Three architecture options
A managed model service from a cloud provider in an EU region, an open model on rented GPU in your account, or a hybrid where only sensitive queries stay internal. We cost each option using your figures.
Not every task needs a dedicated GPU. We pick the smallest option that meets your data protection and quality requirements.
A managed model service from a cloud provider in an EU region, an open model on rented GPU in your account, or a hybrid where only sensitive queries stay internal. We cost each option using your figures.
Text recognition for scans, splitting into meaningful sections, metadata such as date, matter or department, and regular refreshes as new documents arrive.
Anyone barred from a folder in SharePoint or on the file server gets no answers drawn from it in the AI search either. Sign-in runs through Entra ID or another identity provider.
A question set with known correct answers, compiled with your team. Every change to the model or the preparation pipeline is tested against it first.
No publicly reachable endpoints, encrypted storage, keys held in your vault at the cloud provider and logs recording who asked what and when.
Budget limits, alerts, automatic scale-down outside working hours and a monthly usage and cost report.
A data-flow description for your record of processing activities and the DPIA, plus an overview your legal adviser can use to judge what the EU AI Act means for the deployment.
A first environment covering part of your documents is usually ready within a few weeks. Expansion follows data volume and results.
Which questions should be answered, which data is involved, which constraints apply? Plus two or three costed options.
A defined collection, such as one practice area, is indexed and checked against the question set.
Further collections, permission inheritance and connection to Teams or the intranet.
Ongoing management by us, or transfer to your IT team with everything documented as code.
Access rights tend to vanish during indexing. Load every document into one shared search index and the AI will suddenly answer questions from salary lists or personnel files for anyone who asks. That is why we inherit permissions from the source systems and, before go-live, test deliberately with accounts that should not see certain folders.
That depends on model size and GPU. For searching internal documents, a few seconds per answer is normal. We measure it during the pilot with your real questions before you commit to an option.
We test it against your question set and switch only if it proves better or cheaper. Because the infrastructure is described as code, a switch stays manageable.
Yes, and we recommend it. One practice area, a few thousand documents and a managed model service in the EU let you see value and cost before larger sums are committed. You never have to buy hardware; capacity is rented.
You decide. The usual set-up is restricted access for IT staff for troubleshooting and a defined retention period. The arrangement should also be covered in the information given to employees.
The build is charged as a project or at €110 per hour plus VAT. For running costs you receive a monthly estimate with several usage scenarios in advance; cloud charges are paid directly to the provider.
Describe which documents should become searchable and what data protection and location requirements apply. We will suggest suitable options.
Your enquiry has arrived
Our reply reaches you within one working day. Outages that leave your staff unable to work are dealt with first.
We could not find that town. Try another spelling, or choose whichever provincial capital lies closest; as everything is handled remotely, you get the same service in all nine Austrian states.