Changelog

Why we built automated orchestration

The Orchestrator has been part of neuland.ai HUB for a while, but invisibly: it worked behind the scenes, and you instructed it anew with every request. So you had to describe recurring processes again every time. That is exactly what we wanted to change. With automated orchestration, we lift this function to a level you can see and control: you define a process in a chat dialog, have it displayed and then run it as often as you like, with no code, no node editor, no workflow builder. The naming may still change when the beta ends; the way it works will stay.

How to create an orchestration

Describe the process in the chat

Describe in plain everyday language what should happen, for example: “Every Monday, get all open Jira issues from project HUB, summarize them and send them to me as a structured email.” Where something is unclear, neuland.ai HUB asks follow-up questions: which project, which recipient, which format.

Save as an orchestration

As soon as the chat delivers the results you expect, have the process behind it created as an orchestration, using the “Create Orchestration” button at the top of the chat or the corresponding command directly in the chat. Converting it into a workflow can take anywhere from a few to several minutes.

Review the flowchart and run it again

The finished process is shown to you as a flowchart: you see every step, every condition and every tool used. The workflow stays saved and can be run again or adjusted at any time, either in the associated chat or in the Orchestration tab.

Important for using it: you do not edit the steps directly in the diagram. You describe changes in the chat again, for example “Remove the step with the Excel export and just send the text instead”. neuland.ai HUB then rebuilds the workflow. This is by design: you should focus on the what, not on wiring up building blocks.

What orchestration can do

• Every tool the chat can use. Whatever your users reach today in the chat via a connector is also available to orchestration, including Jira, Confluence, HubSpot, Outlook (mail & calendar), Microsoft Teams, Fireflies and SQL databases, as well as the AI tools of neuland.ai HUB such as Code & Data Analyst, summaries and text drafts.
• Not just linear chains. Depending on how you describe the process, you get simple sequences or workflows with conditions (“if there is no result, then …”) and loops, such as “go through each contact in the list one by one”.
• Deterministic execution. The steps run as a fixed process through the respective interfaces. Results vary only where AI is deliberately involved: writing texts, summarizing content.
• Management in one place. In the orchestration view, you see all saved workflows and can view, adjust, start manually and delete them. During each run, you see live which step neuland.ai HUB is working on.

Beta: what you should know

We are deliberately starting as a beta. Automated orchestration is a big step for the platform, and it is new. In the first weeks, errors may occur, especially with very complex processes, when several systems are used in one step, or when sending to multiple recipients. Manually refining individual branches of very complex workflows is not yet possible at launch; for now, the only option is to describe them again in the chat.

Orchestration is enabled per organization and is not available by default. If you want to get started, please coordinate with your organization and contact support@neuland.ai. The same address also reaches us with any bug report, and each one helps us harden orchestration quickly.

Automatic model selection (beta) and Effort

Which language model is the right one? And how long should it think? You no longer have to ask yourself these two questions with every message in neuland.ai HUB. Both switches below the input field are now set to Auto by default: neuland.ai HUB decides on its own what fits each request.

Automatic model selection

When model selection is set to Auto, each individual message is routed to a suitable language model based on its complexity. A short follow-up question goes to a fast model, a demanding analysis to a more powerful one, all within one ongoing conversation. So you don’t need to predict in advance where a conversation is heading. Only models available to you are used: whatever your administrator has not enabled via Resource Access, automatic selection won’t use either. You stay in control: as usual, you can set a specific model manually at any time via the model selector, for example if you want to compare results or deliberately prefer a particular model for a task.

Effort: how long the model thinks

Modern language models can take more or less time to “think” before answering. More thinking time produces better results on tricky tasks, but costs time and compute. That is why you now find the Effort switch next to the model selector. One click opens a slider between “Faster” and “Smarter”; the setting applies immediately to your next message. Move it left for short, clear tasks such as rephrasing, translations or quick follow-up questions: the answer comes noticeably faster. Move it right for truly complex tasks such as multi-step analyses, demanding calculations, tricky code or problems with many constraints: longer response times are normal here.

Our recommendation: the default sits in the middle and is just right for typical tasks: summarizing, writing, researching, questions about documents. In the vast majority of cases, you don’t need to change anything. Only turn the slider up if the result of the default setting did not convince you. Working permanently on “Smarter” slows down simple requests unnecessarily and increases costs without improving the result.

Send files directly from the chat

Until now, neuland.ai HUB could write messages and emails via the connectors, but not attach files. That changes: attachments are now supported, initially for Outlook (email) and Microsoft Teams. Upload a file in the chat as usual or have one generated, for example an Excel analysis by the Code & Data Analyst, and ask neuland.ai HUB to send it: “Email this analysis to the team.” Before sending, you see the finished draft including the attachment and approve it; the human-in-the-loop principle applies here too.

Good to know: the default limits are typically 20 MB per file, 25 MB per send and a maximum of 10 files; your organization may differ. For technical reasons, Microsoft Teams provides the file via OneDrive/SharePoint and links it in the message. And on the first send, your users may need to confirm permissions with Microsoft again. Also new: emails now keep the displayed formatting when they are opened as a draft via the corresponding option in the email client.

Agents as tools in the chat

Your agents are no longer just separate chats: they can now be called as a tool within an ongoing conversation. So you keep working in the regular chat and bring in the right specialist when you need it, instead of switching conversations. A typical example: you are researching a contracting party and, in the middle of the chat, have your NDA review agent assess the attached document; result and context stay in one place.

How to enable it: open the Agents area, find the card of the agent you want, click the three dots (…) in the top right corner and select “Make available as tool”. Once provisioned, the agent is available in the chat under the plus sign for the users it is shared with. Visibility follows the agent’s existing sharing: exactly the users you have already shared it with will see it; anyone without access to the agent won’t be offered it as a tool either. So you don’t have to maintain permissions twice.

This is especially useful for agents with a clearly defined task: reviewing, assessing, rephrasing, researching. Very broad all-rounders benefit less, because the same capabilities are available in the chat anyway.

Usage & costs at a glance: the new Analytics dashboard

Usage reporting in neuland.ai HUB has been significantly expanded. Instead of a single total, administrators get a detailed breakdown: usage and costs by level (tenant, user, group, agent and API key), plus file and storage statistics on uploaded content. The views are role-dependent: who may see which report depends on their permissions. Also new is a usage forecast that projects your current monthly trend so you can take action early. You find the reports under Settings → Consumption.

Resource Access: control who may use which models, tools and connectors

Until now, whatever was enabled in neuland.ai HUB was available to all users. With Resource Access, administrators can now restrict access specifically, to individual groups or individual people. You find the area in the administrator settings under Roles. By default, everything is initially available to everyone: if you don’t want any restriction, you don’t need to do anything, and neuland.ai HUB behaves as before. The restriction only applies once you assign specific groups or users to a resource.

Governance: central policies for your tenant

Until now, retention rules, AI guidelines and guardrails were spread across different places. In the new Governance area, you manage these policies together in one place, valid for the entire organization. You find it under Settings → Governance; changes only take effect once you click “Save”.

• Document retention. Define after how many days unassigned documents are deleted automatically; the default is 90 days. Documents assigned to a chat, a project or a knowledge base are not affected; only orphaned uploads are. This keeps your data lean and supports your deletion policies.
• Email notifications. You decide what administrators are notified about automatically: budget alert emails as soon as a budget threshold is exceeded, and soft limit warning emails as soon as a user reaches the soft spending limit.
• Instruction Prompt. Store an overarching instruction for how the AI should respond in your organization, such as tone (friendly or formal), a professional role or a preferred language. This instruction applies tenant-wide and no longer needs to be repeated in every single agent.
• Guardrails. Guardrails can be enabled and disabled centrally. With the inbound guardrails, you describe in your own words how the AI should handle incoming requests, for example which topics it should reject or flag with a notice. Because the rules are written in natural language, you can tailor them exactly to your compliance requirements.

Our recommendation: test new guardrail rules first with typical phrasing from your everyday work before enabling them organization-wide. This ensures that legitimate requests are not blocked by mistake.

Improvements

Alongside the new features, this sprint, including an interim hotfix, brought more than 30 bug fixes. The most important ones at a glance:

• Stable chat. Fixed an issue where the chat got stuck with an endless loading indicator in certain situations, aborted with sporadic error messages or stopped responding entirely after long load times. The “Continue” button for very long, truncated answers also works reliably again.
• More accurate guardrails. Guardrails used to block conversations too often and with unclear reasons. Detection and the message text have been reworked.
• Display fixes. Minor issues with code blocks, table rendering and voice input are fixed.
• More reliable agents. Agents restarted when they hit the output limit or hung indefinitely without an error message; both are resolved. Agent settings also show the model actually in use again.
• Automatic model fallback. If a model fails, a fallback model now takes over reliably, and you are notified.
• Code & Data Analyst. The tool no longer appends raw output to its summaries.
• Sign-in. Several sign-in issues are fixed, such as “Email not found” despite a valid account, failed logins for external addresses and faulty SSO configurations for individual tenants.
• Connectors. Confluence and Fireflies connect without problems again, and authorization errors with individual integrations are a thing of the past.

For bug reports, feature requests or tool approvals, write to us any time at support@neuland.ai. We look forward to your feedback: it decides what we build next. 🙌

Why we built the SQL connector

Your data doesn’t only live in documents and wikis; much of your most valuable business information sits in databases. Until now, getting answers from a database meant writing queries yourself or switching to another tool. We wanted to change that. With the new SQL connector, you simply ask a question in natural language and let the AI do the rest, without ever leaving the chat.

How to use the SQL connector

Connect your database
Go to Settings → Integrations and look for the new SQL connector card. Enter your database connection string, and you’re connected. A clear status indicator shows whether the connection is active. You can update or disconnect it at any time.

Natural language queries
Once connected, just activate the SQL tool in your chat and ask your question naturally, for example: “What were the best-selling products last quarter?” or “Show me all open orders for customer X.” The AI translates your question into a safe, read-only query and returns the results directly in the chat.

Security by design

The SQL connector was built with security as its top principle. It runs read-only queries exclusively; your data can never be changed through the chat. So you can safely give your team access to data-driven answers without worrying about unintended changes.

Web search safety banner

We have also introduced a warning banner that cannot be dismissed and appears automatically as soon as web search or Deep Search is active in a chat. The banner reminds users not to enter confidential information, such as customer names or account numbers, that could be passed on to external search services. The banner appears and disappears in real time depending on whether the tool is switched on or off, so users are always informed at the right moment. Tenant admins can enable or disable this banner globally in the tenant settings panel, giving them full control over when and how the notice is shown.

Both features are still in an early version and we are actively working on them. We look forward to your feedback: it decides what we build next. 🙌

Improvements

• Chat UI state fixes: Fixed an issue where the loading animation in the chat title kept spinning after a complete answer and the auto-generated chat name in the overview did not update without manually reloading the page. Both now update immediately and reliably.
• Confluence integration fixed: A persistent error that prevented the Confluence tool from working correctly has been fixed. Users no longer need to switch models or reconnect the integration to get results.
• SharePoint for multi-tenant environments: Fixed an issue where users from a different organizational tenant could not connect the SharePoint integration even though SSO worked fine. The connector now supports these cross-tenant scenarios correctly.
• Deleting chats works again: Fixed a long-standing bug that caused an error when deleting chats with stored usage history. Chats can now be deleted smoothly and without unexpected errors.
• Stability with long outputs: Fixed an issue where generating very long AI answers caused the chat to freeze silently mid-sentence. The platform now handles output limits much more robustly.
• Code & Data Analyst improvements: Fixed a compatibility issue that caused the Code & Data Analyst tool not to work with certain AI models. The tool now works reliably with all supported models. We also fixed an encoding error that displayed exported files incorrectly, and improved the placement of generated images within the answer: they now appear exactly where the AI intended, no longer all at the beginning.
• Central integration settings for admins: Admins and super admins can now configure integrations centrally for the entire tenant, including endpoint URLs, API keys and other parameters, directly from the integrations page in the settings, without every user having to set them up themselves.
• Group management: Administrators can now create and manage user groups directly in neuland.ai HUB, without Active Directory. Groups can also be synced automatically from your existing identity provider, and agents can be assigned to groups so the right people always see the right tools.
• Mistral models added: Mistral models are now available as an option when choosing the AI model in the chat and in agents.
• Performance improvements: Fixed a platform-wide performance issue that caused long load times in projects, the prompt library and agents for some users. The platform should feel noticeably faster and more reliable.
• OpenDesk / Nextcloud connector: We added support for Nextcloud in the openDesk Sovereign Workplace Suite, a popular solution in the German public sector. Users can now connect their Nextcloud files and access them directly from neuland.ai HUB, in a GDPR-compliant, data-sovereign environment.