Sally - AI Meeting Assistant

NOVEMBER 2026

On-Premise vs. Cloud for AI Meeting Assistants: A Decision Guide for IT and Data Protection

Data protection wants on-premise, IT dreads the operations, management wants a decision. This guide shows when a German cloud is enough, when running it yourself is mandatory and when it is overkill.

IT manager and data protection officer discussing self-hosting next to a server rack in a server room

Whether an AI meeting assistant should run on-premise or in the cloud is rarely a GDPR question. It is a question of requirements, risk and your ability to operate it. For most companies, a cloud hosted in Germany with a data processing agreement and no third-country transfer is enough. Running it yourself becomes mandatory or sensible where a policy, a regulator, a client contract or network separation requires that meeting content never leaves your own network.

Tuesday morning, a meeting room on the third floor. Management wants an AI meeting assistant and the budget is approved. The head of IT has compared three vendors, and the data protection officer has one condition: "On-premise only, or I won't sign off." The head of IT does the maths in his head. Servers, compute, updates, on-call duty. And who fixes it when the system is down at eight in the morning before the board meeting? Both have a point. And both are talking about different things.

This article is the decision guide for exactly that table: IT, data protection and management. It explains what on-premise vs. cloud for an AI meeting assistant actually means, which rules require what, when running it yourself is mandatory and when it is overkill. Why server location matters in the first place is covered in the comparison German servers vs. US cloud. This article goes one step further: whether you should run the software yourself.

What on-premise really means for an AI meeting assistant

Many discussions fail because everyone says "cloud" but means something different. Before you compare vendors, it pays to agree on a shared vocabulary for the operating models.

Four operating models that are often confused

  • US cloud: The vendor or its parent company is based in the US. Even if the servers are in Frankfurt, it is subject to US law.
  • Cloud in Germany: A vendor without US group ties runs the software in German data centres. You use it like any other SaaS application.
  • Hybrid model: The application runs at the vendor, while the content (recordings, transcripts, analyses) sits on your servers.
  • On-premise: The complete software runs in your data centre or private cloud, including the processing. You provide hardware, platform and operations.

The decisive difference between the first three models and true on-premise: with on-premise, it is not only the storage location that changes but the operational responsibility.

Which building blocks of an AI meeting assistant run where

An AI meeting assistant is not a single application but a chain of recording, speech recognition, AI analysis, storage and integrations. A bot joins the online meeting or an app records in the room. Speech recognition turns audio into text. A language model creates the summary, decisions and tasks. A database and object storage keep everything, and connectors send the results to calendars, CRM or project tools.

Every link matters for the hosting question. A vendor that stores transcripts in Germany but calls a US service's language model for the summary still processes the content there. So do not just ask where the data is stored, ask where every single processing step runs.

What running it yourself means for your IT team

On-premise means the software comes from the vendor and everything underneath comes from you. That usually includes a container platform, compute capacity for speech recognition and AI, storage, backup, monitoring and a directory service for single sign-on. Speech recognition and language models are compute-intensive. Whether you need your own graphics processors (GPUs), and how many, depends on the model, the languages and your meeting volume. Serious vendors give you figures for your environment, not generic ones.

Then there is the ongoing work: testing and applying updates, patching operating system vulnerabilities, adding capacity, fixing incidents. Choosing on-premise means taking on part of what a cloud subscription already covers. That is not an argument against it, but it belongs in the decision before the contract is signed.

What the GDPR and the CLOUD Act actually require

The GDPR does not prescribe a specific operating model. It requires processing to be lawful, secure and traceable. For meeting data, three areas of regulation determine what that means in practice.

Art. 28 GDPR: data processing agreements in almost every model

If a service provider processes personal data on your behalf, you need a data processing agreement under Art. 28 GDPR. That applies to every cloud, including a German one. In many cases it also applies to on-premise: as soon as the vendor can access the system for support or updates, it generally acts as a processor again. On-premise does not automatically spare you the agreement, it only changes what the agreement has to cover.

Art. 32 GDPR: security based on risk, not location

Art. 32 GDPR requires technical and organisational measures appropriate to the risk, taking into account the state of the art and the costs. A professionally run German data centre with certifications can meet that requirement better than a server nobody patches in day-to-day business. The reverse is also true: if you run on-premise, responsibility for encryption, access control, logging and recovery sits with you. Running it yourself is only as secure as the IT team running it.

Third-country transfers under Art. 44 et seq. and the CLOUD Act

Art. 44 et seq. GDPR only allow transfers to third countries under certain conditions, for example on the basis of an adequacy decision or standard contractual clauses. On top of that comes the US CLOUD Act: it obliges US providers to hand over data when ordered by the authorities, even if that data is stored in Europe. This is the real difference between a US cloud and a German cloud. A vendor without US ties whose entire processing chain is in Germany solves the third-country problem without in-house operations. You do not need on-premise for that.

When a cloud in Germany is enough

For the vast majority of companies, the honest answer is: a cloud in Germany with a clean contract is enough. That applies to sales calls, project meetings, team rounds and client appointments, and to many conversations with confidential content as well.

Companies without special requirements

If no internal policy, regulator or client contract requires data to stay inside your own network, there is no legal obligation to go on-premise. Typical examples are mid-sized companies, agencies, consultancies and sales organisations. Their risk lies less in the operating model than in a vendor that transfers data to the US or uses it for AI training. The effort of running it yourself would be out of all proportion to the security gained.

What to demand from a cloud vendor

A German cloud is only as good as its details. At a minimum, check the Art. 28 agreement, the full list of sub-processors with locations, the exclusion of AI training on your data, deletion periods and evidence such as ISO 27001. How to check each of these points is covered in the GDPR checklist for AI meeting assistants. Public-sector clients in Germany sometimes also require an attestation under the BSI C5 criteria catalogue for cloud services. If that applies to you, ask for it specifically.

Works council: co-determination is not a hosting question

A common misconception is that the works council demands on-premise. In fact, under § 87(1) no. 6 of the German Works Constitution Act (BetrVG), it has a right of co-determination over technical systems capable of monitoring behaviour or performance. That covers usage, access rights, analyses and deletion periods, not automatically the server location. What that can look like for an AI meeting assistant is explained in the article on whether Sally is works council compliant. It is a different matter if an existing works agreement explicitly requires in-house operation. Then it is binding until it is changed.

When on-premise is mandatory or makes sense

For some organisations, running it yourself is not a preference but a prerequisite. What they have in common: an external or internal requirement says that meeting content must not leave the organisation's own network.

Public authorities, municipalities and defence

Many public bodies have their own data residency rules and their own data centres, often run by municipal or state IT service providers. In-house operation is the norm there, not the exception. It gets stricter where classified information is discussed. Classified material is subject to its own security regulations, and whether an AI meeting assistant is even an option for such meetings is decided by the security officer, not by IT alone. For classified conversations, an external cloud service is generally ruled out.

Critical infrastructure, banks and the supply chain

Operators of critical infrastructure and entities within the scope of NIS2 must manage cybersecurity risks, explicitly including their supply chain. Banks and insurers have been subject to the Digital Operational Resilience Act (DORA) since January 2025, which puts ICT third-party providers under close scrutiny. Neither regulation prescribes on-premise. But they make every outsourcing subject to review, and some organisations decide that another external service for meeting content is not worth the review effort. In that case, running it yourself is the pragmatic choice.

Healthcare, law firms and professional secrecy

Hospitals, practices, research institutions and law firms process health data under Art. 9 GDPR or professional secrets protected in Germany under § 203 of the Criminal Code (StGB). External service providers are not prohibited in principle: German criminal law allows them to be involved if they are bound to confidentiality. Many organisations nevertheless have internal policies stating that patient or client conversations must not leave their own network. Where such a policy applies, on-premise is no longer an option but a condition.

Client contracts, confidentiality and air-gapped networks

Sometimes the requirement comes from a client rather than the law. Suppliers in sensitive industries sometimes sign confidentiality agreements that tie certain development topics to their own network segment. The strictest form is an air gap, a network with no connection to the outside at all. Any function that needs a connection drops out there, such as a bot joining meetings on external platforms, and updates only arrive through approved channels. Whether and how a vendor supports fully disconnected operation is something you should ask explicitly rather than assume.

Boiled down to two questions, the decision looks like this:

Decision tree on-premise vs. cloudQuestion 1Does a rule requiremeeting content to stayin your own network?Question 2May the applicationrun externally if thecontent stays with you?Full on-premiseApplication, processingand data with youCloud in GermanyVendor without US ties,no in-house operationsHybrid modelContent with you,operations with the vendorIn every model:a DPA under Art. 28 andsecurity measures underArt. 32 GDPRYesNoNoYes
Two questions decide the operating model. Only then does it matter whether your IT team can carry the operations.

US cloud, German cloud and on-premise compared

The table puts the three basic models side by side. Hybrid models sit between the German cloud and on-premise: the data is with you, the operations with the vendor.

CriterionCloud (US vendor)Cloud (Germany)On-premise
Where is the data?Often selectable, including an EU regionIn the vendor's German data centresIn your data centre
Vendor's jurisdictionUS law, CLOUD Act appliesGerman and EU lawYou control the infrastructure, the vendor supplies the software
Third-country transfer (Art. 44 et seq.)To be assessed: adequacy decision, standard contractual clauses, transfer impact assessmentNone, if all sub-processors are also in the EUNone from operations, support access to be regulated separately
Agreement under Art. 28RequiredRequiredRequired as soon as the vendor has access
Operations and updatesVendor, continuouslyVendor, continuouslyYour IT team, with maintenance windows and sign-offs
Hardware and computeNone of your ownNone of your ownYour own platform, compute capacity and storage
Getting startedImmediately after sign-upImmediately after sign-upA project with installation and a pilot phase
Cost structureLicence per userLicence per userLicence plus infrastructure and staff
New featuresAvailable immediatelyAvailable immediatelyWith the next approved update
FitsTeams without sensitive content and with a properly assessed transferMost companies, including for confidential conversationsPublic authorities, critical infrastructure, healthcare, law firms, separated networks

To be fair: US vendors often have the largest ecosystem and ship new features quickly. If you do not discuss sensitive content and have assessed the transfer properly, you can work with them. And a German cloud is not a free pass. It also has to disclose its agreement, sub-processors and deletion concept, otherwise the location is just a label.

Cost and operations: what on-premise really costs

Flat figures are of little use here because every environment is different. More meaningful is a look at the cost over the full period of use, the total cost of ownership (TCO).

What the cloud licence already covers

In the cloud you usually pay a licence per user, and operations, hardware, updates and scaling are included. Costs are predictable and grow with the number of users. With Sally, plans start at €8 per month, and you can test it free for 30 days before you commit.

What running it yourself adds

  • Infrastructure: servers or private cloud capacity, compute for speech recognition and AI, storage for recordings and transcripts, backup.
  • Platform: container environment, monitoring, logging, network approvals and single sign-on.
  • Staff: time for installation, updates, testing, incidents and capacity planning, often including on-call duty.
  • Project: requirements, architecture, pilot and acceptance before the first user works productively.
  • Licence: usually an individual enterprise offer rather than a list price.

Often the biggest item is not the hardware but the working time permanently tied up in operations. It never shows up on a purchase order, which is why it is missing from many comparisons.

When the numbers turn around

On-premise can make economic sense when much of it already exists: a running container platform, spare compute capacity, a team with experience operating AI applications and many users across whom the fixed costs are spread. There is also a factor that appears on no hardware list. If every new outsourcing arrangement triggers a long review by data protection, information security and audit, running it yourself is sometimes simply the faster route to productive use.

Hybrid models and how Sally implements them

Between "everything at the vendor" and "everything with you" there are intermediate steps. Alongside the cloud, Sally offers two routes into your own infrastructure, described on the page about an AI meeting assistant inside your own infrastructure. Both use the same software and differ in how much you provide and run yourself.

Standard: cloud hosted exclusively in Germany

In standard operation, Sally is run by Aliru GmbH from Mannheim and hosted exclusively in Germany, with no transfer to third countries. Meeting data is not used to train AI models, and the Art. 28 GDPR agreement is available directly in the app. Sally joins Google Meet, Zoom, Microsoft Teams and Webex, understands 99+ languages and connects to 8,000+ tools. Details on encryption and certifications are on the page about GDPR and security.

Shared hosting: content with you, operations with Sally

Often the point is not to run the whole application yourself but only to keep meeting content inside your own network. With shared hosting the application keeps running at Sally, while recordings, transcripts, summaries, decisions and tasks sit on your servers. Operations, updates and support stay with Sally. You provide the storage, not the entire platform.

Full on-premise: everything inside your network

With full on-premise, the complete Sally installation runs in your infrastructure, including transcription and AI analysis. On request, you can connect your own language model. Users, permissions and logs stay with you, and single sign-on runs against your directory. On your side you need a container platform, compute capacity for transcription and AI analysis, storage for recordings, transcripts and the database, and defined network approvals for the integrations you want. Updates are agreed with your IT team and applied in a maintenance window.

From the first call to the pilot

An on-premise project with Sally runs in four steps: capture requirements, architecture and quote, installation and pilot, operations and rollout. IT and data protection belong in the first call, because that is where it is clarified which data, systems and rules apply. After that you get a proposal for setup, operating model and effort as a basis for your internal decision, and none of it happens without sign-off from your IT team. The cost is quoted individually as part of the enterprise offering, because the effort depends on your environment. The process in detail is on the page about Sally on-premise.

Checklist: the questions for your decision paper

Answer these questions together before you commit to a model. Put the answers in writing in the paper for management, so the decision can be justified later to a regulator or auditor as well.

  1. Requirements: Is there a policy, regulator, works agreement or client contract that specifies where data is stored or processed? If so, quote it verbatim.
  2. Content: Which conversations will be recorded, and do they include health data, professional secrets or classified information?
  3. Processing chain: Where do speech recognition, AI analysis and storage run, and which sub-processors are involved?
  4. Jurisdiction: Does the vendor or its group have US ties that trigger the CLOUD Act?
  5. Contracts: Is there an Art. 28 GDPR agreement, including for support and maintenance access?
  6. Security: Which Art. 32 measures are in place, who is responsible for which of them with on-premise, and is a data protection impact assessment under Art. 35 required?
  7. Operations: Does your IT team have the platform, compute capacity and staff for installation, updates and incidents?
  8. Features: Which meeting platforms and integrations do you need, and do they work in the chosen model with your network approvals?
  9. Cost: How do licence, infrastructure and staff compare over several years?
  10. Co-determination: Has the works council been involved early, and does the works agreement regulate usage, access and deletion?

When on-premise is not worth it

An honest assessment: on-premise is overkill when no requirement demands it and no IT team can carry it. A company with three admins who already look after workstations and the network, and are now supposed to run an AI platform with its own compute capacity, mainly buys itself outage risk. The data is then formally in-house but in practice often less well protected than in a professionally run German data centre.

If the wish for on-premise comes from a vague sense of unease, it is worth naming the cause. Usually it is the fear of US access or of AI training on your own data. A German vendor without US ties that excludes training solves both. And if only a few meetings are truly sensitive, such as the supervisory board or HR conversations, it is often simpler not to record those at all than to run a separate platform for every other meeting.

Conclusion: settle the requirement first, then the operating model

The question of on-premise vs. cloud for an AI meeting assistant is not decided by gut feeling but by two points. Does a requirement say that content must stay inside your own network? And can your IT team carry the operations? If the first answer is no, a cloud in Germany with a clean contract is enough. If it is yes, the second answer decides whether a hybrid model is sufficient or full on-premise is needed.

Back to the table on the third floor: the data protection officer is right that location matters. The head of IT is right that operations are work. Keeping the two apart leads to a decision that still holds in three years. If a requirement calls for in-house operation, take a look at Sally's on-premise options and book a consultation with your IT and data protection team. If the cloud in Germany is enough, you can test Sally free for 30 days, from €8 per month. Create your account now and let Sally join your next meeting.

FAQ

Gianni Piruzza

Gianni Piruzza

Marketing Manager

“Especially in today's age of Zoom meetings, automatic meeting transcription is becoming increasingly important. Without the support of a suitable AI tool, you get lost in the documentation chaos.”

Learn more about the author