Go back
How to run a SIEM RFP in Germany: an evaluation framework built around sovereignty, cost control and the BSI Act

How to run a SIEM RFP in Germany: an evaluation framework built around sovereignty, cost control and the BSI Act

Securevisio
17.11.2026

Many SIEM requests and tenders are built around a list of features: correlation, UEBA, SOAR, dashboards, an AI assistant. That leaves plenty of room for interpretation, and it is easy to lose sight of what really matters. A buyer who cares about digital sovereignty may ask on which servers in Europe a “sovereign” SIEM is deployed. That is a valid question, but what if the platform relies on dozens of third-party tools and services that send customer data outside Germany and Europe?

The questions that later decide whether a platform works for a German organisation are usually missing from the list: under which jurisdiction the provider operates and where the AI sends data, how the licence behaves when log volume doubles, who actually runs the service and with whom you sign, what the platform contributes to your duties under the BSI Act, and whether the works council will agree to behavioural analytics at all.

This guide turns those questions into an evaluation framework: nine criteria, each with the questions to put in the RFP, what a strong answer contains, the evidence to demand, and the answers that should worry you. It ends with a weighting model and a script for a proof of concept.

What exactly are you buying?

First, establish which kind of SIEM offering you need, and do it before you write a single requirement: the categories overlap, and vendors will answer the question they prefer.

  • A SIEM platform you operate yourself. You or your integration partner handle deployment, data collection, detection rules, SOAR playbooks and data retention. If you also want a SOC, you build it yourself: recruiting experts, setting up incident handling, running your own team.
  • A platform plus a managed service. The same technology, operated around the clock by a service provider. The partner buys the hardware to run the SIEM, deploys it, onboards the data and writes the detection rules. If the partner runs a SOC, it can also provide 24/7 monitoring together with analysis, response and other required services.
  • A platform plus leadership as a service. Everything in the previous option, plus a professional CISO service for organisations without a CISO: someone who owns direction, reporting and risk analysis.

Write the primary objective into the first page of the RFP. “Detection and response for a 24/7 operation, run by a partner, on-premises, with evidence for the BSI Act” produces very different answers from “central log retention”.

SecureVisio’s answer: a single platform with SIEM, SOAR, UEBA and CMDB in one engine, plus XDR, threat intelligence, vulnerability management, business impact analysis and risk analysis. Managed SIEM is provided by European partners; CISO as a Service and other Advanced Services are provided by SecureVisio itself. SecureVisio’s partners maintain their own detection rules written for SecureVisio, which are then adapted to each client’s environment.

Who has to be at the table before the RFP goes out?

In Germany, more people than in most markets. Bring them in before the requirements are fixed, not after the shortlist:

  • CISO or IT lead: the risks of each operating model, how far the platform can be adapted to your needs, platform capabilities (detection, analysis, correlation, integrations) and ongoing operation and maintenance.
  • Procurement: the TCO calculation, the licence model and the contract structure, especially if a service partner is involved.
  • Management: under §38 BSIG, the management bodies of essential and important entities must implement the risk-management measures and oversee their implementation. They will want to understand what the platform contributes and what remains their responsibility.
  • Works council: if behavioural analytics are in scope, co-determination is likely to come up (criterion 9).
  • Data protection officer: for what leaves the organisation, especially in AI steps.

A requirement the works council blocks in month four costs more than a requirement discussed in week one.

Criterion 1: who can legally reach your data, including through the AI?

Why it matters in Germany. Where data is stored (residency) and who can legally access it (sovereignty) are different questions. A data centre in Frankfurt answers the first and not the second if its operator is subject to a non-European jurisdiction. The US CLOUD Act, 18 U.S.C. § 2713, obliges providers to disclose, preserve and retain data in their “possession, custody, or control” “regardless of whether such communication, record, or other information is located within or outside of the United States”. With AI in the platform, there is a second path: the language model.

Questions to put in the RFP:

  1. Where is the provider headquartered, where are research and development, and where is the capital? Is there a parent company outside the EU?
  2. In each operating model you offer, where is data stored and processed?
  3. In the AI steps (summaries, triage, response selection), which model processes our data, where does it run, and who chose it?
  4. In the configuration you deliver, does any step of the analysis send data outside the EU? If a customer can choose a model outside the EU, how is that decision made and documented?
  5. Can the AI run entirely on our own infrastructure?
  6. What can be masked before data reaches a language model, and who sets the scope?
  7. Is any of our data used to train models for other customers?
  8. Besides storage and the AI, which components or services receive our data or metadata, for example threat-intelligence lookups, sandboxing, telemetry, licence checks, update servers or crash reports? List each one with its destination and jurisdiction.
  9. From which countries can your support staff, or a partner’s, access our environment? Is there follow-the-sun support from outside the EU?
  10. Is any subcontractor or component supplier subject to the US CLOUD Act or FISA Section 702?

A strong answer separates the provider’s jurisdiction, the data path per operating model and the data path of the AI. It says plainly what changes if you choose a different model.

Evidence to request: a data-flow description per operating model including the AI steps; the provider’s corporate structure; a list of all outbound connections the platform makes in the delivered configuration.

SecureVisio’s answer: no US parent company; headquarters, research and development, and capital in the EU; European support. In the configuration delivered in the EU, no step of the analysis sends data out of the EU. An AI model outside the EU is a deliberate decision of the customer, not the delivered state, and if a customer makes it, that changes the state: data processed by that model then leaves the EU. Running a model locally on the customer’s own infrastructure is the alternative. Before data is handed to a language model, IP addresses, hostnames and account names, among other data, can be masked, with the scope set by the customer’s administrator.

Criterion 2: can the platform run on-premises, air-gapped or only in the cloud?

Why it matters. Public authorities, operators of critical facilities and many mid-sized companies have hard constraints on where systems may run. An AI feature that only works in a vendor’s cloud may be unusable for them.

Questions to put in the RFP:

  1. Which operating models do you offer: on-premises, as a hosted service, air-gapped?
  2. Do any of them depend on public cloud services?
  3. Which functions work in an air-gapped environment, and specifically: does the AI work there, and with which model?
  4. For the hosted model: who is the operator, where is the data centre, and under which jurisdiction do both fall?
  5. How long does a typical production implementation take, and on what does that depend?
  6. Which log formats and integration protocols are supported as standard?
  7. Does the platform offer a built-in set of detection rules that can be freely customised?
  8. How do you prepare for the obligations of the Cyber Resilience Act (CRA)? Do you provide an SBOM, how long is the support period, and how are vulnerabilities disclosed?

Evidence to request: a reference architecture for your size; an implementation plan with dependencies.

SecureVisio’s answer: on-premises, as SIEMaaS or air-gapped, without dependence on public cloud services. On-premises keeps all data in the customer’s infrastructure. In air-gap operation the platform works without a connection to external servers; the AI agent then uses a local model server. SIEMaaS is operated by European partners in European data centres; for a specific offer, ask the partner about the operator and the data centre’s owner. A typical production implementation takes two to three months, depending on the complexity of the IT environment. Standard formats include syslog, CEF and LEEF, and REST API as an integration protocol. A free proof of concept is offered in the customer’s own IT environment.

Criterion 3: what happens to the SIEM bill when your data grows?

Why it matters. When a licence is priced by data volume, every new log source becomes a budget discussion, and logging decisions start to follow cost rather than risk. Under §30 (1) sentence 2 BSIG, the cost of implementation is one of the factors in assessing whether measures are proportionate, so a predictable cost model is not just a procurement preference.

Questions to put in the RFP:

  1. What is the licence metric: data volume, events per second, assets, users?
  2. Are there limits on EPS, log volume or concurrent operators? What happens when we exceed them?
  3. Which modules are licensed separately? Is UEBA an extra licence?
  4. Which assets count, and which do not (network devices, OT elements)?
  5. Can we store more data than we correlate, and is retention independent of correlation?
  6. Does the inventory show assets that are not licensed?
  7. Along which dimensions is a managed service priced?
  8. Are there additional costs for custom detection rules or for onboarding new data sources?
  9. Is the annual price increase capped, and which index is it based on?
  10. For the on-premises model: which hardware and storage do we need for our growth scenario?
  11. For public-sector buyers: do you accept EVB-IT contract terms (e.g. EVB-IT System or EVB-IT Cloud)?
  12. What do exit and data export cost at the end of the contract?

A strong answer lets you build a three-year cost model with your own growth assumptions. Ask every vendor to fill in the same model: today’s assets and log sources, plus 30 % more sources in year two. The number is yours to calculate; the vendor’s job is to make the metric transparent.

SecureVisio’s answer: the SIEM licence has no limits on EPS or log volume and is based on the size of the environment, specifically the number of assets, not on the volume of data processed. Network devices are not counted as assets; OT elements are counted per element. There is no limit on concurrent operators. UEBA needs no separate licence. All modules are sold as one licence; SIEM and SOAR are also available on their own. Log management is a separate layer: logs are stored raw and normalised, with independent retention. A managed SIEM is priced by number of assets (not ingestion), scope of the SOC service, SLA, 24/7/365 operation, contract term and functional scope.

Criterion 4: does the data your SIEM needs for detection and analysis actually arrive?

Why it matters. Detection is only as good as collection and parsing. Without a parser there are no fields; without fields there is no detection.

Questions to put in the RFP:

  1. Which collection paths do you support, and which do you recommend for Windows endpoints and servers?
  2. If we decline an endpoint agent, how are Windows logs collected?
  3. Can the endpoint agent provide extended telemetry, for example via Sysmon?
  4. How are parsers provided, built and maintained? What happens when no parser exists for a source?
  5. How long does onboarding a new data source typically take?
  6. How do OT and industrial systems get into the platform?
  7. How many detection rules are available from day one, and how are they mapped?
  8. How does the platform detect data lost in transit, for example dropped UDP syslog messages, gaps in sequence numbers or delayed timestamps?
  9. For OT and critical facilities: which OT security vendors have a tested integration, and which protocols (e.g. Modbus, PROFINET, IEC 60870-5-104) reach the platform through them?

Evidence to request: your three most awkward log sources, parsed live in the PoC.

SecureVisio’s answer: three collection paths in a fixed order of preference. Syslog listening is the default (UDP 514 as standard); the XDR agent is recommended for Windows workstations and servers and can provide extended telemetry via Sysmon; if a customer declines an agent, Windows logs are collected through a connector using Windows Event Forwarding; active connectors (API, ODBC, file reading) are used where the first two are not enough. Parsers map raw text into the platform’s own schema; the AI assistant can create a parser from sample data and update existing ones. Connecting a new data source takes days, at most weeks. Over 500 ready-to-use detection rules, mapped to MITRE ATT&CK, are active from day one of implementation.

Criterion 5: does a SIEM alert arrive with its business context?

Why it matters. Two identical alerts on two different assets are not equally urgent. A queue sorted by rule severity treats a test laptop and the payroll server alike. The same applies to vulnerabilities: a CVSS score describes the vulnerability, not what it means on a particular system.

Questions to put in the RFP:

  1. Where does the asset and process context come from, and how is the inventory filled?
  2. Is context attached to every incident automatically, or only via playbooks?
  3. How is an incident’s priority set, and does it take the importance of the asset into account?
  4. How are vulnerabilities prioritised beyond CVSS? Can rules lower a priority as well as raise it?
  5. How is behavioural analytics built: which algorithms learn, from what, and which numbers are defaults?
  6. Does the platform show why a rule fired?
  7. Can the criticality of KRITIS-relevant assets and processes be flagged separately and used in prioritisation?

Evidence to request: in the PoC, the same alert on two assets of different importance, and two vulnerabilities whose order should differ from their CVSS order.

SecureVisio’s answer: the CMDB is filled through manual entry, templates, detection rules, vulnerability scans, firewall data and XDR detection, with discovery switched on or off per asset; it guarantees no complete inventory, and context is only as good as the CMDB. When an incident is created, CMDB and Active Directory context is written into its history automatically, without a playbook. Priority rules set the incident’s priority, based on the importance of the asset, and prioritisation draws on the built-in risk analysis. Vulnerability priority combines the scanner’s severity, an internal priority from CMDB, business impact analysis and risk analysis, and configurable rules that raise priority (exposure, missing WAF, indicator of compromise) or lower it (isolation, compensating controls, low data criticality); live log data can correct the priority. UEBA uses six documented algorithm families; parameters such as the 30-day learning period of the built-in limiters are defaults and configurable. The incident form explains why a rule fired. More at CMDB and BIA, risk management, vulnerability management and UEBA.

Criterion 6: what does the AI in the SIEM do, and where does it stop?

Why it matters. “AI-powered” is on every datasheet. For an RFP, the label is of little use unless the vendor states which kind of model does what, within which limits, and with what record.

Questions to put in the RFP:

  1. Which AI roles exist in the platform, and in which order do they act on an event?
  2. Does an AI agent act on incidents? Can it close incidents or execute responses on its own?
  3. How is its autonomy configured, and what limits it?
  4. Can it invent actions, or does it choose from a defined list?
  5. Is every step the agent takes recorded and reproducible?
  6. What happens if the model, or another required service, is unavailable?
  7. Does the AI decide whether an incident has to be reported?
  8. What measured effect do you claim, and how was it measured?

A strong answer to question 8 is either a number with its method or a plain “we have no measurements”. A number without method is a red flag.

SecureVisio’s answer: machine learning filters the raw event stream first; Deep Value Learning, the central innovation in version 6.0, learns the value of concrete behaviours and processes; language models summarise, explain and generate queries; reasoning models are used selectively, after filtering. The AI agent picks up each new incident by itself, works within the limits of its playbook, classifies it, and either closes it with its reasoning or hands the analyst a reasoned verdict. It chooses responses from a predefined list. Whether it acts on its own or asks for approval is configured per case, and the asset’s context limits what it may do. Every agent step is stored and reproducible. If the language model is unavailable, playbooks have a fallback per step; if a required service is unavailable, processing stops. The classification is a working assessment, not a reporting decision. SecureVisio has no internal benchmarks on detection or response times and says so. Details: SecureVisio SOAR.

Criterion 7: how does it support your duties under the BSI Act?

Why it matters. No tool is “NIS2-compliant”. Entities comply; tools and services support them. An RFP should ask what the platform contributes to each duty and where it stops.

Questions to put in the RFP:

  1. How does the platform support documentation under §30 (1) sentence 3 BSIG?
  2. For operators of critical facilities: how does it support the attack-detection duty under §31 (2) BSIG?
  3. How does it support reporting under §32 BSIG (early warning within 24 hours, notification within 72 hours, final report one month after the notification)? Does it provide ready-made BSI reporting templates?
  4. What can you provide for the evidence operators of critical facilities owe under §39 BSIG?
  5. Which parts are product functions, and which are services, and provided by whom?
  6. Can you map platform functions to the risk-management measures in §30 (2) BSIG and state for each whether the product, a service or the entity itself covers it?
  7. Can incident data be exported in a form that helps fill in a report to the BSI, so that the early warning within 24 hours does not start from a blank form?
  8. For financial-sector customers: how does the platform support DORA (classification of ICT-related incidents and reporting to BaFin)?

Red flag: any vendor that calls its product “NIS2-compliant” or promises to report for you.

SecureVisio’s answer: the platform supports the fulfilment of the attack-detection duty with continuous real-time monitoring, automatic correlation and anomaly detection; whether an implementation suffices is decided in the evidence process, not by the tool. The incident history, in which every agent, operator and system step is recorded, serves as the evidence trail; incident data can be carried over into the reporting workflow. There are no ready-made BSI reporting templates and no ready-made audit pack for §39. The entity reports; reporting workflow, templates and coordination with authorities are part of SecureVisio’s Advanced Services.

Criterion 8: who operates it, and who is your contractual partner?

Why it matters. With a managed service, the vendor of the software and the provider of the service are often different companies. That changes who you sign with, who is liable under the contract, and whom you ask about the data centre.

Questions to put in the RFP:

  1. Who operates the managed service, and with whom do we sign the contract?
  2. Which operating models are possible: self-operated, co-managed, fully managed?
  3. What does the SLA cover, is it 24/7, and how is escalation structured?
  4. What does the service cover in incident response?
  5. If we have no CISO: who can take over that role’s tasks, and what stays with management?
  6. Which certifications does the provider hold, and which legal entity and which service do they cover?

SecureVisio’s answer: managed SIEM is provided by European partners, who are your contractual partners. Two models: the partner operates, or you operate with partner support (parsers and rules stay with the partner). The service covers the full incident response framework, 24/7, with an SLA tailored to you and an L1 to L3 escalation path. No minimum size; a small installation is introduced in two to six weeks. CISO as a Service is provided by SecureVisio itself, as a configurable package; legal responsibility stays with management under §38 BSIG. SecureVisio is ISO 27001 and ISO 9001 certified as a company. Partners: partner network; services: services.

Criterion 9: will the works council and the data protection officer sign off?

Why it matters in Germany. Under §87 (1) No. 6 of the Works Constitution Act (BetrVG), the works council has a right of co-determination over the “introduction and use of technical equipment designed to monitor the behaviour or performance of employees”. If no agreement is reached, the conciliation committee decides (§87 (2) BetrVG). Behavioural analytics that profile users are the obvious candidate for that discussion. Plan for it in the RFP, not after signing.

Questions to put in the RFP:

  1. Which user-related data does behavioural analytics process, and per which profiles?
  2. Can profiles or groups be excluded from machine learning?
  3. How is visibility of events and incidents restricted to the teams responsible?
  4. What can be masked before data reaches a language model?
  5. Can you provide documentation that the works council and the data protection officer can review?

SecureVisio’s answer: behavioural analytics learn per user or host profile; machine learning can be included or excluded per profile, and an inactive profile is used neither by UEBA nor by rules nor by machine learning. Access to a logical map is granted per responsible team, and those with access see the events, incidents and vulnerabilities linked to its assets. IP addresses, hostnames and account names, among other data, can be masked before a language model sees them. Whether a specific deployment requires an agreement with the works council is for the organisation and its works council to determine.

How should you weight the criteria?

There is no universal weighting; your constraints decide. As a starting point that reflects what German buyers most often treat as decisive:

CriterionSuggested weight
1. Jurisdiction and data path, including AI15 %
2. Operating models10 %
3. Cost control and licence metric15 %
4. Collection and parsing10 %
5. Context and risk-based prioritisation15 %
6. AI: roles, limits, record10 %
7. Support for BSI Act duties10 %
8. Operation and contractual partner10 %
9. Works council and data protection5 %

Two rules make the scoring honest. Score each criterion only on evidence you saw, not on answers you read. And treat some answers as knock-out criteria regardless of score: for an operator that must run air-gapped, criterion 2 is not 10 %, it is pass or fail.

How do you run a proof of concept that actually decides something?

Vendors are good at demos. A PoC should be run on your data, with the same script for every vendor:

  1. Same scope for everyone. The same log sources, the same assets, the same time frame.
  2. Your awkward sources. Include the three log sources nobody wants to touch.
  3. One incident, end to end. Inject a scenario (for example a logon from an unusual network on a server with a critical business process) and follow it from log line to closure. Note at each stage which question gets answered, and by whom.
  4. The same alert on two assets. Check whether priority and permitted response differ.
  5. Two vulnerabilities. Check whether the order follows exposure and context, not just CVSS.
  6. Pull the plug on the AI. Make the model unavailable and see what happens.
  7. The evidence test. Three weeks after the PoC incident, ask for the record of who decided what, and why.
  8. The bill test. Ask each vendor to price your growth scenario in the same template.

Measure what you can observe yourself, such as the time until the incident shows the right context or how many steps the analyst needs. Do not accept vendor-supplied effect numbers as a substitute.

Run the framework against your own environment:
Request a demo

Table of contents


Please contact us if you have any questions.

Learn more about SecureVisio and the benefits it offers.
Poland
Poland
+48 17 779 6246
Germany
Germany
+49 4186‑895991‑0

Fill out the form to contact us