// what we build
Build specs, not buzzwords
Each kind of build below is described the way we’d write it up in a proposal: what goes in, what comes out, where a person signs off, and what it can’t do. Example scenarios are illustrative.
01 AI agents
An agent handles a defined job across several steps and systems. Unlike a chat assistant, it doesn’t wait to be asked: it’s triggered by something arriving — an email, a form, a new record — works through the job, and stops at the point where a person needs to decide.
| Triggered by | New email in a shared mailbox, a web form, a new record in your CRM or job system, a schedule |
|---|---|
| Reads | The incoming item and attachments, plus lookups in your systems (customer history, price lists, availability) |
| Produces | A prepared action: draft reply, draft quote, new record, task, summary |
| Human step | Approve, edit or reject before anything leaves the business or changes a record of consequence |
| Logs | Every run: input, what it looked up, what it proposed, who approved |
Example scenario
A commercial cleaning company receives tender requests and site enquiries in one inbox. The agent reads each one, pulls out site size, frequency and location, checks the company’s rate card, drafts a first-pass quote and response, and puts it in front of the operations manager in Teams with a short note on anything unusual. The manager edits and sends. Turnaround drops because the first draft is already there when they open it.
Limits
- It follows your rules; it doesn’t invent pricing policy. If the rate card doesn’t cover a case, it flags it.
- It isn’t given authority to pay, sign or commit without a person.
- Jobs with many judgement calls and few patterns are poor candidates.
02 Assistants on your own documents
Often called a “chatbot on your data”. Your documents are indexed so that when someone asks a question, the assistant finds the relevant passages and answers from them, citing the source. The technique is usually called retrieval-augmented generation; what matters to you is that answers trace back to your own files.
| Sources | SharePoint or Google Drive folders, policy manuals, product sheets, past proposals, knowledge-base articles |
|---|---|
| Users | Your staff, inside Teams or a simple web page; public use only after a staff-only phase |
| Permissions | Respects who can see what: a person shouldn’t get answers from a folder they can’t open |
| Answers | With a link to the source document, and “I don’t know” when the documents don’t cover it |
| Upkeep | New and changed documents picked up automatically; out-of-date ones need someone to archive them |
Example scenario
A building services firm has years of method statements, safety procedures and product data sheets spread across shared drives. Site supervisors ask the assistant “what’s our procedure for working on a fragile roof?” and get the relevant section with a link to the current document, rather than ringing the office.
Limits
- It’s only as good as the documents. Contradictory or outdated files produce contradictory answers — tidying them is part of the build.
- It summarises; it doesn’t replace a qualified person’s judgement on safety, legal or financial questions.
03 Email and document automation
The narrowest and often most valuable build: take a document or message of a known type and turn it into structured data or a filed, named, logged item.
| Inputs | Supplier invoices, purchase orders, application forms, contracts, timesheets, delivery dockets, scanned PDFs |
|---|---|
| Outputs | Draft bills in Xero or MYOB, rows in a spreadsheet or system, renamed and filed documents, exceptions list |
| Checks | Totals add up, supplier and ABN match records, required fields present; anything uncertain goes to a person |
| Human step | Review the exceptions and approve the batch |
Example scenario
A wholesaler’s accounts team receives a few hundred supplier invoices a month as PDF attachments in every layout imaginable. The automation reads each, matches the supplier, codes the lines against the usual accounts, creates draft bills and lists the ones it wasn’t sure about. The team reviews the exceptions instead of keying every invoice.
Limits
- Poor scans and handwriting reduce accuracy; we measure it on your real documents before promising anything.
- Drafts, not final postings, until you’ve seen enough runs to trust a category.
04 When a ready-made platform beats a build
Sometimes the honest answer is that you don’t need a custom build. If what you want is an AI assistant for the whole team that works across email, calendar, files, Xero and your job system, a ready-made product may cost less than building one. The company behind Gantry, SG1 Consulting, makes one: The Everything, which runs in Microsoft Teams and costs $150 a month for the first user and $90 a month for each extra user, plus GST. We’ll tell you when that’s the better option, and that it’s ours.
Not sure which of these your problem is?
That’s what the scoping call is for.
Book 15 minutes