Flowing Mind is a sovereign AI consultancy for UK defence, central government and regulated organisations. We design, build and govern AI on infrastructure you control, up to fully air-gapped, and we produce the evidence your accreditor will ask for.
Fixed fees, published. No proprietary runtime and no licence on work you paid us to build.
The problem
The pilot works and the board likes it. Then it reaches information governance, the DPO or a procurement panel, and the questions start. Where does the data go? Who trains on it? What happens when the vendor changes its terms? Can you show us the evidence?
Those are fair questions. They are also very hard to answer about a system built around someone else's API, and by that point the money has been spent.
We settle them in the first week, before any architecture is chosen.
What we do
Most clients start with the assessment and go no further until they are convinced. That is how it is meant to work. It should be cheap to find out whether we are the right fit.
A review of your data, systems, controls and people. You get a ranked list of AI opportunities with costs, risks and a build order, and a written note on what to leave alone.
Read more 02Open-weight models on your own hardware or in a UK region you control, up to fully air-gapped on NVIDIA DGX or HGX B300. Hosts hardened to DISA STIG and CIS Level 2.
Read more 03Retrieval over your own documents, agents for work that varies, and automation inside the systems you already run. Built in your repository and handed over.
Read more 04The policy, the risk register and the evidence a regulator, an auditor or a procurement panel will ask for. Built so that AI proposals get approved instead of parked.
Read moreDeployment tiers
The word gets used for everything from an air-gapped network to a hosted API with a US parent company and a promise. These are the three tiers we use with clients.
Your hardware, your network, your control plane.
Open-weight models on machines you own, up to a network with no internet egress at all. Every dependency is mirrored inside the boundary: images, packages, drivers, licensing, model weights. The strongest position, and the heaviest to operate.
Right for: defence, segregated government networks, patient records, privileged material.
A UK region, on a contract you hold.
Dedicated model instances in a UK region of a cloud you already hold a data processing agreement with. Data stays in the country and out of shared inference pools, at a fraction of the operational cost of running your own accelerators.
Right for: most regulated production workloads, where the constraint is residency and physical control is not required.
A frontier model under enterprise terms.
Sometimes the best model is a hosted one and the data is not sensitive. That is a reasonable choice if you have written zero-retention and no-training terms, a documented residency position, and an accurate classification of what goes in.
Right for: public information, marketing, internal drafting, low-classification workloads.
We recommend Tier 3 where it fits. Most engagements end with a mix of tiers, routed by data classification.
Who we work with
Inference on hardware inside your perimeter with no egress path. Hosts hardened to STIG and CIS Level 2, and the Secure by Design evidence your accreditor expects.
MOD Secure by Design, JSP 604 and Def Stan 05-138
Departmental AI that stands up to GovAssure, Secure by Design and ATRS, running on infrastructure the department controls.
GovAssure against the NCSC CAF, and Secure by Design
Model risk documentation, explainability and Consumer Duty evidence, with customer data kept inside a controlled environment.
SS1/23 model risk expectations and Consumer Duty evidence
Patient data cannot go to an offshore API. We build inside your boundary and produce the DPIA and information governance evidence a project needs to be approved.
DSPT, DTAC and information governance sign-off
Privilege and client confidentiality rule out most hosted AI. Private retrieval over your own matter files keeps the material inside the firm.
Legal professional privilege and SRA supervision duties
Councils, NDPBs and wider public services. Transparency duties and public scrutiny set a high evidence bar, so we build to the standards your assurance team already works to.
ATRS entries, AIME self-assessment and procurement scrutiny
Delivered work
Central government
An on-premise AI platform on NVIDIA DGX B300, running with no internet egress, hardened to DISA STIG and CIS Benchmarks Level 2, and mapped onto the NCSC Cyber Assessment Framework while it was being built.
Defence
An NVIDIA HGX B300 platform inside a defence perimeter with no egress path of any kind, delivered under MOD Secure by Design with a live security case, DISA STIG and CIS Level 2 hardening, and every artefact provenance-checked at the transfer boundary.
How we work
Every engagement starts by drawing the line your data must not cross. Everything we design sits inside it. Where a hosted model is the right call for a low-classification workload, we document the residency and retention terms and put them in the evidence pack.
No proprietary runtime of ours, no per-seat licence on middleware, nothing that stops working if we part company. Open weights and open standards where they do the job, and a repository with tests and documentation your own team can run.
An evaluation set built from your own material, measured results, a risk register entry and the documents an auditor or procurement panel will ask for. These are produced during the build, while the people who made the decisions can still explain them.
A fair share of proposed AI projects should not go ahead. Where the answer is a database query, a fixed rule or a better process, that is the recommendation, and it is written down with the reasoning so you can take it to your board.
Insights
Regulation
There is still no UK AI Act, but that does not mean no obligations. What binds UK organisations using AI in 2026, and what to put in place.
ReadSovereignty
A working definition of sovereign AI, the four questions that separate real data sovereignty from a residency claim, and where each deployment tier fits.
ReadBuying AI
Real UK market rates for AI consulting in 2026: day rates, fixed-fee audits, project costs and retainers, plus where the money usually gets wasted.
ReadCommon questions
One that designs AI systems around your data boundary first and chooses vendors second. In practice it means open-weight models on infrastructure you control, either your own servers or a UK cloud region on a contract you hold, so prompts, documents and outputs stay in your jurisdiction and are never used to train someone else’s model.
No. There are three workable positions: self-hosted on your own hardware, private deployment in a UK cloud region you control, or a hosted model under enterprise terms with UK residency and no training on your inputs. Each suits a different class of data. We work out which of your workloads belongs in which tier and write down the reasoning for audit.
Our readiness assessment starts at £4,500, fixed fee, over two to three weeks. Governance starts at £7,500, private deployment at £18,000 and implementation at £25,000. All four are on the pricing page with what is included.
Both. Advice that has never had to survive production tends to be wrong, and a build without governance tends to be blocked at the assurance gate. We assess, we build, we produce the evidence, and we hand the lot over to your team.
Yes. We have delivered air-gapped platforms on NVIDIA DGX B300 and HGX B300 for central government and defence. Every dependency is mirrored inside the boundary: container images pinned by digest, operating system and language packages, GPU drivers and firmware, licensing served locally, model weights imported as hashed artefacts, and vulnerability feeds brought in on a set cadence. The first deployment is the easy half. Keeping the platform current in year three is the half we design for first.
Hosts, container runtimes and orchestration are built from DISA STIG baselines reconciled against CIS Benchmarks Level 2, in code, and scanned before handover. Every deviation is documented with its compensating control. Platforms are mapped onto the NCSC Cyber Assessment Framework and the NCSC Guidelines for Secure AI System Development, and traced to NIST SP 800-53 and the NIST AI Risk Management Framework where a partner or supply chain expects that vocabulary.
No deck. Tell us what you are trying to do and we will tell you whether AI is the right tool, roughly what it would take, and roughly what it would cost.