Your AI should be yours.

No, and we would usually advise against that. Most companies end up hybrid: sensitive workloads served by models running inside their own environment, everything else routed to frontier APIs under zero-retention terms with sensitive fields redacted before the call. Which workloads sit on which side is a decision for your risk owner. We build the boundary, and we make it visible enough to be audited.

Not if it is done with evaluations, which is why we will not ship a routing change without them. Before anything moves, we build an evaluation set from your own real cases and score the current setup on it. A cheaper route ships only when it holds quality on that set. Where it does not, the traffic stays where it is and we say so.

You do, from the first commit. Work happens in your repository under your accounts, model weights and inference run in your environment where the architecture calls for it, and your provider keys stay yours. At handover you receive the code, the runbooks, the evaluation suite and the operating knowledge required to run and change the system without us.

No. That is the normal starting position, and it is what the first two days are for. Provider invoices arrive as one number with no breakdown by team, feature or workload, so almost nobody can answer this from a dashboard. You do not need to arrive with the answer. You need someone who can open doors across departments, because the spend is rarely all in one place.

The honest answer is that we do not know until we have metered you. Savings depend on the traffic mix, model choices, prompt size, cacheability and the quality threshold each workload must hold. Some workloads are already on the right model and we will tell you to leave them alone. The only number we will stand behind is the before-and-after on your own invoice, on the same traffic.

For a large share of production traffic, yes: classification, extraction, summarisation, routing, retrieval-grounded answers over your own documents. For the hardest reasoning, frontier models still win. That is the whole point of routing: the question is never which single model is best, it is which model is sufficient for each request. We settle it with evaluations on your cases rather than benchmarks on someone else's.

We assess your AI paths against the OWASP GenAI LLM Top 10 2026. You receive a versioned findings register with evidence, exposure, remediation priority and retest status for each applicable risk. It is an engineering assessment, not a penetration test or certification, and the report says so explicitly.

No. We are not a router vendor and we take no margin on what you spend. We will happily deploy an open-source gateway or an existing commercial one if it is the right answer for you, in your cloud, under your accounts. Nothing we build depends on Zylen staying in the picture, and none of your traffic passes through anything we own.

Then we tell you that, and you still have a metered baseline and a versioned OWASP findings register. A young AI footprint on a sensible model with little repeat traffic may have nothing meaningful to recover yet. The engagement is staged so the cheapest step comes first precisely so that finding this out costs you ten days rather than a sprint.