AI that helps your numbers, not just your slide deck
I build AI on your real data and real workflows — so it ships and stays useful. I also say when a simple repeatable tool is better than AI.
Common signs you need help
- Your team keeps asking "what should we do with AI?" but nothing ships
- You tried ChatGPT on the side and it doesn't connect to your real tools
- Vendors sell AI features you don't need while the real problem stays stuck
Who this is for
This is for owners who have watched a team trial ChatGPT for three months without anything reaching a customer. Usually there is real enthusiasm, a few clever prompts saved in a document, and no connection between any of it and the CRM, the job data, or the inbox where work actually arrives. If you can name a task someone does daily that involves reading, summarising, classifying, or drafting, there is probably something here worth building.
Where this applies
Many Australian businesses trial ChatGPT in isolation while customer data and ops tools stay disconnected — I build AI that sits on your real workflows.
AI copilots and document workflows for teams under 50 staff.
What the assessment covers
Map where AI would actually save time or make money
Pick the one task where AI would pay for itself first
Honest go/no-go — I'll say when AI isn't the right fix
A build, end to end
A first build often looks like this: quoting rules, product specifications, and warranty terms live across a shared drive, several PDFs, and one long-serving staff member everybody interrupts.
- Agree the question set the tool must answer, and the documents that are actually authoritative for each one.
- Connect those documents as the only source the assistant may draw on, so it cannot invent a specification.
- Test it against real questions your team asked last month, and compare the answers to what the expert would have said.
- Put it where the work already happens — the inbox or the tool your people already have open — rather than behind another login.
Outcome: Specification questions get answered without interrupting the one person who knows, and every answer cites the document behind it — so when a rule changes you update one file instead of retraining people.
What I build
In-tool helpers, ask-your-files bots, lead sorting, AI-assisted quoting
A typical first build is an answer bot over your own documents: quoting rules, product specs, warranty terms, whatever your team currently asks the one person who knows. It sits in the tool they already have open, cites the source document, and says it does not know rather than inventing an answer. Others I have scoped include inbound enquiry classification that routes to the right person with a drafted first reply, and quote drafting that pre-fills from historical jobs for a human to check and send.
How a build runs
Timings assume you can give me a few hours of subject-matter time in the first and third weeks. The evaluation step is the one most projects skip and the reason most AI pilots quietly stall.
Free call and working sample30 min, then 2 business days
Map where AI would genuinely save time, and rule out the tasks that need a deterministic flow instead.
Data and accessWeek 1
Agree which documents and systems are authoritative, and get read access without opening anything you should not.
Build and evaluateWeeks 2-4
Wire the model to your data, then score its answers against questions you already know the answer to.
Pilot with one teamWeeks 4-6
One group uses it for real work while I watch failure cases and tighten the prompts and sources.
What it costs and how I scope it
It starts with a free 30-minute call: you show me the task, and I tell you honestly whether AI is the right fix. If it's one process, you get a working sample on demo data within two business days, then a fixed price from $2,500, agreed in writing. The cost driver is rarely the model — it is the plumbing to your data and the evaluation work to prove the output is reliable enough to trust. Bigger AI work that spans several systems is scoped with a paid Opportunities Report first. Ongoing model usage is billed at cost and is usually small relative to the build.
Every fix starts with a free 30-minute call and a free working sample, so you know the fixed price before work begins. Bigger projects get a paid Opportunities Report first.
When this is the wrong fix
AI is the wrong fix when you need the same answer every time. Approvals, compliance checks, pricing rules, and anything auditable should be a deterministic flow, not a language model. It is also the wrong first purchase if your underlying data is scattered across three systems and nobody agrees which one is correct — clean that up and the AI build gets cheaper and better. I will tell you when that is the case rather than sell the more interesting project.
Related reading
FAQ
- Do you train models from scratch?
- Rarely. I use proven models and focus on wiring them into your tools and workflows.
- Is AI always the right answer?
- No. Sometimes you want the same result every time — a checklist, a fixed flow, a plain tool. I use AI when it helps; when repeatability is better, I build that instead.
- Will I own what you build?
- Your data is always yours. You get a permanent licence to use and change your system, and nothing locks you in. I keep the building blocks I reuse across clients, which is a big part of how I can build it from $2,500.
Got a process you've outgrown?
Show me on a free 30-minute call. You'll see a working sample of the fix within two business days, before you pay a cent.
Book my free 30-minute call