in, draft the tier-1 response in PropertyMe and queue it for the property manager to review" — that is.

Most failed Australian AI rollouts we've encountered share the same pattern: the business bought enterprise AI access, trained the team on prompts, and waited for productivity gains that never arrived. The team uses the AI for ad-hoc questions, occasionally drafts an email, then drifts back to old habits. Six months later there's nothing to show.

The workflows that compound are the ones with sharp edges: defined trigger, defined output, defined reviewer, defined exit. They're harder to scope and they look smaller. They also actually ship.

AI is the 70%, not the 100%

On any business workflow worth automating, AI handles roughly 70% of the grunt — classification, drafting, summarisation, format-conversion, lookup. The remaining 30% — judgement calls, regulated decisions, client-facing finality — stays with humans. The business that tries to push past 70% usually breaks something expensive; the business that takes the 70% wins the time back and moves on.

We hold this number loosely — 70/30 is a useful frame, not a measurement. Some workflows are 90/10 (document field extraction with high confidence thresholds). Some are 30/70 (anything involving final legal advice, final tax positions, or sensitive client comms). The point is that the boundary is real, named in advance, and visible inside the workflow itself.

The 30% the human keeps isn't waste. It's where the firm's actual expertise gets applied. The 70% the AI absorbs is the part where the firm was wasting expertise on grunt work.

Live inside the software, not on top of it

Every business already has a CRM, an inbox, a document store, a finance system and a sector platform. AI workflows that live inside those systems get adopted. AI workflows that require the team to log in somewhere new get ignored within 30 days. There is no exception to this in any Australian SMB we've worked with. The platform you've spent the last 5 years getting your team fluent in is the platform the AI needs to inhabit.

This is why Horizon AI builds inside Xero, MYOB, LEAP, Smokeball, PropertyMe, Reapit, HotDoc, Iress Xplan, HubSpot, Microsoft 365 and Google Workspace, rather than asking firms to adopt a new AI platform. The model bill is the same either way; the adoption rate is not.

The corollary: if a workflow genuinely can't live inside the existing stack — because the stack doesn't have the right surface, or the integration doesn't exist — the right answer is usually to wait, not to build a parallel destination the team will resent.

Australia matters more than people think

Australian AI implementation is not just US AI implementation with a different time zone. The software stack is different (Xero dominates accounting, LEAP dominates legal, PropertyMe dominates rentals, Iress dominates financial advice). The compliance load is different (Privacy Act, AHPRA, ASIC, AFSL, Australian Consumer Law, state-specific real estate licensing). The language is different (BAS, EOFY, body corporate, conveyancing, trust accounts). An implementer who treats AU as a generic English-speaking market gets it wrong in expensive ways.

The Australian-software-fluency gap shows up most painfully on integration. A US-trained AI implementer who hasn't worked inside LEAP doesn't know which fields trigger trust accounting consequences. A generic offshore dev team doesn't know that PropertyMe's owner portal is the destination most rental workflows need to surface into. The build still ships; it just doesn't do the right thing.

We're not arguing local is always better — for some technical workflows, location is irrelevant. We're arguing that the closer the implementer sits to the specific software, compliance and language conventions of the business, the fewer surprise rebuilds happen at month three.

Boringly competent beats novel

The AI implementation partners that compound value year over year are the ones that ship reliable, well-documented, audit-trailed workflows that the team can operate without re-explaining. Novel approaches and bleeding-edge model demos are seductive and unreliable. We aim to be the most boringly competent AI implementation partner an Australian business will work with. The bar is unglamorous and high.

Boring competence means: every workflow has a runbook. Every AI call has an audit log. Every approval gate is a human, not an opaque rule. Every change has a clear diff. Every handover includes training. Every retainer includes monthly maintenance, because vendor APIs change and edge cases surface.

It also means saying no to AI workflows we don't think will hold up. Frontier model demos are exciting; six-month-old workflows that are still working without intervention are what actually moves the business. We choose the latter every time, and we'd rather lose an engagement than ship something that breaks in a way that costs the client more than it saved.

Strategy decks aren't AI

An AI strategy document that doesn't end in a working workflow is theatre. We've reviewed enough $40,000–$120,000 'AI roadmap' engagements for Australian businesses that produced a deck, three workshops, and exactly zero production AI to be honest about it: most of that money would have been better spent on one scoped build. We sell strategy only when the business genuinely doesn't yet know what to build first — and we sell build the rest of the time.

There is a real role for AI consulting — workflow audits, opportunity prioritisation, integration assessments, compliance-aware framing. These deliverables matter when the business has real ambiguity. They become busywork when the answer was already obvious and what the business actually needed was a builder.

The litmus test we apply before quoting strategy work: "if the recommendation is going to be 'build the obvious thing first', would you want to pay us $20K to confirm that, or $20K to build the obvious thing?" Most clients pick build. We respect that signal.

Maintenance is the work, not the warranty

A shipped AI workflow that nobody maintains is a six-to-twelve-month problem waiting to happen. Vendor APIs change, edge cases surface, business rules evolve, regulatory positions update. Implementation partners that don't include or recommend ongoing maintenance are leaving a time bomb behind. We charge for maintenance because we believe in maintaining things; if a client genuinely doesn't want maintenance, we'd rather not ship the workflow at all.

The maintenance cadence we recommend for most Australian SMB workflows: a monthly check-in (30 minutes plus async), a quarterly substantive review (1-2 hours), and unbounded reactive coverage for vendor-driven breakage. For sector-regulated workflows (legal, financial advice, medical) we recommend a slightly higher cadence and explicit annual compliance review.

Maintenance is not failure recovery; it's how the workflow keeps earning out. We'd rather an AI workflow get gently better over 24 months than ship hot and decay quietly.

What we will not do

There are workflows Horizon AI declines. Anything that bypasses a platform's terms of service. Anything that presents AI as supervised professional advice (final legal, final tax, final medical). Anything that fabricates client testimonials or experience. Anything that knowingly violates Australian Privacy Act obligations. We'd rather lose an engagement than ship a workflow that hurts the client downstream. This is not moral signalling — it is how the work actually has to be done if it's going to keep being useful in 18 months.

We're also cautious about workflows that scale faster than oversight can scale. An AI-drafted email batch that goes out once a week to 50 clients with practitioner review is one thing; the same workflow firing to 5,000 contacts with no review is another. The Australian businesses we work with are usually professional services firms where reputation compounds — and where a single bad AI-generated communication can do meaningful damage. We design for the cautious end of that spectrum.

Where this leads

The Horizon AI bet is that the AI implementation work that holds up over time will look like operations consulting more than like technology installation. It will be deeply software-fluent, compliance-aware, judgement-led, and unspectacular. The firms that win on AI in Australia over the next 24-36 months won't be the ones that adopted the most AI; they'll be the ones that quietly removed the most friction with AI and kept their professional judgement intact.

If that's the work you want done in your business, the next step is a free 30-minute strategy session. Bring one workflow you want to improve. We'll map the cleanest implementation path before the call ends. If we don't think we can help, we'll say so.

Common questions

Answered, before you buy.

Is Horizon AI hostile to AI strategy work?
No — we do strategy and audit work, and we recommend it when the business has real ambiguity about what to automate first or where the compliance constraints are. We push back on strategy engagements that exist primarily to justify their own fees, not on strategy engagements that produce a real decision.
Do you take a position on which AI model is best?
We use Claude, ChatGPT and Gemini in production and pick per workflow. Claude tends to be stronger on long-context document and compliance work; ChatGPT tends to be stronger on broad drafting and code; Gemini tends to be stronger on multimodal and large-context retrieval. Most Horizon builds route across more than one model rather than committing to one.
Where do you draw the line on AI doing client-facing communication?
Drafting is fine; sending without human review is not, for any workflow that touches a client by name, makes a commitment, or interprets professional advice. We bias toward review-on-every-touch for regulated sectors (legal, financial advice, medical) and review-on-exception for volume operations (ecommerce support, basic admin).
What's the smallest engagement Horizon AI takes on?
A scoped Flex build of one workflow, typically $5,000-$15,000 depending on integration complexity. We don't usually take engagements smaller than that because the discovery, scope and handover overhead doesn't fit cleanly in a smaller package. For genuinely tiny tasks, the right answer is often a no-code workflow built in-house rather than an external engagement.

Want to talk through this for your business?

Book a free 30-minute strategy session with Adam. Bring one workflow you want to improve and we will map a concrete first build before the call ends.