Why teams enable the Collabora Online AI assistant with a self-hosted model
A document platform team is often asked to offer AI help inside Writer, Calc and Impress without sending confidential content to a third party. Collabora Online 26.04, announced on July 23, 2026, adds an optional AI assistant across those three applications. It is off by default, and the administrator chooses the AI provider and whether that provider is hosted externally or runs on infrastructure the organization controls. That choice is what makes the feature usable in regulated settings, but it is only a starting point. Before anyone turns it on, confirm in the 26.04 release notes and the COOL 26.04 release post exactly which providers and configuration keys are supported in your build.
The decision belongs to the owner of the document platform, usually the platform or IT lead, working with information security and the data protection officer. Security should review where prompts and document excerpts travel, what the provider retains, and how access is limited to user groups. Legal and compliance should confirm that the processing fits existing data processing agreements and classification rules. Application owners in finance, legal or HR should be consulted because their documents carry the most sensitive content, and they will be the first to ask whether an assistant can see it.
The feature is not needed if your users rarely edit documents in the browser, if your data classification forbids any AI processing of office files, or if no provider you control or have approved is available. In those cases, leaving the assistant disabled is a complete and defensible policy. Document that decision so it is reviewed again when the provider landscape or your classification rules change.
Provider options and when each fits
This table compares the three paths open to an administrator: an external hosted model provider, a model served inside your own infrastructure through a gateway, and keeping the assistant off. Read each row to see where prompts and document content travel, who runs the model, what to confirm before enabling anything, and how much governance work each path adds. The main takeaway is that the assistant is off by default, so the decision to turn it on is itself a governance decision.
| Option | Data residency and operator | What to confirm before enabling | Governance burden |
|---|---|---|---|
| External hosted provider | Prompts and document excerpts leave your network for the provider's infrastructure, and the provider operates the model | Where requests are processed, retention and training terms, sub-processors, and whether the provider's contract covers your data classification | Highest. Needs a data processing agreement, a documented data flow, and periodic review of provider terms |
| Model in your infrastructure via gateway | Prompts and content stay inside your network if the model endpoint and gateway are hosted there, and your team operates the model | Which provider integration Collabora Online supports for a self-hosted endpoint, required API compatibility, and whether logging is kept inside your boundary | Moderate to high. You own model hosting, access control, patching, capacity and logging of prompts and outputs |
| Assistant off (default) | No prompts or document content are sent to an AI provider through the assistant | Confirm the setting remains disabled in the 26.04 configuration and that users cannot enable it through their own accounts or integrations | Lowest. Document the decision, review it when the platform or data policy changes, and keep it in the change record |
| Choose when data is restricted | Use the hosted provider only for documents classified as public or internal, if your policy allows it | Classification rules, user guidance, and whether the integrating platform can enforce the split by group or document | High. Requires classification labels that users apply consistently and periodic audits of usage |
Choosing external or self-hosted by data sensitivity and operating model
Documents under strict residency, legal privilege, patient or customer data, or classified handling rules carry the highest risk when a prompt leaves your network. Under an external provider, the text you select and the surrounding context the assistant sends are transmitted to the provider's infrastructure, so the question becomes which region, retention terms and sub-processors apply. A self-hosted path keeps prompts and documents inside your boundary, but it moves the work to your team: you choose a model, provision accelerators or CPU capacity, patch the serving stack and monitor latency under load.
- Data flow: Map every endpoint the Collabora Online server calls for the assistant and confirm it resolves only to the provider you approved.
- Compute: Plan accelerator or memory capacity for the model you select, not for Collabora Online itself, which is sized separately.
- Operations: Assign an owner for model updates, security fixes in the serving stack and capacity reviews.
- Logging: Decide whether prompts and responses are logged, where those logs live and who can read them.
- Gateway and serving: A gateway such as LiteLLM or a serving engine such as vLLM or Ollama may be relevant as separate decisions to validate against your model choice.
Before approving this tab, confirm with your legal and privacy teams which data classes may enter a prompt at all, and ask the provider in writing for its retention and training terms.
For drafts, internal memos and non-sensitive working documents, an external provider may be a reasonable starting point because the content carries lower risk and the team gains capability quickly. The trade-off is that each prompt still leaves your network, so the approval scope should be written down by content type rather than assumed. Operating cost in this model shifts toward usage-based fees and governance of who may send what, which needs its own policy.
- Data flow: Document which Writer, Calc and Impress features send text to the provider, since each can send different amounts of context.
- Cost: Ask the provider how usage is billed, whether there are minimums, and how cost is capped per user or group.
- Access: Limit the assistant to approved groups so it is not switched on for everyone by default.
- Review: Set a periodic check of which teams use the assistant and for what content.
- Exit path: Confirm you can switch providers later without changing the document workflow.
Before rollout, ask whether the provider's terms allow you to restrict data use and whether a data processing agreement covers the assistant feature.
Many organizations run Collabora Online well but have no staff who operate machine learning serving. For these teams, self-hosting the assistant adds a new production service with hardware, upgrades, capacity planning and incident response, all outside their current skills. An external provider removes that operational load, at the cost of the data flow discussed above, so the decision depends on whether the risk accepted is smaller than the capability you can staff.
- Skills gap: List who would patch, monitor and troubleshoot the model service, and whether that work fits an existing platform team.
- Sequencing: Start with the default-off setting and enable the assistant for a small pilot group before any production use.
- Partner support: If you engage a partner for serving or gateway work, scope it in writing with named deliverables and support boundaries.
- Recurring work: Model and engine updates recur, so budget for them as ongoing effort, not a one-time build.
- Fallback: Decide what happens when the provider or your self-hosted service is unavailable.
Before choosing, ask your platform team to estimate the operational hours they can commit each month, and compare that figure with the support terms the provider or partner offers.
How to enable the assistant safely in Collabora Online 26.04
This sequence moves the optional AI assistant from a default-off installation to a pilot group with an approved provider path, a documented data policy and evidence that documents and prompts stay where you expect. Each step closes with a check you can record before the next one starts.
Confirm the running branch is 26.04
Your platform or infrastructure lead checks the Collabora Online version on every server and container that will serve editing sessions, not only the one used for testing. The AI assistant is part of the 26.04 branch announced on July 23, 2026, so any node still on an older branch cannot offer it. Record the exact version string for each node in your change record. Done means every node reports 26.04 and no node is left on a mixed version that could give different users different editing features.
Record the data classification for each user group
Your security or compliance lead, working with data owners, writes down which classifications of content each user group may send to an AI assistant. Include Writer, Calc and Impress separately, because spreadsheets and presentations often hold different sensitive data than word processing files. Ask the business owner which groups handle regulated personal data, contracts or unpublished financial information. Done means a signed table maps each group to an allowed classification level, and any group without an entry is treated as not permitted.
Select and validate the model provider
Your architecture team chooses between a provider hosted outside your infrastructure and a model you run on infrastructure you control. Collabora lets administrators choose the provider and whether it is hosted externally or on their own infrastructure, so confirm the provider options and their data handling terms with Collabora and the provider before you commit. Ask the provider in writing where prompts and document content are processed, stored and logged, and how long. Done means a written provider decision names the endpoint, the hosting location and the data terms that your data classification table relies on.
Confirm the assistant is off by default
Before you change any setting, your platform engineer confirms that the assistant is disabled on the server and that no user group has it enabled. The feature is off by default in 26.04, so this check verifies that your deployment has not inherited an enabled state from a template, a Helm value or a previous test. Open a Writer, Calc and Impress document as a user outside the pilot group and confirm no assistant control appears. Done means a dated screenshot or configuration record shows the feature is off for everyone.
Enable the assistant for a pilot group only
Your platform engineer turns on the assistant using the provider configuration approved in the previous steps, and limits access to a named pilot group that matches the classification table. Keep the pilot to the smallest group that answers your questions about usefulness and data handling. Confirm that the Nextcloud or other WOPI host still restricts editing rights for the pilot group as your access model requires. Done means the pilot group can use the assistant and every other group still cannot see it.
Verify network egress matches the approved path
Your network or security engineer watches outbound connections from the Collabora Online servers while a pilot user sends test prompts. Traffic should go only to the provider endpoint and hosting location recorded in the approved decision. Any connection to an unlisted destination, including a fallback model or telemetry endpoint you did not approve, stops the rollout until it is explained. Repeat the capture after a server restart, since configuration can reload with different defaults. Done means the captured destinations match the approved path and the capture is filed with the change record.
Test document fidelity and prompt logging
Your compliance lead and a set of pilot users open representative Writer, Calc and Impress files, including Microsoft formats, and run the assistant on them. Confirm that the assistant does not alter the saved file unless a user accepts a change, and compare the output file to the original for layout, formulas and slides. Check what prompt and response logs are kept, where they are stored, who can read them and when they are deleted. Done means the test results are recorded, any fidelity defects are logged with owners, and the logging behavior matches your data policy before the pilot widens.
Security controls for prompts, documents and provider access
The AI assistant adds a second outbound path from your editor: document content and user prompts may leave the Collabora Online server for a model provider. The cards below follow that path from the editor to the provider, then to the admin settings that switch it on, so each boundary has a named owner and a check before rollout.
Map what the editor sends
Before enabling the assistant, list the content each request carries: the user prompt, the selected text or the surrounding document context, and any metadata Collabora Online attaches. Confirm with Collabora exactly which parts of a document reach the provider for Writer, Calc and Impress, since a spreadsheet may expose far more cells than the user intended. Treat this map as the input to your data handling policy and re-check it after every major release, such as the 26.04 branch that introduced the feature.
Confirm provider retention and logging
Retention and logging are set by the provider, not by Collabora Online. Ask the chosen provider how long prompts and completions are stored, whether they are used for training, which staff can read logs, and where those logs are held. For a hosted provider, check whether the contract allows deletion on request. For a self-hosted model, you control these settings, but you still need to define them in writing and test that logs do not capture document text by default.
Store provider credentials securely
An external provider needs an API key or equivalent credential that the Collabora Online server presents on every request. Keep it out of image layers, shared configuration files and chat logs, and load it from your secrets store or a Kubernetes secret with restricted read access. Scope the key to the assistant only, rotate it on a schedule you set, and have a tested way to revoke it if the server is compromised. Confirm how the chosen version of Collabora Online stores this value before writing your runbook.
Encrypt traffic between components
Document content moves between the browser, the Collabora Online server, the integrating platform and the model endpoint. The Nextcloud administration manual recommends using the same protocol for the Collabora server URL as for the Nextcloud installation, preferably HTTPS. Apply the same rule to the model endpoint: use TLS with certificates your team validates, and reject plain HTTP for any provider address, including internal model servers in your own network.
Restrict outbound network egress
Allow the Collabora Online server to reach only the provider hostnames or private model addresses you approved, and block all other outbound traffic from that host. Pair this with the Nextcloud guidance to restrict WOPI requests to known Collabora server IP addresses, so that only your integrating platform can reach the editor. Test the rules from the server itself, because a firewall that looks correct on paper may not block a DNS-based path to an unapproved provider.
Limit who can change AI settings
Turning the assistant on or changing its provider is a data-flow decision, so restrict those settings to a small administrator group. In Nextcloud Office, administrators can limit which groups use the editor and which can edit, which gives a starting point for limiting who opens documents with the assistant available. Log every change to provider settings, review them in your change process, and check the admin console access controls separately, since it shows open documents and per-document activity.
Track security advisories and patches
Collabora publishes security advisories for Collabora Online. The 25.04 release notes, for example, record that 25.04.7.5 fixed CVE-2026-23623 and that package signing keys were updated for RPM and DEB users. Subscribe to these notices, schedule point releases in your patch window, and verify package signatures during upgrades. An assistant component added in a new branch also needs its own review, so do not treat a routine point release as proof the AI path has been checked.
Limits to confirm before you enable the assistant
The AI assistant is optional and off by default in Collabora Online 26.04, so the decision to enable it should follow confirmation of what the feature supports in your environment. The cards below cover the questions to put to Collabora or test internally before any user sees the assistant.
Language and document coverage
Confirm which languages the assistant handles in Writer, Calc and Impress in 26.04, and whether it works on every document type your users open, including DOCX, XLSX, PPTX, ODF and PDF. Ask whether support differs between the three applications. Test with a sample set of real business documents from each department, and record which prompts return usable output and which fail or degrade.
Provider options per deployment method
Confirm that the external or self-hosted provider choice you need is available whether you run Collabora Online from Docker images, system packages or Kubernetes. Ask whether the provider settings are the same across all three methods and whether any option requires a specific build or the COOL Controller. Document the chosen method, since it determines how provider configuration is applied and upgraded.
Resources for a self-hosted model
A self-hosted model adds compute, memory and storage demands on top of the editing server. Ask Collabora which model sizes and hardware it validates, and benchmark response time under your expected number of concurrent editors. Keep model hosting on separate capacity from the editing pods so that a model load does not affect document sessions.
CODE versus supported subscription
The free CODE edition is a rolling release that Collabora does not recommend for production. Confirm whether the assistant is available on CODE at all, and if it is, whether it receives the same fixes and support as a Business or Enterprise subscription. Ask what happens to provider settings when you move from CODE to a subscription, and whether a supported subscription changes the update channel you should plan around.
Data flow and logging behavior
Before enabling any external provider, confirm exactly what leaves your network: the selected text, the whole document, metadata, and any user identifiers. Ask whether prompts or responses are logged by Collabora Online, by the provider, or both, and how long each keeps them. Verify the answers with a packet capture or egress log in a test environment.
Versions and patch cadence
Confirm the assistant's dependencies and which Collabora Online point releases include fixes for it. Check how the 26.04 branch receives security updates compared with earlier branches, and ask how the update process handles provider configuration. Record the version you test so you can reproduce any problem when Collabora publishes a later release.
Questions buyers ask before enabling the AI assistant
What does enabling the AI assistant cost, and what drives the cost?
The cost comes from three areas: the Collabora Online subscription or support, the model provider path you choose, and the internal or partner effort to govern and test the feature. Collabora's pricing page lists the Business plan at 3 euros per user per month for up to 99 users, which covers long term support, a service level agreement, signed security updates and customer portal access. Model costs depend on whether you use an external provider, which bills by usage, or a model you host, which adds compute, storage and operations work. Ask for a quote that separates one-time setup from recurring licence, model and support charges.
Which AI provider path should we choose: external or self-hosted?
The choice depends on where prompts and document content may travel under your data policy. Collabora 26.04 lets administrators pick the AI provider and whether it runs externally or on their own infrastructure. An external provider usually means less operations work but requires a data processing review and contractual terms on retention and training. A self-hosted model keeps prompts inside your environment but adds accelerator capacity, model updates and monitoring. Confirm with Collabora which provider types the assistant supports before you fix the design.
How long does a pilot take, and what sets the duration?
A pilot length is set by the number of applications and users involved, the provider path, and how much evaluation you need before approval. Start with a single group of users in Writer, Calc and Impress, then extend once the governance checks pass. The main drivers are whether a self-hosted model must be sized and deployed, how many approval steps your security and legal teams require, and how much document-fidelity testing you want. Ask any partner for a fixed-scope plan with dated milestones before work begins.
What must our team provide before the work starts?
Your team needs to provide an existing Collabora Online deployment or a plan to install one, the Nextcloud or other WOPI host integration, administrator access to the hosts, and a named owner for AI policy. You also need the approved list of model providers, a written data handling position on prompts and documents, and test documents that represent your real workload. For self-hosted models, expect to provide the compute platform, storage and network access rules. A partner will normally ask for these inputs in writing before preparing a scope.
Can our internal staff run the AI assistant on their own?
Internal staff can run it if they already operate Collabora Online and have capacity for model operations, security review and user support. The work that often needs experienced help is provider configuration, governance settings, load and quality testing, and documenting who may enable the feature. Plan for ongoing ownership of model versions, provider terms and Collabora release updates. A retainer can cover recurring support, while the initial build can be handled internally or scoped as a fixed project.
Is CODE suitable for an AI assistant rollout?
CODE is not recommended for this rollout. Collabora describes the free CODE edition as suited to testing, home use or small teams and not recommended for production. A feature that handles prompts and business documents needs the long term support, service level agreement and signed security updates that come with the Business or Enterprise plans. You can use CODE in a sandbox to evaluate the assistant, then move to a supported plan for production. Check the CODE page and pricing page for current terms.
How do we keep AI use governed and tested after launch?
Keep the feature off by default, enable it per group, and review the provider settings whenever Collabora publishes a release or a security advisory. Test output quality and data handling with your own documents before widening access, and record which groups may use the assistant. Ongoing testing should cover provider changes, prompt logging rules and model updates. Ask for a written governance procedure that names the owner for each setting, so audit evidence stays in your hands.
How DeployOpen scopes and delivers AI assistant rollouts
DeployOpen delivers the AI assistant rollout as a fixed-scope engagement, with the scope, deliverables and timelines agreed in writing before any work starts. For this topic, the written scope should name each application in play (Writer, Calc and Impress), the WOPI hosts that will expose editing, the provider path chosen (externally hosted or on infrastructure you control), and the default-off setting that will be confirmed in every environment. Because the assistant ships in Collabora Online 26.04 and is off by default, the scope should state explicitly which users or groups, if any, will be allowed to turn it on, and how that will be verified. A quote follows from that written scope, so the more precisely the decisions are recorded up front, the more precisely the work can be priced and timed.
The engineering work covers the Collabora Online deployment itself and the parts of the integration that touch model traffic: network paths to the chosen provider or private model endpoint, egress rules, administrator configuration, logging of assistant activity where your policy requires it, and hardening of the surrounding service. DeployOpen implements and hardens this inside your own cloud accounts, data centers or air-gapped networks. DeployOpen can also run security testing of the integration, such as checking that prompts and document content leave only through the paths you approved, and that the assistant cannot be reached by users outside the groups you defined. Ongoing support, version updates and periodic retesting are available as retainers, which suits a feature that a provider or Collabora may change after go-live.
DeployOpen does not issue audit reports, attestations or certifications. Evidence produced during the engagement, such as configuration records, test results and change logs, can help your compliance assessor, but the assessor and any certification body make their own determinations. The governance policy for the assistant, including whether it is enabled at all, which data classes may be sent to a model, who approves a provider, and how usage is reviewed, remains your organization's decision. DeployOpen can help you document that policy, implement its technical controls and test them, but the policy itself should be owned and signed off by your security, legal and data protection leads before the rollout is scoped.
