Useful AI, without your data leaving the building.
Some businesses hold records that should never sit on someone else's computer — a nursery's incident notes, an accountant's client files. That shouldn't lock you out of AI; it should change how yours is built. There are three honest ways to do it, and we'll tell you which one you actually need — including when the cheapest one is fine.
Same AI, three places for your data to live.
Most businesses
Cloud with safeguards
Your AI runs on Anthropic's Claude — the engine behind everything we build. Their commercial terms mean your data is never used to train their models, we'll sign a DPA, and everything travels over enforced TLS.
- What leavesThe task and its text, encrypted in transit
- Who sees itAnthropic, as a contracted sub-processor
- Best forMost service businesses — these safeguards genuinely are enough
- The honest trade-offYour text is processed on someone else's infrastructure, under contract
The middle path
Hybrid — private by design
Before anything leaves your premises, a local layer strips out the people: names, contact details and account numbers become anonymous placeholders. The cloud AI does its thinking on text that identifies nobody, and the answer is re-personalised back on your side.
- What leavesThe work — with every identity removed before it goes
- Who sees itThe cloud sees ⟨PARENT_1⟩. It never sees the name
- Best forSafeguarding notes, client records, financial details
- The honest trade-offA detection step we build and tune to your data, per client, as part of your project
Full cloud intelligence, applied to text that could be anybody's. For sensitive records, this is usually the setup you actually want.
Ask us about itNothing leaves
Fully private
The AI model itself runs on hardware in your building, or on a private server only you control. No cloud, no sub-processor, nothing to redact — the data simply never goes anywhere.
- What leavesNothing. There is no outbound path to a cloud AI
- Who sees itYou. That's the list
- Best forRules — or nerves — that say the data cannot leave, full stop
- The honest trade-offOn-site models trail the frontier cloud ones, and the hardware costs real money. We'll say so if it isn't worth it for you
These are build architectures we scope into your project, not boxes on a shelf — the setup, like the price, is fixed in writing before we build.
Here's what the cloud actually sees.
On the hybrid setup, this is what happens to a nursery manager's note before it goes anywhere:
Text description of the film
A 26-second animated film. No speech, no audio track — everything below is on screen as text and animation.
0:00–0:05 — "A note full of personal details." A wide shot: a large box labelled "Your premises" on the left holds a note and a padlocked box marked "Never leaves"; a dashed boundary line separates them from a cloud labelled "Anthropic · Claude" on the right. The camera moves in on the note, which reads "Priya Sharma grazed her knee in the garden. Let her mum Anita know on 07700 900123, and attach March's invoice" — an orange marker sweeps across the two names and the phone number.
0:05–0:09 — "Stripped out — before anything leaves." The camera pulls back. The three highlighted details lift out of the note and arc, one by one, down into the padlocked box, each landing with a small ripple. Grey stickers reading ⟨CHILD_1⟩, ⟨PARENT_1⟩ and ⟨PHONE_1⟩ stamp over the gaps they left, and the padlock clicks shut.
0:09–0:12 — "Only the anonymised work crosses the line." An orange line draws itself from the note, across the boundary, to the cloud. A small copy of the note — grey lines and a ⟨CHILD_1⟩ placeholder, nothing readable — travels along it and is swallowed by the cloud, which swells briefly.
0:12–0:15 — "The personal details? They don't get out." The camera moves to the boundary. While the cloud pulses in thought, an orange dot of personal data races from the vault at the wall — and slams into it. The wall flexes and holds; an orange ring bursts at the impact; the dot squashes, drops, bounces twice on the floor and rolls back into the vault.
0:15–0:20 — "The answer comes back — names restored, on your side." A small "Draft reply" card travels back from the cloud across the boundary. Inside the premises it opens into a full card: "Draft reply — ready for your approval. Hi Anita — Priya's absolutely fine. She grazed her knee playing in the garden this afternoon; we cleaned it up and she was straight back out there. March's invoice is attached." The names Anita and Priya highlight in orange as they are restored, and an orange tick draws itself in the corner.
0:20–0:26 — The scene fades to near-white and the Zone Ten AI wordmark draws itself letter by letter, the orange tick lands in the O, the orange dot drops onto the i and bounces to rest, and the closing line "Private by design." appears.
"Draft this afternoon's incident note: Priya Sharma grazed her knee in the garden. Let her mum Anita know on 07700 900123, and attach March's invoice."
"Draft this afternoon's incident note: ⟨CHILD_1⟩ grazed her knee in the garden. Let her mum ⟨PARENT_1⟩ know on ⟨PHONE_1⟩, and attach March's invoice."
The list matching ⟨CHILD_1⟩ to a real child exists only on your premises. The drafted reply comes back, the real names go back in on your side, and the cloud never knew who anyone was.
Most businesses don't need the expensive one. Anthropic's safeguards are real, and for most of what we build, cloud with safeguards is genuinely enough — we'll tell you when it is. The other two setups exist for the records where "a contract says it's fine" isn't the answer you want to give a parent, a client or a regulator.
What this looks like when it's your data.
Incident notes, safeguarding records, parents' contact details — data you'd never forgive yourself for leaking. On the hybrid setup the AI drafts the parent updates and chases the unpaid fees, and no child's name ever reaches the cloud. The writing is as good as the frontier model that wrote it; the identities stayed in the building.
Client financials, payroll, HMRC references. The AI summarises the file, drafts the letter and builds the January chase list — working entirely on ⟨CLIENT_1⟩ and ⟨REF_1⟩. Your client list is the business, and its privacy doesn't rest on a contract, because it never left.
Decided at scoping, like everything else.
Map the data
We look at what you actually handle. Not every business with a spreadsheet needs a bunker — and we'd rather prove that than sell you one.
Pick the setup — per job
The chatbot answering opening hours can live in the cloud while the safeguarding notes stay home. It's not one decision for the whole business.
See it working
Before anything goes live you watch it run — and on the private setups, that includes exactly what left the building and what didn't.
Tell us what you handle. We'll tell you what should leave the building — and what shouldn't.
Twenty minutes, straight answers — including "the standard setup is fine" if it is.
Zone Ten handles it.