Where public AI is the right call
Let's start where most of these pages won't: by telling you when not to call us.
For a great deal of work, public AI is exactly what you should reach for.
Need to write or debug code? The coding assistants — tools built around Claude and
the GPT models — are superb, and getting better by the month. Synthesising a market
scan or a literature review? Research tools like Gemini and Perplexity will pull
together a cited report in minutes. Tracking something breaking in real time? Grok,
wired directly into the live feed, is sharp at it. Drafting, brainstorming, thinking
out loud? The general assistants like ChatGPT are hard to fault.
If your work doesn't touch anything you'd be unhappy to see leave your building,
this is where you should be, and there's no shame in staying there. Buying private
AI to do what a public tool does perfectly well isn't prudence; it's AI for the
sake of AI, the expensive mistake a lot of organisations are making right now because
the word is everywhere. Don't make it.
The line where it stops being a choice
For a person, public AI is a tool. For an organisation holding other people's
confidential information, the same tool behaves differently — because the risk
isn't yours to carry alone.
A sole trader experimenting with ChatGPT is risking their own work. A firm turning
over millions, with hundreds of clients, sensitive records, and a regulator watching,
is risking everyone else's. At that scale the calculation flips: the choice is no
longer "which AI tool" — it's "use AI safely, or don't use it at all." And "don't
use it at all" isn't really an option any more, because your people will use it
anyway, quietly, on their own accounts.
That's the line. Below it, public AI is fine. Above it — once the work involves
information that can't leave your walls — there are only two honest answers: ban it
and lose the benefit, or run AI you actually own. Private AI is what makes the
second answer possible.
This isn't theoretical. Samsung's engineers crossed exactly this line in 2023 —
and how they crossed back is the clearest case study there is.
Read how Samsung handled this →
But doesn't an enterprise agreement cover this?
It's the fair question, and the honest answer is: partly. A proper enterprise
contract does give you real protections — confidentiality terms, a commitment not
to train on your data, detailed logs, a legal basis to act if something goes wrong.
That's genuinely better than a free account, and for a lot of organisations it's
enough.
But a contract protects your rights. It doesn't change where your data physically
is. Under an enterprise agreement, a copy of your confidential material still
leaves your environment, still sits on infrastructure you don't run, still falls
under the provider's jurisdiction — and if that provider is a US company, still
may sit within reach of US law regardless of which data centre holds it. A contract
gives you the right to sue after a breach. It can't make the breach impossible.
And some information — privileged, special-category, genuinely sensitive — can't
be un-leaked by winning a case afterwards.
That's the whole point of owning the system instead of renting access to someone
else's. Not a better promise. No promise needed.
What this actually feels like to use
Forget the word "self-hosted" for a moment. It sounds like a server room and a
team of engineers. For the person actually using it, it's nothing of the sort.
Your people open a tab in their browser. There's a familiar chat window — they
type, it answers, they get on with their work. It feels exactly like the public
tools they already know. The only difference is the one that matters: nothing
they type goes anywhere.
Think about how you already trust Word and Excel. You write a confidential letter,
you build a spreadsheet of real figures — and it never occurs to you to worry,
because that document isn't being sent off to anyone, isn't being read to train
someone else's product, isn't leaving for a server you've never heard of. It's
just your work, on your machine, doing what you told it to.
Your private AI sits in exactly that category. It's the same quiet, trustworthy
kind of tool — one that happens to think. So you can give it the document you'd
never paste into a public chatbot — because with this, there's nowhere for it to go.
Beyond the general assistants, there's a whole category of excellent
sector-specific software — tools built by talented teams for legal work, finance,
healthcare, and more. They're genuinely good at what they do, and they're built
for exactly the kind of professional work we're talking about.
Which is precisely where the dilemma bites. The moment you reach the point where
you can't afford to avoid AI, but the nature of your work obliges you to be
cautious with every document that passes through it, you're caught between two
instincts that pull in opposite directions: adopt the capable tool, or protect the
confidential material. Most of these specialist services are excellent right up
until you look underneath them — at whose cloud they run on, whose models they
call, and where your data travels in the process.
Private AI resolves the dilemma instead of forcing you to pick a side. The
specialist capability can live on the system you own. We can build a workspace
tuned to your kind of work — the same sort of sector-specific tooling, structured
around how your people actually operate — running inside your environment, on your
infrastructure, under your control. You get the specialist tool and the
confidentiality, rather than trading one for the other.
Not all private AI is actually yours
Here's where it gets subtle, because plenty of organisations think they've already
solved this. You can run private AI on a big cloud platform — a dedicated deployment
on Azure or AWS in a London region. Vendors offer it, and companies buy it. Your
data stays in the UK, there's a contract, there's a tenant that's notionally yours.
Isn't that the same thing?
Not quite. And the cleanest way to see the difference is to think about renting.
A private deployment on a big cloud is like renting through a platform like Airbnb.
Yes, you have a place to stay, and yes, it's yours for the duration. But you're
inside someone else's system, on their terms — the platform sets the rules, can
change the price, can change the conditions, and ultimately answers to a head office
on another continent. You don't hold a key to the building. You hold an account on
a service.
Owning your own system on dedicated infrastructure is like renting directly from
the landlord, on a contract you both signed. You still don't own the building —
nobody's pretending you bought the data centre. But there's no platform in the
middle, no one who can quietly change the terms underneath you, and — this is the
part that matters most — everything inside the place is yours and comes with you.
The furniture, your files, the way it's all arranged. If you ever leave, you move
it all to a new address and carry on exactly as before.
That portability is the whole game. Picture the provider vanishing overnight —
impossible in practice, but useful as a test. If your AI lived on a big platform's
managed service, it would vanish with them: the model, the configuration, the way
it was tuned were never yours to take — they're the platform's trade secret. If
your AI runs on your own system, the hardware is just hardware. Your system and
your data are yours. You redeploy with another provider, and you're running again —
same setup, same data, nothing to rebuild.
And it isn't only about disappearing. It's every “what if the platform
decides”: doubles the price at renewal, changes what's permitted, or is
compelled by its own country's law to hand over what passes through it. Rent
directly, with everything portable, and that entire category of risk simply doesn't
apply — there's no platform above you to make those calls.
One detail that makes it concrete: your system doesn't talk to servers across the
ocean — not for inference, not even for monitoring or support. There's no quiet
background connection to a head office somewhere else. Whether your infrastructure
sits in the UK or in the EEA, it answers to you and to its own jurisdiction,
full stop.
| Public AI | Private AI |
|---|
| Capability today | Bleeding edge, newest models first | Close to frontier on most professional tasks, on proven open models |
| Setup & effort | Instant, nothing to run | We build it and run it for you |
| Best for | Anything where a leak isn't a risk you carry | The same work — when a leak is a risk you can't carry |
| Your data | Leaves your environment | Never leaves it |
| Who controls the model | The vendor's — closed, their trade secret | Runs on infrastructure you control |
| Jurisdiction | The vendor's — often US, regardless of data residency or contract | Yours. Sovereignty itself |
| Cost shape | Per-seat or per-token — can spike with use | Fixed — doesn't grow with your team |
| If the vendor changes course | Your model can be deprecated, repriced, or restricted | Your system keeps running, unchanged |
Where it really pays off is the long game
Everything so far is true on day one. But the case for owning your AI gets stronger
the longer you look — because renting access to someone else's model exposes you to
a set of risks that compound quietly over years.
The bill that grows on its own.Public AI is usually priced per seat or per token, which means it scales with use, not with value. The teams who get the most out of it — the heavy users running document after document — are the ones who cost the most, so success quietly inflates the invoice. It's a well-documented trap: a monthly bill can swing tenfold on nothing more than a few power users and a busier month, and because providers bill by consumption with no application context, finance often reacts the only way it can — by throttling or banning the very usage that was working. Your own system is the opposite. You fix the cost for years at a time, flat, and it doesn't move when your team grows or your busiest people lean on it hardest. Add fifty staff; the number stays the same.
The model that gets pulled out from under you.When you build a workflow on a public model, you're building on something the provider can change or retire whenever it suits them. Models get deprecated, repriced, or quietly swapped for whatever the platform wants you on next — and the version your process depended on simply stops being available. On your own system, the model you chose keeps running for as long as you want it, exactly as it was. Your workflow doesn't break because someone else shipped something new.
The rules that aren't yours.Every public AI carries its provider's guardrails, content filters, and system instructions — and they can refuse your perfectly legitimate work because it tripped a rule written for someone else's risk appetite. On a system tuned to your standard, the tool does the work you need it to do, judged by your context, not a vendor's lowest common denominator.
The asset versus the rental.Rent, and at the end of every term you have nothing but the next invoice. Own, and the system compounds in value: your own documents become a searchable, answerable resource; the setup is tuned closer to your work over time; the whole thing becomes infrastructure your organisation holds, not a subscription it services. One is a cost that recurs forever. The other is something you build once and keep.
And the part people brace for — the running of it — isn't yours to carry. We build
it and we keep it working. You don't need a data centre, an engineering team, or a
fortune; you need a predictable monthly cost and us. It's far less than the scale
of the word "infrastructure" suggests.
