JDCS
Start a conversation
Start a conversation Or call

jdcs.au · Mid North Coast, NSW

Local & private AI · Australia-wide

Can you actually put client files through ChatGPT?

Sometimes yes, once it's set up properly. Sometimes no, and then the answer is AI that runs on hardware you own, where the files never leave the building. I work out which one you're looking at and put it in writing you can hand to a client, a board or an insurer.

Start with a free look at where your data goes Want it in writing? The AI audit

5.0 on Google

Most businesses don't need to buy a server. A few genuinely do.

Public AI tools are not designed for every risk profile, and that's a different statement from saying they're bad. I'm James, a family-run business on the Mid North Coast (who you'll work with). If the tool you're already paying for is fine once it's configured properly, I'll tell you that and charge you nothing for the relief.

Three answers

There are three ways to fix this. Only one needs hardware.

Ranked cheapest and least disruptive first, which is also the order I work through them. Plenty of businesses stop at the first one.

  1. 01

    Keep the tool you have, set up properly

    Business and enterprise plans generally do not train on your data, and most consumer plans do. Half the risk people worry about is a settings problem, a plan problem or an over-shared folder. Often this is the whole fix, and it costs nothing.

  2. 02

    Same tools, held in Australia

    Some workloads can be pinned to Australian infrastructure so the processing happens here. That answers a data residency question. It does not answer a question about which country's laws could compel access, and I will not pretend otherwise.

  3. 03

    Private AI on hardware you own

    An open model running on a machine in your office or your rack. Nothing is sent to any AI company, because there is nowhere for it to go. No per-seat bill, no usage meter, and nothing that can be deprecated out from under you.

Australia's privacy regulator has said that deploying AI locally is "likely to be more privacy-preserving as it limits the risks of third party access to the data". That's a reason to consider option three, not a reason to skip one and two.

What it's good at

Where a private model earns its keep.

These are the jobs where an open model running on your own hardware is genuinely good enough, and where you would struggle to pick it from a cloud service.

  1. Meeting and consult notes

    Recording and transcription that happens on your own machine. Usually the best first project: quick to stand up, obviously sensitive, and the local quality genuinely competes with the cloud services.

  2. A searchable company brain

    Plain-English questions answered from your own files, with citations back to the source document. The hard part is permissions, not the AI, and that is where the work goes.

  3. Pulling data out of documents

    Invoices, purchase orders, forms and contracts turned into clean structured data. Header fields are close to solved. Line items still need a person checking the ones the system is unsure about.

  4. Sorting and routing at volume

    Tagging, triage and classification across thousands of items. This is where per-token cloud bills quietly compound, and where a small local model often beats a large rented one.

  5. Redacting before anything leaves

    Stripping names, numbers and identifiers out of a document before it goes anywhere near an outside service. This is what makes the safe use of cloud AI possible for the rest of the work.

  6. Scanned archives you cannot search

    Decades of paper turned into something you can actually find things in. Unglamorous, low risk, and it usually pays for itself faster than anything else on this list.

The honest part

Where private AI falls short.

Open models sit roughly one release cycle behind the best cloud models. On the everyday work above you will not notice. On hard multi-step coding, long autonomous tasks and complex reasoning across a whole contract, the gap is real and it shows. Under about thirty staff, owning the hardware usually costs more than renting the software, so buy it for control rather than savings. Somebody has to spend a few hours a month keeping it healthy. And a model that talks to your customers needs guardrails around it, because open models refuse far less than the commercial ones do. If any of that rules it out for you, that's a useful thing to learn before you spend money, not after.

What most people land on

Private for the sensitive work. Rented for the rest.

Telling a business to abandon its whole stack is telling it something it will never do. The version that actually works sorts your information into three buckets. The genuinely confidential material never leaves your network. The middle tier can go out once names and identifying details are stripped from it. The public material goes wherever is cheapest and best. One gateway sits in front of all of it, enforcing the rule and keeping a log of what went where.

The part that makes this safe to commit to: local systems speak the same interface the big providers use. Pointing an existing tool at your own model is usually a change of address rather than a rebuild, so you can pilot something in the cloud, prove it works, and bring it in-house later without starting again.

What it costs

Real numbers, including the bad ones.

A single capable machine for a small team starts around $3,500. A box that will comfortably serve a whole office runs to roughly $18,500 for the graphics card alone. Power is almost irrelevant, a few hundred dollars a year, so ignore anyone who leads with the electricity bill. The cost people miss is the five to ten hours a month of looking after it, which at consulting rates is often larger than the hardware.

One more thing worth saying out loud in 2026: memory prices have run hard this year, and graphics cards are selling well above their list price. If you can wait six months, waiting is a legitimate strategy. The full working, with the break-even seat counts, is in what it costs to run AI on your own hardware.

Start here

A free look at where your data goes.

Tell me what your team uses and what sort of information goes into it. I'll tell you which of the three answers you're most likely looking at, and whether it's worth going any further. No obligation.

If you need the full answer in writing, with every tool checked and your obligations verified, that's the AI audit: $2,450 fixed, and it comes off a build.

Prefer to just talk? Call 0418 858 937.

Step 1 A few quick questions

What prompted the question?

Whatever's driving it. This shapes what I check first.

What AI tools are in play?

Official or not. "Not sure" is a common answer, and a useful one.

What sort of information does the business hold?

Roughly is fine. The proper tiers get drawn in the audit itself.

Tell me about the business.

How many of you?

Where should I get back to you?

I'll come back with whether the audit is worth doing for you at all. No obligation, and sometimes the answer is one setting.

Questions

The ones that come up most.

What is local AI?
AI that runs on a computer you own instead of a service you rent. The model is a file sitting on your own drive, and the work happens on your hardware. Because there is no outside service involved, nothing you type or upload is sent anywhere, and you can confirm that by watching the network or simply unplugging it.
Is my data really private if I run AI locally?
Yes, on the point that worries people most. An open model is a static file of numbers, not a program that can call home. Once it is running on your machine, your documents stay on your machine. What still needs attention is who inside your business can see what, which is an access control question rather than an AI one.
Do I need to buy a server?
Often not. For a lot of businesses the honest answer is that the tool you already pay for is fine once it is configured properly, and a policy plus a settings change gets you there. Buying hardware makes sense when you have a real obligation that cloud cannot satisfy, or enough volume that the maths turns over.
How much does private AI cost in Australia?
Hardware for a small office generally runs from about $3,500 for a single capable machine to around $18,500 for a card that will serve a whole team, plus setup. Running cost is small, in the low hundreds of dollars a year for power. The bigger ongoing cost is the few hours a month somebody has to spend keeping it healthy.
When is local AI cheaper than paying per seat?
Later than most people expect. Around ten staff, cloud usually wins on cost. Around thirty it is close, and which way it falls depends on how much admin the system actually needs. Past roughly sixty or a hundred seats it is not close, and owning the hardware wins clearly. Below that, buy it for control, not to save money.
Are open AI models good enough?
For summarising, extracting, classifying, transcribing, translating and answering questions from your own documents, yes, and you would struggle to pick the difference. For hard multi-step coding and long autonomous tasks, the best cloud models are still clearly ahead. Measured across the industry, open models sit roughly one release cycle behind, and that gap has held steady rather than widened.
Are Chinese open models a security risk?
The risk people picture belongs to the hosted app, not the model file. Running the weights on your own hardware sends nothing to anyone. What is worth managing is where you download from, checking the file against its published checksum, and firewalling the machine. Worth knowing: when the Australian Government directed agencies to remove DeepSeek's products and web services in February 2025, that direction expressly excluded open models deployed locally with appropriate mitigations in place.
Can I move to local AI later without redoing everything?
Usually yes, and this is the part that de-risks the whole decision. Local AI systems speak the same interface the big providers use, so pointing an existing tool at your own model is normally a change of address rather than a rebuild. You can start on a cloud service, prove the idea works, and move it in-house afterwards.
What happens if you get hit by a bus?
Fair question for a one-person business, and the answer is that everything is documented and handed over. You hold the credentials, the runbook is a deliverable rather than a secret, and the system is built from open components so any competent IT person can pick it up. If you want to stop working with me, it keeps running.