Every Australian SME I talk to in 2025 is asking the same question: "Should we have an AI chatbot?" The honest answer is: maybe, but probably not the one you're imagining. The bots that work look very different from the ones founders typically scope.
What AI chatbots are actually good at in 2025
- Answering questions from your existing knowledge base. Help docs, product manuals, policy documents. RAG (retrieval-augmented generation) makes this genuinely good now.
- Triaging and routing. Understanding what a customer is asking, then handing to the right human or workflow.
- Looking up structured data on demand. Order status, account info, appointment availability, anything an API can answer.
- Drafting human-quality responses for human approval. Especially good for support teams handling high volume.
What they're still bad at
- Open-ended sales conversations. Real selling is hard. AI bots that try to sell typically annoy users.
- Anything mission-critical without a human in the loop. Legal, medical, financial, never let the bot be the final answer.
- Brand voice in a way that holds up. Bots can sound like your brand 80% of the time. The 20% off-brand answers undo the goodwill.
- Answering questions outside their training data. Hallucinations are mostly fixable with good retrieval, but never fully gone.
Where Australian SMEs are getting real ROI
The patterns that work:
- Customer service deflection on tier-1 tickets. A bot that answers "where's my order" and "can I change my address" deflects 30–60% of tickets at most online retailers.
- Internal knowledge base for staff. "What's our policy on X?", saves manager time, improves consistency.
- Lead qualification on the website. Replaces or augments contact forms; collects info, qualifies, books a meeting.
- Booking and rescheduling. Especially in services businesses (clinics, trades, beauty) where the front desk is overloaded.
The chatbot worth building is the one that does one specific job extremely well. The one that "talks to your customers" in general almost always disappoints.
What they cost
- Knowledge bot (Q&A on your docs): $3.5k–$8k build, $50–$300/month running.
- Support agent with tool use (lookups, refunds, scheduling): $12k–$30k build, $200–$1,500/month running.
- Custom domain agent (specific workflow, high accuracy needs): $32k+ build, $500–$3,000/month running.
Running cost is split between the LLM API (Claude, GPT-4) and your hosting. For most SMEs, $300–$700/month is realistic.
How to scope one well
- List the top 20 questions customers ask. If 80% of volume is in 20 questions, you don't need a bot. You need a better FAQ. Add the bot only if the volume is much wider.
- Pick one specific job. "Answer order status questions and route everything else to a human." Not "be our AI assistant."
- Define what good looks like. "Resolve 50% of order-status enquiries without escalation, with under 10% wrong answers." Measurable.
- Decide the escalation path. Bots fail. The escalation needs to be clean, fast, and frequently used at first.
- Plan the eval suite. A test set of 50–100 real customer questions, with expected answers, that you re-run on every update.
Hidden costs founders miss
- Content cleanup. Your docs are probably not bot-ready. Plan 1–2 weeks of editing before launch.
- Eval discipline. Without a test suite, you can't tell if your prompt change helped or hurt. Build it from day one.
- Brand voice prompt engineering. Tuning the bot to sound like your brand takes longer than non-engineers expect.
- Monitoring. Every bot installation needs ongoing review. Plan a few hours per week of looking at conversations and tightening prompts.
The "is it worth it" check
Three numbers that decide whether a chatbot pays back:
- Volume of repetitive questions per month (>1,000 = probably yes; <200 = probably no).
- Cost per ticket today (>$5 = probably yes; <$2 = probably no).
- Customer tolerance for bots in your specific industry.
The bottom line
AI chatbots work when they have a specific, measurable job and a clean escalation path. They fail when they try to be your AI assistant. Scope narrow, ship a knowledge bot first, then expand.
If you want three Australian developers to weigh in on whether a chatbot is right for your specific case, FindDevs gets you those quotes. Free, in 24 hours.