A private AI workspace that runs on your own knowledge.
A curated base of your organisation's own strategy, systems, obligations and market, served to AI. Ask it a question. Put a decision to it. Have it produce the board paper. Every claim carries its source, and nothing has to leave your walls.
For insurers, payers, providers, banks, holding groups and public bodies: organisations with real regulatory exposure and a standing advisory bill.
One knowledge base, and three things you can do with it.
The base is the product. Curated notes on your organisation and its market, linked into a graph, with each note traced back to where it came from and forward to the deliverables that depend on it. Not a folder of PDFs. Not a chat history. A store that is maintained, versioned, and yours.
Ask
Answers questions from your base, and cites the note behind every claim.
An answer with citations.
declines anything outside the base
Advise
Reasons over the same base, weighs the trade-offs, and recommends.
A recommendation with its trade-offs shown.
applies your precedents and standards
Generate
Turns the base into finished, branded deliverables.
A board paper, market map, register or diligence pack.
audited write · approval before accept
Ask, Advise and Generate all read from and write back to the same base. Every answer traces to a note, and every note traces to its origin. That is the whole difference between a tool your teams use and an asset your organisation holds.
You are not paying for the report. You are paying to learn the same ground again.
Every advisor you bring in will leave. That is the arrangement, not a complaint about advisors.
For each decision you bring in a team. They spend weeks learning your situation, hand you a conclusion, and go. The conclusion is worth having. The understanding underneath it, which is the part the weeks actually paid for, walks out with them.
Then the next question arrives. New team, same ramp-up, same fee, from zero. Nothing accretes. Five engagements later you hold five documents, each one ageing, and no more institutional understanding than you had at the start.
The question was never whether to bring in expertise. It is whether any of it stays.
Rented · resets each time
what you keep → nothing
Owned · compounds
what you keep → all of it
And the AI tools your teams already use keep nothing either.
While the advisory spend runs, the work is quietly moving into public chatbots. It is faster than waiting for anyone. It also fails in four specific ways, and they compound.
Confidential material you cannot un-send
To save an afternoon, someone drops a target's numbers, a regulator submission or a draft board paper into a public chatbot. That material now sits on a server you do not own, inside a product you cannot audit. There is no recall.
Fluent answers built on none of your facts
These tools answer from the public world. Your policies, your standards, your history are not in there. What comes back reads beautifully and cites nothing. Plausible is not the same as right, and it is far harder to catch.
Ten people, ten answers
Everyone prompts differently, so the same question returns a different answer each time and none of them is the version of record. When two teams disagree, there is nothing to check either one against.
Spend that never becomes an asset
Seats spread team by team and the annual bill climbs. At the end of the year you own no knowledge of your own, no audit trail, and nothing the next answer can be built on. Only a renewal.
Neither failure is a discipline problem, and no policy memo fixes either one. Both come out of how the tools are built, which is also where the fix comes from.
Grounded, governed, and reviewed before anything counts.
Two facts about the machine explain most of what goes wrong with it, and both point at the same fix.
Prediction, not retrieval
A model predicts the next word from patterns it learned elsewhere. It does not look anything up. That is why it is fluent, and why it can state something false with complete confidence. Nothing in the machinery distinguishes a fact it absorbed from a plausible shape that fits the sentence.
A desk of fixed size
Everything a model can hold at once sits on one desk: your question, the documents, the back-and-forth. Pile enough on and the earliest material slides off the edge while its attention thins across what remains. Late in a long session it forgets and drifts, and it does not announce either.
Which is why the knowledge sits outside the chat
Keep the durable knowledge in a governed store outside the conversation. Pull in only what the question needs. Make it cite what it used, so the work can be checked in seconds rather than trusted on faith. That is what Studio is, and it is the reason the same machine behaves differently inside it.
A model predicts the next token. It does not look anything up.
Next word, not a lookup
most likely next — not looked up
The desk of fixed size
earliest material falls off · focus thins
The fix, in three moves
The parts, plainly.
None of these is the model. That is the point.
guardrails · audit · approvals span every layer
Curated knowledge base
Notes on your organisation, its systems, its obligations and its market, written to be read by people and by machines. Maintained, not dumped.
Relationship graph
Notes are linked, so a question about a system also reaches the obligation, the decision and the risk attached to it.
Provenance trail
Every note records where it came from and which deliverables rely on it. When a fact changes, you can see what else changes with it.
Cited answers, honest refusals
Each claim carries its source. Questions the base cannot support get a refusal, not an answer invented to fill the space.
Voice and disclosure rules
Studio writes in your house style and holds to your disclosure boundaries. What it must not say is configured, not hoped for.
Audited write and execute
Producing a document is a permissioned, recorded action, not a chat message. You can see what ran, on what, and who asked for it.
Versioned candidates with approval gates
Regenerated work arrives as a candidate and waits for a named human to accept it. The review queue is the control, and it is visible.
Read-only MCP endpoint
The base is exposed to AI over an open protocol, read-only. Bring your own MCP-capable tools and ground them on it.
Isolated instance, cloud or on-prem
Your own workspace in your own identity, on its own instance, hosted or on your hardware. Data does not mingle between organisations.
Cloud, or entirely inside your network.
Same workspace either way. If your requirement is that nothing crosses the boundary, that is a deployment choice here, not a line on a roadmap.
Managed cloud
Fastest to stand up. Your base and your workspace stay yours and isolated; the model call reaches out to a managed provider.
one hop crosses the boundary
On-premises
A local model on your own hardware. Every part of the loop sits inside your walls, including the model itself.
nothing crosses the boundary
What actually changes once you own the base.
Outcomes, not features. Each of these is something that is true of your organisation today and stops being true.
The next question starts from somewhere
The work done for the last decision is still there when the next one arrives. Ramp-up shrinks, because your situation no longer has to be explained from the beginning to someone new.
One answer of record
Asked twice, in two teams, the same question draws on the same base and the same standards. Disagreement becomes a matter of judgement rather than a matter of whose version is correct.
Confidential material stays inside
There is a sanctioned place to do the work, so there is far less reason to paste anything anywhere else. Run it on your own hardware and the question stops existing.
Spend becomes an asset
A consulting engagement buys a document that starts dating on delivery. A fraction of that spend, redirected, buys a base that deepens every time it is used. Worst case it never compounds and you still own the base.
Deliverables arrive as deliverables
Board papers, registers, market maps and packs come out in your format, sourced and ready to read. Gathering, first-drafting and formatting stop eating the week, so your people spend it on the argument instead.
The audit trail exists before anyone asks for it
Sources, versions, approvals and what produced what are recorded while the work happens. Regulatory and internal-audit questions get answered out of the record instead of reconstructed from memory months later.
| Dimension | An engagement you rent | A base you own |
|---|---|---|
| Who holds the understanding | The team that leaves | Your organisation, in the base |
| What remains after delivery | A document, ageing from day one | The document and the reasoning behind it |
| The second question | New team, same ramp-up, same fee | Starts from what the first one established |
| Answer consistency | Depends who you asked | One base, one set of standards, sources shown |
| Where the data sits | On their laptops and in their tenant | Your instance, cloud or your own hardware |
| What the spend becomes | An expense, renewed each time | An asset that deepens with use |
Organisations carrying regulatory exposure and a standing advisory bill.
Both conditions matter. Regulatory exposure is what makes grounding, citation and an audit trail worth paying for. Recurring advisory spend is what makes owning the base cheaper than renting the answer. If only one of the two is true of you, Studio is probably premature, and we will say so.
Insurers and payers
The questions that return every renewal and every review: what the policy actually covers, what the last three vendor assessments concluded, what was filed with the regulator and when. Sourced answers instead of a rebuilt spreadsheet.
Providers and health groups
Service-line business cases, licensing and accreditation paths, capacity and staffing plans. The standards and precedents sit in the base, so the next site is not argued from first principles.
Banks and financial services
Obligation registers, control mappings, vendor and model risk. Ask what was committed, to which supervisor, in which submission, and get the paragraph rather than a recollection.
Holding groups and family offices
The knowledge sits inside the operating companies and rarely reaches the centre, which is where capital is actually committed. Studio gives the centre a base of its own: screening, benchmarks across the portfolio, first-hundred-days plans that start from what the group already learned.
Public bodies and regulators
Policy positions, prior determinations, published commitments. Answers that cite the document, and an on-premises deployment for material that cannot leave the network.
The questions we get every time, answered straight.
Why not the assistant we already have?
Enterprise chat assistants are useful and most organisations should have one. They answer per person, in a vendor's tenant, from whatever that person happened to paste in. Nothing accumulates, nothing is cited, and there is no version of record. Studio is the missing half: the governed base underneath. If your assistant speaks MCP, it can read that base too.
Why not build it ourselves?
You can, and some of you will. The model is the easy part, and the first API call is a day's work. The cost lives in the grounding, the citations, the audit trail, the review gates and the discipline to keep the base current. That is plumbing. Plumbing is worth buying. The data and the knowledge that accrues on top of it are the parts worth owning, and those are yours either way.
Why not wait a year?
The models will be better next year. The base will not have built itself in the meantime, and the base is the part that takes time. Whatever you learn now is what the better model gets to work with.
Are we locked in?
The base is notes and a relationship graph in readable formats, served read-only over an open protocol. Point another AI at it, export it, or run the whole thing on your own hardware. If you stop working with us, you keep the base and it keeps working.
Who maintains it?
Someone has to, and that is a real commitment rather than a footnote. Hildens stands the instance up and does the first curation, senior-led. After that, keeping the base current can sit with us, with your team, or be split between the two, and it gets scoped with you rather than assumed. What matters is that it is named. An unmaintained base decays quietly, and we would rather say that now than have you find it in month six.
Where does our data actually sit?
Each instance is isolated and data does not mingle between organisations. If your requirement is that nothing leaves your network at all, that is the on-premises deployment with a local model, not a promise about somebody else's terms of service. We would rather you deploy it the way your obligations require than take our word for it.
Four steps. The first real output is work you were going to pay for anyway.
You will spend on the next significant decision regardless. Build it inside Studio instead of renting it, and you keep the machine that built it. We have deliberately put no durations on these steps: the shape depends on how much of your material already exists in usable form, and we would rather scope that with you than advertise a number.
Load your organisation and its market into a private base
Strategy, operating model, systems, obligations, prior decisions, and the market around you. Internal and public, your side only. This is the step people underestimate and the step everything else rests on.
Produce the point of view on a decision that is already live
Take the decision in front of you now, the one you were about to put out to a firm. Studio produces the analysis and the paper, grounded and cited. You judge it against what you would have received.
Stand up agents for the work that repeats
Screening, licence and approval paths, benchmark refreshes, board-paper drafts, register upkeep. The recurring work stops being a new project every time it comes round.
The method is owned, so the next question starts from a running start
New market, new decision, new joiner. The base is already there, and so is the reasoning behind the last three answers.
Running today, on real data, in-market.
See it running on your own data.
Thirty minutes. We will ask what decision is in front of you and what you were about to spend on it, then tell you whether Studio helps. If it does not, we will say so on the call. No deck.
No commitment. One call, one honest answer.