V3Q
Contact us

V3Qtor AG/Systems/AI Governance

Sovereign. On-premise. Deterministic.

AI that an enterprise can rely on has three properties: it runs where the data lives, it gives the same answer to the same question, and every answer can be replayed. We build and operate AI systems with all three, in your racks or a sovereign cloud, with forward-deployed engineers who stay until the output is trusted.

Book a margin reviewRead: why 80% of AI projects fail
A server stack inside a sealed glass enclosure on a steel plinth, with a red boundary line marking the edge of the jurisdiction

Sovereign AI: the model, the data and the logs stay inside a boundary you control.

The numbers

Most AI projects fail. The reasons are known.

80%fail

More than 80% of AI projects fail, twice the rate of IT projects without AI1. The five root causes are misunderstood problems, missing data, technology over problem, missing infrastructure, and problems too hard for the method.

RAND Corporation, 2024

95%no P&L impact

About 95% of generative AI pilots show no measurable P&L impact; the failure sits in integration into workflows rather than in model quality2.

MIT NANDA, 2025

30%abandoned

Gartner's expected abandonment of GenAI projects after proof of concept by end-2025: weak data, insufficient risk controls, rising cost, unclear value3.

Gartner, 2024

What it reports

Every run has a cost, and the board shows it.

Cost per agentic run, per task completed and in total. GPU utilisation across the cluster. Requests per week and per month, for every workload. The log that makes a run replayable is the same log that prices it.

AI Governance Operations board

On-premise cluster · 8 GPUs30 days
Cost per agentic runCHF 0.84Median across workloads
Cost per task completedCHF 2.31−18% against the prior month
Total cost, 30 daysCHF 18,420Compute, storage, licences
GPU utilisation71%Peak 94%, weekday mornings
WorkloadsModelsEvaluationsRun logCosts
WorkloadRequests, weekRequests, monthCost per runCost per taskBudget
8D fault classification1,2845,310CHF 0.62CHF 1.90✓Within
Claim validation9,76041,200CHF 0.31CHF 0.44✓Within
Discharge letter draft2,1158,940CHF 1.12CHF 3.05Over
Maintenance work-order triage6402,710CHF 0.48CHF 2.60✓Within
Illustrative. Every run is logged with model version, inputs and cost, so the same record that makes it replayable makes it billable.Requests this week: 13,799 · this month: 58,160

Three properties

Sovereign, on-premise, deterministic. In that order.

  • 1
    SovereignData, models and logs stay in your jurisdiction, under your keys, with no dependency on a provider outside it. Residency is a setting you can show the auditor.
  • 2
    On-premiseDeployed in your data centre or a sovereign cloud, air-gapped where the plant or the regulator requires it. Capacity sized for the workload, models kept current under your change control.
  • 3
    DeterministicWhere a decision depends on the output, the output is reproducible: fixed decoding, pinned model versions, schema-constrained responses, validators before anything is written, a confidence floor below which a person decides.
  • 4
    ReplayableEvery run stores its inputs, model version, decoding parameters, validators and an output hash. A reviewer, an auditor or a court can run it again and get the same answer.
  • 5
    GovernedA model registry with owners, evaluations that run before every change, policies per task, and a kill switch. The evidence files into ShieldOS.

AI Governance Run log

On-premise · ZürichModel registry v31
Runs today48,210All replayable
Schema-valid outputs100%0 free-text decisions
Replay agreement100%Same input, same output
Escalated to a person1.3%Below confidence floor
RunsModelsEvaluationsPoliciesData residency
RunTaskModel · versionDecodingValidationOutcome
#48210938D fault classificationvision-ind · 2.4.1T=0 · seed 17Schema + lot rule✓Accepted
#4821094Claim validationclaims-lm · 1.9.0T=0 · seed 17Schema + eligibility✓Accepted
#4821095Discharge letter draftclin-lm · 3.1.2T=0 · seed 17Schema + clinicianEscalated
#4821096Rate recommendationrev-model · 5.0.3DeterministicBounds + replay✓Accepted
#4821097Concierge answer, DEguest-lm · 2.2.0T=0 · seed 17Citations required✓Accepted
Illustrative. Every run stores inputs, model version, decoding parameters, validators and the output hash.Data residency: CH · EU

How AI output becomes reliable

Reliability is engineered.

Language models are probabilistic. The systems around them do not have to be. Six measures make an AI output something an operation can rely on.

  • Fixed decoding

    Temperature at zero, fixed seeds, pinned model builds. The same prompt and the same context return the same tokens. Where hardware non-determinism remains, outputs are canonicalised and compared.

  • Schema-constrained outputs

    The model fills a structure, never free text. A fault class, a lot number, a claim decision with reason codes. Anything that does not fit the schema is rejected before it reaches a system of record.

  • Validators before writes

    Business rules check every output against the master data and the policy: a lot that exists, a tolerance that applies, a rate within bounds. The validator decides what gets written; the model only proposes.

  • Grounding with citations

    Answers that rely on documents carry the passage they came from. No citation, no answer. The concierge, the SOP assistant and the discharge draft all run this way.

  • Confidence floors and escalation

    Below a threshold the run goes to a named person with the evidence attached. The threshold is a policy, reviewed with the evaluations.

  • Evaluations and regression tests

    A fixed test set per task, run before every model or prompt change. A change that lowers agreement with the reference set does not ship. Results file into the audit trail.

The EU AI Act

What a high-risk system has to show, from August 2026.

Regulation (EU) 2024/1689 entered into force on 1 August 2024. Most obligations for high-risk systems apply from 2 August 2026, with fines up to €15 million or 3% of worldwide turnover for failing them4. The six measures above are how we meet the technical ones.

  • Art. 10
    Data governanceTraining, validation and test data with documented provenance, relevance and bias examination. The Master Data Management Layer is the record of it.
  • Art. 11
    Technical documentationKept up to date before the system is placed on the market, and for ten years after. Generated from the pipeline, not written afterwards.
  • Art. 12
    Record-keepingAutomatic logging of events over the system’s lifetime. Every run stores inputs, model version, decoding parameters and an output hash.
  • Art. 14
    Human oversightA named person who can understand, override and stop the system. Confidence floors route the run to them with the evidence attached.
  • Art. 15
    Accuracy, robustness, cybersecurityDeclared accuracy levels and resilience to errors and attacks. The evaluation set per task is the declared level, re-run before every change.

67%versus 22%

MIT NANDA found purchased or partnered AI deployments succeed about 67% of the time; internal builds about a third as often2. The difference is who sits in the operation while the system learns it.

MIT NANDA, 2025

4quarters

A forward-deployed team owns one P&L line for four quarters: from the first control or the first model to a number the CFO signs.

V3Qtor engagement model

Forward-deployed engineers

We deploy engineers, not decks.

The learning gap MIT names is closed in the plant, on the ward, at the claims desk. Our engineers sit with the operation, build against its data, run the evaluations with its people, and stay until the output is trusted and the number is in the P&L.

  • 1
    One P&L line, one team on site, four quartersScope small enough to measure, long enough to bank.
  • 2
    Built on your data foundationThe Master Data Management Layer first, the model second. RAND's second root cause does not get a chance.
  • 3
    Governed from day oneRegistry, evaluations, policies and ShieldOS evidence from the first run.

Talk to us about an AI project.

Tell us about one that is stuck in pilot or one you are about to start. We will look at where its data lives, whether its output can be reproduced, and who would own the number, and give you a view of what it would take to get it into the P&L.

Book a margin review
Notes and sources
  1. 1RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RRA2680-1, August 2024. More than 80% of AI projects fail, twice the rate of IT projects without AI; 65 experienced practitioners interviewed. Source
  2. 2MIT NANDA, The GenAI Divide: State of AI in Business 2025, August 2025. About 95% of generative AI pilots show no measurable P&L impact; purchased or partnered deployments succeed about 67% of the time, internal builds about a third as often. Source
  3. 3Gartner, August 2024: at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, citing data quality, risk controls, cost and unclear business value. Source
  4. 4Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal 12 July 2024. Entry into force 1 August 2024; prohibitions apply from 2 February 2025, general-purpose AI obligations from 2 August 2025, most high-risk obligations from 2 August 2026. Article 99: fines up to €35 million or 7% of worldwide turnover for prohibited practices, up to €15 million or 3% for other obligations. Source