Enterprise
Dependence is cheap to acquire and expensive to leave, and the price is only set once you cannot.
The operating reality #
Every enterprise AI decision taken this year is a decision about who owns the capability in five. The commercial pattern behind that is well established and it is not accidental. Adoption is subsidised while it is optional. Integration deepens until it is not. The data model, the identity estate, the tooling and the skills all take the shape of one supplier's interfaces, and only then does anyone find out what it costs, because only then is leaving quoted as a rebuild.
Jurisdiction is the second exposure and the one least often priced. A platform operated under another country's law can be compelled, restricted or reprioritised by a government you did not elect and cannot lobby. That is not a question about privacy policy. It is a question about whether your operations continue unchanged when the relationship between two states does not.
The third is the consultancy model, where capability is rented rather than transferred. An engagement that ends with the supplier holding the knowledge, the code and the ability to operate the thing has not built you an asset. It has sold you a subscription that was presented as a project.
Agentycs is the opposite trade. The platform installs into a boundary you control, every interface it presents is one you could build against elsewhere, and the software factory hands delivery capacity to your team instead of metering it back to you. You can leave. That is what makes staying a decision rather than a condition.
Nobody sets the price while you still have somewhere else to go.
What that requires of the platform #
- The whole stack — lakehouse, model serving, application runtime, software factory and the servers under them — installs into a boundary you control.
- Models are imported once into your own hub and served from your own accelerators, so inference carries no runtime dependency on an external provider's price list or current line-up.
- The interfaces you build against are standards: PostgreSQL wire protocol, Arrow Flight SQL, S3-compatible storage, OIDC, SCIM, MCP and Kubernetes. That is what makes leaving possible rather than contractual.
- Federate the identity provider you already run and keep membership current through directory synchronisation, so adopting the platform does not fork your identity estate.
- An autonomous software factory turns a small team's intent into shipped, tested services, so delivery capacity is transferred to you rather than rented back indefinitely.
- Per-tenant policy decides which models may serve a request and whether data may leave the cluster, enforced before any work is dispatched.
Where the argument is written out in full #
Financial services carries the sharpest version of every constraint on this page, so it has a route of its own with the detail behind it.
Financial services
Provenance and containment as platform properties, so a number that reaches a decision can be traced to the query, the data and the identity behind it.
The same requirements, different vocabulary #
These verticals arrive with their own language for the same constraints. We have not written a page for each of them and will not until we can say something specific, so here is the short version instead.
Energy and critical infrastructure
Control rooms, substations and remote assets sit behind operational-technology boundaries and often at the end of a constrained link. Inference runs at the site rather than in a central region, a zero-trust overlay carries what has to move between them, and forecasting and anomaly detection work over decades of operational history without any of it leaving your estate.
Healthcare and life sciences
Patient records and trial data cannot become a training input for somebody else's model, and consent is specific about purpose. A per-tenant lakehouse with row-level security keeps cohorts separated, models are served privately on your own accelerators, and layout-aware extraction turns clinical documents and scanned records into data you can actually query.
Industrial and manufacturing
A plant generates more event and image data than it is economic to ship anywhere. A high-throughput ingestion path absorbs it locally, inference runs on the line where a decision has milliseconds rather than seconds, and durable workflows carry a process across a shift change or a restart without losing where they were.
Legal and professional services
The work product is documents and the duty is confidentiality, which makes a general-purpose assistant a professional risk. Document intelligence turns matter files into structured, searchable data, retrieval cites its sources so an answer can be checked, and client and matter separation is enforced by the platform rather than by convention.
Telecommunications
Thousands of sites, a distributed estate and network telemetry at a volume that punishes any architecture built on central collection. Multi-site deployment puts compute next to the data, geo-distributed storage keeps it durable across regions, and the mesh brings its own routing and internal naming rather than depending on an external edge provider.
Research and education
Institutions need to collaborate without pooling data or handing custody to a commercial model provider. Per-tenant isolation lets separate groups share one platform safely, federated identity means people sign in with the account they already have, and your own model hub keeps open weights available and reproducible for as long as the work needs them.
Read next #
Sovereign AI platform
Own the whole stack, from the accelerators up to the applications.
AI software delivery
How intent becomes shipped, tested and observed software without a hiring round.
Where it can run
The UK sovereign cloud, and the on-premises model everywhere else.
Deployment patterns
Every pattern compared, including who operates what in each one.
Start with what you cannot currently leave
Name the platform, the contract or the supplier relationship you would struggle to walk away from, and what it would cost you to try. We will show you what the same capability looks like when you own it.
Agentycs is pre-revenue and signing its first deals with its first customers. We are collecting the record as we go, and we will not dress a design decision up as traction.