AI Implementation
We come in, find where the hours actually go, and build agents against the fastest wins.
No strategy deck. No twelve-month transformation. A two-week audit, a ranked shortlist with hours and dollars attached, and then we build the top items and measure what came back.
The problem
Most businesses that 'tried AI' picked the wrong first project.
They put a chatbot on the website, found it embarrassing in front of a customer, and concluded the technology was oversold. That is a fair conclusion from a badly chosen starting point.
What actually works
Pick a task that is high frequency, low judgement and text-shaped. Build against it. Measure the hours before and after. Then do the next one. Every business we have seen get real value followed that sequence, and every one that did not, did not.
Why an outside view helps
You cannot see the task that eats six hours a week, because you have been doing it for nine years and it feels like the job. The value of someone sitting with your team for two weeks is that they have not stopped noticing.
Step one
The AI opportunity audit
Two weeks, fixed fee. We work inside your business — sitting with the people doing the work, not interviewing the executive about their vision.
- Half a day shadowing each functional area — sales, admin, finance, operations
- A task inventory with frequency and honest time-per-instance
- Data and systems review: what exists, what is accessible, what is a mess
- Every candidate scored on frequency, hours, judgement required and cost of error
- A ranked shortlist with estimated hours recovered and dollars attached
- A build plan for the top three, with cost, timeline and what could go wrong
What you walk away with
A document you could hand to any competent implementer — us or someone else. That is deliberate. The audit is worth having even if you never build anything with us.
- Task inventory
- Every repetitive task, with hours per year attached
- Ranked opportunities
- Scored by payback speed and risk, not by how impressive they sound
- What not to do
- The candidates we assessed and rejected, and why
- Baseline measurements
- Timed before-figures, so the after-figure means something
- Build plan
- Scope, cost and timeline for the top three
Where it pays
The tasks that come up again and again.
This is not a menu to order from — your audit will surface your own list. But these are the areas where we most often find the first win.
Sales & enquiries
- Enquiry triage and routing by intent
- Quote and proposal drafting from your price list and past jobs
- Follow-up sequences that reference the actual conversation
- Lead qualification before it reaches a person
Admin & back office
- Supplier invoice and receipt extraction
- Client and candidate intake with document chasing
- Data entry between systems that do not integrate
- Meeting notes into decisions and assigned actions
Finance & reporting
- Ledger exception review across the whole file
- Monthly report drafting with commentary on variances
- Debtor follow-up drafted in your tone, escalated on a schedule
- Budget-versus-actual narrative for management meetings
Compliance & quality
- File review against a checklist, every file, every time
- Contract and agreement review for missing clauses
- Policy question answering from your own documentation
- Audit-readiness checks before someone asks
Client service
- First-line responses to documented, repetitive questions
- Personalised update drafting at volume
- Knowledge lookup across your own files and history
- Handover notes when work moves between people
Operations
- Roster and timesheet reconciliation
- Job status chasing across systems
- Exception alerts on the numbers that matter
- Procedure documentation from how the work is actually done
Step two
Build, deploy, train, measure.
You choose which items from the shortlist to proceed with. Each one is scoped and priced separately, so you are never committing to a programme.
- 1
Design the guardrails
What the agent may and may not do, what gets validated mechanically, and where a human approves before anything leaves the building.
- 2
Build and test on real work
We run it in parallel with the existing process on live cases, compare outputs, and tune until the error rate is acceptable for that task's risk.
- 3
Deploy and train the team
Documentation, a walkthrough with the people who will use it, and the manual step actually switched off — because running both is how projects quietly fail.
- 4
Measure and hand over
The after-figure against the baseline. Then either we support it or you run it — your call, and there is no lock-in either way.
How we build
Four rules we do not bend.
Least powerful tool that solves it
If a rule can do the job, we use a rule. Rules are cheaper, faster, deterministic and still working in three years. Agents are reserved for the part that genuinely needs judgement — which is usually smaller than the demo suggests.
Read-only until proven
Agents analyse and propose before they ever act. Nothing writes to your ledger, pays a supplier or emails a client without a validated path and, where the cost of error is real, a human approving.
Validate mechanically, never trust output
Does the GST equal one eleventh? Does the ABN exist? Does the total match the lines? Cheap deterministic checks catch the overwhelming majority of model errors, and they cost nothing to run.
Measure or it did not happen
A baseline before, a measurement after, in hours. If a build does not deliver the number, we tell you. An honest failure is more useful to your next decision than a comfortable story.
Why us
We are an accounting practice that runs on this, not a consultancy that reads about it.
Our own client work uses AI agents for ledger exception review, document extraction and report drafting. Everything we recommend is something we have already had to build, deploy and then maintain when it broke.
- We understand your numbers, so we can put a dollar figure on recovered hours that means something
- We know what regulated, audit-trailed work requires, because we do it
- We have made the mistakes already — on our own files, at our own cost
- We stay after go-live, because an agent nobody uses is not a saving
- No lock-in: the build is documented and yours to take elsewhere
Questions
What business owners ask us.
We're not a tech company. Is this for us?
It is mostly for businesses that are not tech companies. The best candidates are ordinary operating businesses — trades, health and NDIS providers, professional services, wholesalers — where a small number of repetitive, text-heavy tasks quietly consume ten to twenty hours a week. You do not need technical staff. You need someone who will do the work of finding the tasks and building against them.
How is this different from just buying an AI subscription?
A subscription gives your team a tool and hopes they work out what to do with it. Most do not, because the hard part is not the model — it is deciding precisely what good looks like, getting clean access to the data, and changing the process so the old manual step actually stops. That is what we do. The subscription is a small line item inside the project.
Will this replace my staff?
That is not what we sell and it is rarely what happens. What changes is the mix of work — the mechanical portion shrinks and the judgement portion grows. In practice most clients redeploy the recovered hours into work they were already behind on, or into capacity they could not previously afford to hire for. If your goal is headcount reduction, we are probably not the right fit.
What about our client data and privacy?
Every build runs on business-tier services with contractual commitments that your inputs are not used to train models, on named accounts with access control, and with a documented data-handling policy. Where personal information is involved we work through the Australian Privacy Principles — particularly APP 6, APP 8 on overseas disclosure, and APP 11 on security — and we tell you plainly if something you want to do is not advisable.
How do you know if it actually worked?
We time the task before we start and time it after. Every engagement includes a baseline measurement, so at the end you have a before-and-after in hours and dollars rather than an impression. If a build does not hit the number, we say so — that is more useful to you than a success narrative.
Do you do this in your own practice?
Yes, which is the main reason we can talk about it usefully. Our accounting work runs on AI agents for ledger exception review, document extraction and report drafting. We only recommend patterns we have already put into production and had to maintain.
Let's find out where your hours are going.
A free thirty-minute conversation about what your team spends its week on. If there is nothing worth automating, we will tell you that — it is a short call and an honest one.