FAST LTA: AI Agents Run in Your Own Data Center
FAST LTA CEO Samir Fadlallah answers six questions about agentic AI on the customer's own appliance: connections, permissions, tests and recovery.
Interview
According to FAST LTA, Silent AI runs as a 4U appliance entirely in the customer’s own data center. CEO Samir Fadlallah answers six questions on external connections, permissions, acceptance tests, running costs and recovery after an attack on the Active Directory.
Key takeaways
- Costs stay predictable. Hardware maintenance costs 8 percent of the list price per year, software maintenance 16 percent, with a minimum term of three years.
- Data and access stay in-house. Connections to source and line-of-business systems run on the local network; external connections are available only with the customer’s explicit approval.
- Recovery is prepared. Veeam Instant Recovery restores a 120 GB VM from Silent Brick Pro in a calculated 30 seconds or so; at 1 GB/s it takes around two minutes.
Portrait photo: FAST LTA
Related:How many coding agents can a review team handle?
No external connections
Question 01
Which connections to external services does Silent AI Pro need for operation, updates and remote maintenance?
Silent AI needs no connection to the outside and can be operated entirely without integration into other systems. Where systems are connected, by default this happens only within the customer’s network.
In Silent AI ES, these connections serve to load data: users from the directory service (e.g. Active Directory) and content from data sources such as file servers, Exchange, Confluence or SQL databases. The connections are not needed for ongoing operation.
Silent AI Pro also embeds the AI in workflows. Using skills and tools, it completes tasks such as creating documents. Via the Model Context Protocol (MCP), Silent AI Pro connects to line-of-business systems such as ERP, CRM or ticketing systems, where agents carry out actions approved by the user. These connections also stay on the local network. Connections to the outside, such as for a web search, are optional and possible only if the customer explicitly approves them.
What is the Model Context Protocol? The Model Context Protocol (MCP) is an open standard that AI applications use to access tools and data sources. Anthropic introduced it in November 2024. An MCP server provides the functions of a line-of-business system, such as creating a ticket. The agent calls them with the permissions of the account through which the connection runs.
Customers receive updates through the administration interface, or as an encrypted file for offline installation if they have no internet connection. Remote maintenance takes place only if the customer gives us remote access.
What happens when permissions are revoked
Question 02
How long does it take in Silent AI Pro until revoking a permission in the source system also blocks access through the AI?
In Silent AI ES, access permissions are inherited from the source systems. Anyone who is not allowed to read a file or piece of information will not see the corresponding content in any AI answer, nor in shared knowledge collections. If a permission is revoked there, this takes effect in Silent AI from the next sync of the respective connector. The interval for this sync can be configured in the system per connector; in addition, synchronization can be triggered manually at any time when needed. In Silent AI Pro, users can create agents and share them with entire user groups. What an agent is allowed to do in a line-of-business system is initially determined by the target system itself: the agent has at most the permissions of the account through which the MCP connection runs.
Anyone who is not allowed to read a file or piece of information will not see the corresponding content in any AI answer, nor in shared knowledge collections.
The acceptance test with your own documents
Question 03
Which concrete acceptance test should a mid-sized company use to verify the factual accuracy of the source-based answers from Silent AI ES?
For source-based answers in Silent AI ES, we recommend testing with your own documents rather than demo data. The business departments draw up 20–30 questions in advance whose correct answer and location they know, deliberately including tables, forms, counting questions over long lists and questions the documents do not answer. Each answer is assessed against three criteria: is it factually correct, does the stated source check out, and does the system honestly say when it finds nothing instead of inventing something? A second pass with a test user with restricted permissions shows whether confidential content really stays hidden.
Fixed maintenance fees instead of per-token costs
Question 04
Which running costs must a mid-sized company calculate for Silent AI Pro over five years in addition to the purchase price?
Silent AI has no usage-based costs: no fees per user, per query or per token. That holds even when agents complete many work steps in the background. Annual maintenance should be budgeted as a fixed running cost: hardware accounts for 8 percent and software for 16 percent of the list price of the respective items in the quote. The minimum contract term is three years. To get started with Silent AI ES (L1000), that is around 5,750 euros per year; for Silent AI Pro in the entry configuration R1000 (128 GB), around 7,350 euros per year. The entry hardware is identically equipped in both cases; the difference lies in the software: Pro adds agents, projects and optional integration with line-of-business systems via MCP on top of the enterprise search functionality. Anyone who does not want to finance the purchase all at once can now also obtain Silent AI Pro as a subscription. Purchase and running costs are then spread across fixed, predictable installments.
Silent AI ES
Search across your own data sources with source citations; the permissions of the source systems continue to apply. Entry configuration L1000.
Silent AI Pro
Additionally agents, projects and integration with line-of-business systems via MCP. Entry configuration R1000 with 128 GB, also available as a subscription.
Recovery after an attack on Active Directory
Question 05
Which recovery time for a business-critical system can be demonstrated with Silent Bricks with a compromised Active Directory, based on a documented test?
The storage itself barely limits the recovery time. With Veeam Instant Recovery, a virtual machine starts directly from the backup on the Silent Bricks, within seconds. For a full copy-back of a 120 GB VM, the calculation yields around 30 seconds on the Silent Brick Pro and, at about 1 GB/s, around two minutes on the Silent Brick Max.
The storage itself barely limits the recovery time.
In an attack on the AD, however, other steps usually determine the total duration: setting up a clean environment with a new directory service and backup server, and above all the forensics. Before systems come back, it must be clear how the attack unfolded and which backup version is still free of attack vectors; if necessary, older versions must be restored. For this, Silent Bricks combine immutable data at the software level, which stays online for forensic analysis, with immutability at the hardware level as a last line of defense, plus local accounts without AD integration and multi-factor authentication. That is why we recommend running through this scenario once a year with your own backup software and documenting the measured time.
How a deletion can be proven
Question 06
How can you prove, with Silent Cubes, that a document was deleted after its retention period expired, including any existing replicas?
Deletion on the Silent Cubes system is usually triggered by the archiving application, such as the document management system (DMS) or SAP, once the retention period of a document has expired; before that, the system rejects any deletion. Since this is a true hardware WORM system, the storage space on the storage unit remains occupied even after deletion.
The system checks whether the period has expired against the Compliance Clock in the storage units, rather than against the system time of the Head Unit, which could be manipulated. With the deletion, the system irrevocably removes the reference to the document; afterwards neither the archiving application nor an administrator can read it. In a replicated system, this applies to both sites.
The deletion log of the archiving application serves as proof, together with the archive reports of the Silent Cubes system, which record every deletion with a timestamp. According to FAST LTA, the Silent Cubes system, including retention periods and deletion after expiry, has been audited by KPMG.
About the interviewee
Samir Fadlallah has been CEO of FAST LTA in Munich since 2025. Before that, he was Chief Information Officer and Chief Technology Officer at Axel Springer, as well as Head of Global Generative AI. As CTO News Media Germany, he led product development of around 60 digital offerings for 25 media brands. As CIO, he was responsible for internal IT covering infrastructure, cybersecurity and cloud platforms. At FAST LTA, he is expected to continue the company’s growth trajectory, with a focus on European expansion and local, sovereign AI systems.
Frequently asked questions
Does Silent AI Pro need external connections?
Silent AI can be operated without integration into other systems. Connections to the outside, such as for a web search, are optional and possible only if the customer explicitly approves them. Updates are available via the administration interface or as an encrypted file for offline installation.
What does it cost to run the appliance?
Hardware maintenance is 8 percent of the list price per year, software maintenance 16 percent. The minimum contract term is three years. To get started with Silent AI ES (L1000), costs are around 5,750 euros per year; Silent AI Pro (R1000, 128 GB) costs around 7,350 euros per year.
How can a deletion on Silent Cubes be proven?
Every deletion generates a deletion log and is documented in archive reports. According to FAST LTA, the deletion process has been audited by KPMG.
Editor's Picks
cloudmagazinWhen AI agents travel: data residency as an operating problemcloudmagazinHow many coding agents can a review team handle?cloudmagazinProxmox move: the way back to VMware stays unbuiltMore from the MBF Media Network
MyBusinessFutureOpenAI agents bypassed their own monitoringDigital ChiefsAI intervenes in processes: Who is responsible here?SecurityTodayAI agents need limited mandatesImage source: AI-generated (November 2026)
Translated from the German original using artificial intelligence. The German version is authoritative.


