Efficiency with AI is not about buying the smartest model. It is about pointing ordinary AI tools at the boring, repetitive work your team already does, measuring what changes, and stopping the things that do not pay. This guide is for owners and managers of small and mid-sized businesses who want real hours back, not a science project.
I run AI workshops and build AI systems for businesses in Wollongong and across the Illawarra. The businesses getting value are not the ones with the biggest budgets. They are the ones that picked a narrow problem, wrote some rules, trained their people and measured the result.
Where AI actually saves time
AI is very good at work that involves reading, writing, summarising, sorting and drafting text. It is decent at analysing structured data if you give it clean data and a clear question. It is poor at anything that requires knowing facts about your business that it has not been given.
These are the areas where I consistently see time come back in small and mid-sized businesses:
| Task | What AI does | What a human still does |
|---|---|---|
| Email replies and follow-ups | Drafts replies from notes or the thread | Checks tone and facts, hits send |
| Meeting notes and actions | Turns a transcript into a summary and task list | Confirms decisions and owners |
| Proposals and quotes | First draft of scope and wording from a brief | Pricing, commitments, final review |
| Marketing content | Outlines, first drafts, variations, repurposing | Opinion, examples, edits, approval |
| Policies and procedures | Turns rough notes into a structured SOP | Checks it matches reality |
| Inbox and enquiry triage | Categorises, extracts details, suggests routing | Handles anything sensitive or unusual |
| Spreadsheet and report analysis | Spots trends, writes formulas, explains numbers | Decides what to do about it |
| Internal knowledge questions | Answers from your documents, with sources | Keeps the documents current |
The common thread: AI does the first 70 percent of a text-heavy task, and a person does the last 30 percent where judgement matters. That split is where the efficiency is.
Example: one admin role
Take an office manager who spends, in a typical week, 5 hours on email, 3 hours writing up meeting notes and actions, and 2 hours turning notes into procedures and client documents. That is 10 hours.
If AI drafting cuts each of those by a third (a reasonable, not heroic, assumption once they are trained), that is about 3.3 hours a week, or roughly 170 hours a year. At a loaded rate of $60 an hour, that is around $10,000 of time from one person, for a tool that costs a few hundred dollars a year per seat.
Illustrative example. Your savings depend on the work, the person and the training.
Where AI wastes money
The flip side matters just as much. These are the patterns where I see businesses burn money:
- Licences for everyone, training for no one. Fifty seats, five regular users. Without training, most people try it twice, get a mediocre answer and give up.
- Building a custom chatbot before the basics. A bespoke AI assistant on top of a messy shared drive will confidently give wrong answers.
- AI on dirty data. Asking AI to analyse a CRM full of duplicates and blank fields gets you a well-written analysis of garbage. Fix the data first (see the CRM data clean-up guide).
- Automating judgement calls. Letting AI make pricing, credit, hiring or legal decisions without review. That is risk, not efficiency.
- Tool shopping. Trialling a new AI app every week. Most of them are a thin layer over the same few models you can already access.
- No measurement. If you cannot say how long the task took before, you cannot say AI helped.
The fastest way to waste money on AI is to treat it as a technology project. It is a people and process project with some technology in it.
A 30-day AI pilot plan
Do not roll AI out to the whole business at once. Run a 30-day pilot with a small group, on two or three specific tasks, and measure it. Here is the plan I use.
Days 1 to 5: pick and baseline
- Choose 4 to 8 people who do lots of text-heavy work and are open to trying something new. Include at least one sceptic.
- Choose 2 or 3 tasks they do every week. Good candidates: email drafting, meeting summaries, first drafts of proposals or reports.
- Record a baseline: how many times a week each task happens and how long it takes. A simple tally sheet is fine.
- Pick one business-grade AI tool for the pilot. One, not three.
- Write a one-page interim AI policy (more on that below) and get everyone to read it.
Days 6 to 10: train
- Run a hands-on session, two to three hours, using their real tasks, not generic demos.
- Build a shared prompt library for the pilot tasks (examples below).
- Set up any shared projects or workspaces with your style guide, templates and reference documents.
Days 11 to 25: use it and log it
- Pilot group uses AI for the chosen tasks every time they come up.
- Each person logs time taken and a quick quality score (1 to 5) for each task.
- Hold a 15-minute check-in each week: what worked, what did not, which prompts need fixing.
- Keep a list of other tasks people discover along the way. That is your next pilot.
Days 26 to 30: decide
- Compare time per task against the baseline.
- Check quality scores and any errors or near misses.
- Decide: roll out wider, adjust and run again, or stop.
- Write down what you learned and update the policy.
Worked example of the decision maths
Say six people each write about 15 client emails a week that take 12 minutes each. That is 90 emails and 18 hours a week. During the pilot, average time drops to 7 minutes per email with no drop in quality scores.
- Before: 90 x 12 minutes = 18 hours per week.
- After: 90 x 7 minutes = 10.5 hours per week.
- Saving: 7.5 hours per week, roughly 390 hours a year.
- At $70 an hour loaded, that is about $27,300 of capacity a year, against tool costs in the low thousands for six seats.
Illustrative numbers. The point is the method: baseline, pilot, compare.
If the numbers are that clear, roll it out. If the saving is marginal, look at whether the problem is the tool, the prompts or the task choice before you give up.
Prompts and policies
A prompt structure that works
Most bad AI output comes from vague prompts. Teach your team this simple structure:
- Role: who the AI should act as. "You are an account manager at a commercial cleaning business."
- Task: exactly what you want. "Draft a reply to this client complaint."
- Context: the facts it needs. Paste the email, the job notes, the relevant policy.
- Constraints: length, tone, what not to do. "Under 150 words. Apologise once. Do not offer a refund."
- Format: how you want it back. "Plain text email with a subject line."
Three prompts worth copying
Meeting actions: "Here is a meeting transcript. List every decision made, every action with an owner and due date if stated, and any open questions. Use three headings. If an owner or date is not stated, write 'not stated' rather than guessing."
Proposal first draft: "Using the brief below and our proposal template, draft the scope, deliverables and assumptions sections. Use Australian English. Do not include prices. Flag anything in the brief that is unclear as a question for me."
Process to SOP: "Turn these rough notes into a standard operating procedure with: purpose, trigger, numbered steps, owner for each step, tools used and common exceptions. Keep it to one page. Ask me questions about any step that is ambiguous."
Note the pattern: tell it what to do when information is missing. "Write 'not stated'" and "ask me" cut down made-up details dramatically.
Your one-page AI policy
Every business using AI needs a short policy. Not a legal tome. One page people actually read. It should cover:
- Approved tools. Which AI tools staff can use for work, on which plan. Personal free accounts for client work should be a no.
- Data rules. What can never go into an AI tool: for example health information, tax file numbers, bank details, passwords, and anything a client has asked you to keep confidential.
- Review rules. AI output that goes to a client, regulator or the public must be checked by a person who understands the subject.
- Disclosure. When you tell clients AI was used, if ever, and how.
- Decisions. AI can recommend, people decide, especially on anything affecting someone's money, job, health or legal position.
- Reporting. Who to tell if something goes wrong, such as sensitive data pasted in by mistake.
- Owner and review date. Someone owns the policy and reviews it every six months.
Claude vs ChatGPT vs Codex, in plain words
These are the three names I get asked about most. They are all good. The differences matter less than how well you use them, but here is the plain version.
| Tool | Made by | What it is | Where I tend to use it |
|---|---|---|---|
| Claude | Anthropic | A general AI assistant, plus models you can build on via an API | Long documents, careful writing, analysis, policies, working through large files, and as the brain inside agents |
| ChatGPT | OpenAI | A general AI assistant with a wide set of built-in features | Everyday drafting, brainstorming, image work, voice, custom GPTs for repeat tasks |
| Codex | OpenAI | A coding agent that writes, edits and tests software | Building integrations, scripts and small internal tools, with a developer reviewing the work |
For most office work, Claude and ChatGPT will both do the job. Pick one for the business so you can build shared prompts, projects and habits around it. Some teams end up with both because different people prefer different tools, and that is fine once the policy covers both.
Codex is a different category. It is for building software, not writing emails. It is very useful for the integration and automation work behind the scenes, but it is not something your office team needs.
Whichever you choose, use a business or team plan rather than personal accounts. Business plans give you admin controls, shared workspaces and clearer data terms. Read the current terms on data use and training yourself, because they change.
What AI costs
There are three layers of cost, and most businesses only budget for the first one.
| Cost layer | What it covers | Typical range |
|---|---|---|
| Seats | Business plans for Claude, ChatGPT or similar | Tens of dollars per user per month. Check current pricing. |
| Training and setup | Workshops, prompt libraries, shared projects, policy | Half-day workshop $3,500, full day $6,500 (up to 12 people) |
| Builds | Automations, integrations and agents connected to your systems | Single automation from $2,500, AI builds from $8,500, AI agents from $12,000 |
| Running costs | API usage for automations and agents, plus maintenance | Usually modest for small volumes. Estimate before you build. |
My prices are AUD. Tool subscription prices are set by the vendors and change regularly.
The pattern I recommend: seats and training first, prove the value on everyday tasks, then spend on builds for the processes where the pilot showed real volume. Do not start with a $30,000 custom build when $5,000 of seats and training would show you where the value is.
Privacy and the Australian Privacy Act basics
This is general information, not legal advice. If you handle sensitive information or are unsure where you stand, get advice from a privacy lawyer.
The Privacy Act 1988 and its Australian Privacy Principles (APPs) set rules for how many organisations collect, use, disclose and store personal information. Some small businesses are exempt based on turnover, but there are significant exceptions, for example health service providers and businesses that trade in personal information. Do not assume you are exempt without checking. Privacy law has also been going through reform, so check what has changed recently.
For AI use, the practical questions are:
- What personal information goes in? If staff paste client details into an AI tool, that is a use, and possibly a disclosure, of personal information.
- Where does it go? Many AI services process data overseas. The APPs have rules about cross-border disclosure.
- Is it used for training? Check the provider's terms for your plan. Business plans generally offer stronger commitments than free consumer accounts.
- Is it accurate? AI can generate wrong information about a real person. If that gets saved into your records, you have an accuracy problem.
- Are you telling people? Your privacy policy and collection notices may need updating to reflect how you use AI.
- Is AI making decisions about people? Automated decisions that significantly affect individuals attract more scrutiny. Keep a human in the loop.
The Office of the Australian Information Commissioner (OAIC) publishes guidance on privacy and AI products. It is worth reading, and worth passing to whoever owns privacy in your business.
The simplest safe default: no sensitive personal information goes into AI tools unless you have checked the terms, updated your policy and decided it is necessary. Most efficiency gains do not need it anyway. You can draft a complaint reply without the client's date of birth.
Getting staff to actually use it
Adoption is where most AI projects die. The tool works, the licences are paid for, and three months later only two people use it. Here is what gets people on board:
- Be straight about jobs. People hear "AI" and think "redundancy". Tell them what the time saved is for: less overtime, more client work, growth without burning people out. Then mean it.
- Train on real tasks. Generic demos do not stick. Their own inbox, their own reports, their own templates do.
- Find champions. One or two keen people per team who share wins and help colleagues.
- Share a prompt library. Good prompts that work for your business, kept in one place and improved over time.
- Lead from the top. If the owner or manager uses it openly, others follow.
- Make it safe to say it did not work. Bad outputs are useful feedback, not failure.
- Check in at 30, 60 and 90 days. Usage drops off without follow-up.
FAQ
What is the best first AI project for a small business?
Usually email drafting, meeting notes or proposal first drafts. They happen every week, they are text heavy, a human reviews the output anyway, and the time saved is easy to measure.
Should we use Claude or ChatGPT?
Either will do most office work well. Pick one as your business standard so you can build shared prompts and habits, use a business plan rather than personal accounts, and cover it in your AI policy.
Is it safe to put client information into AI tools?
It depends on the information, the tool, the plan and your obligations under the Privacy Act. As a default, keep sensitive personal information out, use business plans, check the data terms and get legal advice if you handle health or other sensitive data.
How much should we budget for AI in the first year?
For a small team, start with business seats plus training, which is often a few thousand dollars, and prove the value. Spend on custom automations or agents only for processes where the pilot showed enough volume to pay for them.
Will AI replace our staff?
In most small and mid-sized businesses I see, AI takes over parts of jobs, not whole jobs. The win is usually capacity: the same team handling more work, or growth without the next hire.
Next steps
Pick two tasks, take a two-week baseline and run the 30-day pilot. If you want help, I can run the training, write the policy with you, or do a full AI readiness assessment so you know where the real value is before you spend on builds. Get in touch and tell me what you are trying to fix.