Your private AI workbook.

About an hour with a coffee and a list of the tools your team uses. Fill it in rather than read it. By the end you'll have an inventory of what leaves the building today, your own red, amber and green classification, the obligations that apply to you, a first project worth doing, and the cloud versus private sum run at your actual seat count. Type straight into the boxes and save as a PDF, or print it for a team session.

The short version. Which plan you are on decides what happens to your data: consumer tiers and enterprise tiers are different contracts, and a court in another country can override a vendor's commitments regardless. Residency is storage location; sovereignty is which country's laws can compel access. APP 1.7 commences 10 December 2026 (disclose the kinds of personal information used in significant automated decisions and the kinds of decisions made). The small business exemption has not been removed. Below about 30 seats, on-premise generally loses on cost, so buy it for confidentiality, unmetered use and a bill that doesn't move. Aim for a hybrid: red stays in, amber goes out de-identified, green is cost-routed, with one gateway enforcing it. General information, not legal advice.

1What leaves the building today

List every AI tool anyone in the business touches, including personal accounts used for work. For each, the plan tier is the thing that matters most. If you don't know, write "unknown" and go and find out, because that answer is itself a finding.

Tool	Plan tier	Who pays	What goes into it

For each paid tool: is training on your content on or off by default, and when did you last check? Tools people use that aren't on the official list (be honest, this is where the risk lives):

2Classify your information: red, amber, green

Use your own document names, not generic categories. Red never leaves the network. Amber can go to a cloud model once identifying details are stripped, and the fact it went is logged. Green is fine in the cloud and routed on cost. Do this with the people who actually handle the files.

Tier	Which of our documents (name them)	Where it may be processed
Red
Amber
Green

Australian identifiers that would need stripping from amber material (Medicare numbers, ABNs, TFNs, matter or file references, patient IDs): Today, is there a single point every AI request passes through, or can anyone paste anything anywhere?

3Your obligations checklist

Tick what applies to your business, then note who is going to confirm it. General information only: your own adviser or professional body has the last word on every line of this.

We use personal information in decisions that could significantly affect a person, so APP 1.7 disclosure may apply from 10 December 2026. We are tax practitioners, so client permission is needed before client information goes into AI tools (TPB(GS) 55/2026). We became an AUSTRAC reporting entity on 1 July 2026 (law, accounting, conveyancing, real estate, trust and company services), with seven-year record keeping. We hold health records or other sensitive information. We hold privileged material. A client contract or master services agreement restricts AI processing of their material. We genuinely don't know yet, and finding out is the next job. Who is going to confirm each ticked line (name the lawyer, accountant or professional body, and by when): Decisions in our business where software materially shapes the outcome for a person (the APP 1.7 list): Do our current client permissions and engagement letters match what we actually do with AI? Who is accountable for AI in this business, by name (the first of the six AI6 practices):

4Your first-project brief

One clearly scoped job beats a plan to "sort out AI". Transcription is usually the best first project: fastest visible win, strongest privacy story. Classification and routing has the best economics. Keep long agentic chains on a frontier model.

The first project: Which tier does its material sit in (red, amber or green), and why: Is it single-pass work or a long chain of dependent steps? (Single-pass suits local; long chains suit a frontier model.) Where the human sign-off sits, and whether it blocks the workflow or just sits alongside it: Who is allowed to see what, and how the system will know (permissions are what sink these projects): What "done and working" looks like, in one sentence someone could test:

5Cloud or private: your decision page

Fill in your own numbers next to the reference points. The reference column is the modelled annual cost at three team sizes, including hardware amortised and an allowance for administration. Below about 30 seats, private generally loses on cost alone, so the deciding factors become confidentiality, unmetered use and a bill that doesn't move.

Team size	SaaS per seat	API usage	On-premise
Reference: 10 people	$5,400	$4,435	$7,700 (loses)
Reference: 30 people	$16,200	$13,306	$14,050 (a tie)
Reference: 100 people	$54,000	$44,352	$28,757 (wins)
Our team size:

The costs that get forgotten. Put a number on each.

Maintenance: 5 to 10 hours a month, at our rate, per year
Who does that maintenance, by name
UPS, circuit capacity and where the machine physically lives
Backups of the things that aren't re-downloadable (fine-tunes, embeddings, vector database, prompts)
Who fixes it at 2am, and what the fallback is

The reasons that would justify private even if the arithmetic says no (a contract clause, a professional obligation, a regulator, unmetered use): Our verdict for now: cloud set up properly, hybrid, or private. And the date we'll revisit it:

If the verdict is "keep the cloud tool we have, configured properly", that is a legitimate result and it costs nothing. Remember you can pilot in the cloud and move later: every serious local stack exposes an OpenAI-compatible endpoint, so migrating is a base URL change rather than a rewrite.

Your private AI workbook.

Sent, thanks.