What you need to know
- AI for small business fails when it starts broad. This playbook starts with one repetitive job and one AI agent.
- Week 1 is selection, Week 2 is the build, Week 3 is measurement in real use, Week 4 is the decision to scale or stop.
- Gibson builds the AI prototype on Google Cloud Vertex AI and the Claude API in 48 hours, so most of the 30 days is proving, not waiting.
- The first build is from $1,000 prototype. You extend only what earns its place, which keeps AI automation honest.
- By day 30 you have a working AI agent, a measured result, and a clear decision on the next build.
Most Australian businesses that say “we should use AI” never ship anything, because they start by trying to boil the ocean. They want one assistant that does everything, get lost in options, and quietly give up. The businesses that actually install AI do the opposite. They pick one job, build one AI agent, and prove it before they expand. This is the 30-day version of that discipline.
Week 1: Pick the one job
Do not start with the technology. Start with your week. List the tasks that are repetitive, rule-based, and low-judgement: the ones that eat hours and bore your team. Quote follow-ups. Sorting the inbox. Writing call summaries. Qualifying inbound leads against your criteria. Chasing dormant customers.
Pick one. The test is simple: it should happen often, follow a pattern, and not require a human's judgement on every instance. That is your first AI agent. Write down, in plain English, exactly what a good outcome looks like, because that becomes the spec for the build.
Week 2: Build the AI prototype
This is where most plans stall, and where Gibson's Sandbox is built to move fast. The AI prototype is built on Google Cloud Vertex AI and the Claude API: the Claude model family does the reasoning, Vertex AI provides the enterprise controls, and the agent is wired into the systems you already use. Gibson ships a working build in 48 hours, so this week is mostly about connecting it to your real data and your voice, not waiting on development.
Keep the scope of the build exactly as narrow as the job you picked in Week 1. Resist the urge to add a second task. A focused AI agent that does one thing reliably is the whole point. This is AI automation, not an AI science fair.
The 48-hour build is not the hard part. The discipline of keeping it to one job until it is proven is the hard part. That is what separates AI that ships from AI that gets talked about.
Week 3: Measure it in real use
Now run the AI agent on real work, alongside your team, and measure. Three numbers matter: time saved, error rate, and whether the output is good enough to use without heavy editing. Have a human review every output for the first week, not because the agent is untrustworthy, but because you are calibrating it to your standard.
This is also where AI workflow automation earns or loses its place. If the agent saves real hours and the output is reliable, you keep it. If it needs more editing than doing the task yourself, you either tighten the spec or kill it. Both are good outcomes, because you learned cheaply.
Week 4: Decide, then scale
By day 30 you have a working AI agent and a measured result. Now you decide. If it worked, move it from supervised to trusted, and pick the next adjacent job: maybe the AI agent that drafts follow-ups now also books the reply into the calendar, or an AI phone agent takes the after-hours version of the same enquiry. If it did not work, you spent from $1,000 to find out, which is the cheapest tuition in business.
The compounding comes from repetition. One proven agent every 30 days is twelve a year, each one narrow, measured, and earning its keep. That is how AI for business actually accumulates, not in one big bang.
What to build, in order
A sensible sequence for most Australian SMBs:
- Month 1: Lead qualification or quote follow-up, the task closest to revenue.
- Month 2: Call and email summaries, so nothing falls through the cracks.
- Month 3: An AI phone agent for after-hours enquiries, so missed calls become booked callbacks.
- Month 4 and on: AI marketing automation, reporting, and the long tail of repetitive admin.
How to start this week
If you want help picking the first job and shipping the AI prototype, that is exactly what Gibson's Sandbox does: from $1,000 48-hour build on Vertex AI and the Claude API. Read the Claude API guide for Australian small business for the technical plain-English version, then brief the Gibson team with the one job you want off your plate.
Frequently asked questions
How long does it take to install AI in a small business?
A focused first build can be live in days, not months. Gibson ships a working AI prototype in 48 hours. The 30-day window in this playbook is about choosing the right job, proving the AI agent in real use, and deciding what to scale, not about how long the build takes.
What is the first AI agent a business should build?
Pick the most repetitive, lowest-judgement task that wastes your team's time: quote follow-ups, email triage, call summaries, or lead qualification. A narrow AI agent that does one job reliably is worth more than a broad assistant that does everything vaguely.
What does it cost to install Claude in a business?
Gibson productises the first build as from $1,000 48-hour AI prototype on Vertex AI and the Claude API. Ongoing Claude API usage for a focused agent is usually a few dollars to tens of dollars a month. You extend only what earns its place.
Do I need to replace my existing software to use AI?
No. Good AI automation sits on top of the tools you already run (your CRM, phone system, inbox) through the Claude API. The AI agent reads and writes to those systems; it does not replace them.
