Reference Architecture for an On-Premises AI Assistant for Microsoft Exchange.
Bringing Secure AI Capabilities to On-Premises Microsoft Exchange Many organisations continue to operate Microsoft Exchange within their own data centres because of regulatory requirements, data-residency policies, security considerations, operational dependencies, or the need to maintain complete control over sensitive email information. At the same time, users increasingly expect Artificial Intelligence to help them manage growing email volumes, understand long conversation threads, locate information quickly, analyse attachments, and prepare relevant responses. This creates an important question: Can organisations introduce AI-assisted email capabilities without migrating their on-premises Exchange environment or sending email data to public AI services? With the right architecture, the answer is yes. The opportunity for AI in on-premises Exchange Users spend a significant amount of time reading lengthy email chains, searching for previous discussions, identifying action items, reviewing attachments, and preparing responses. An AI-powered email assistant can help reduce this effort while continuing to operate within the organisation’s controlled infrastructure. Such a solution can be integrated with Microsoft Exchange through an Outlook Web App, or OWA, Add-In. The Add-In provides AI-assisted capabilities within the familiar Outlook Web interface and can be centrally deployed to selected users or Active Directory groups by the Exchange administrator. The purpose is not to replace Exchange or change the way emails are managed. Instead, it introduces an intelligence layer that helps authorised users interact more efficiently with the email information they are already permitted to access. Key capabilities of an AI-powered Exchange Email Assistant 1. Email and conversation summarisation The assistant can summarise an individual email or an entire conversation thread. Instead of reading every message in a long chain, the user can receive a structured summary covering: The purpose of the conversation. Important decisions and commitments. Key action items. Deadlines and pending responses. Relevant information available in attachments. The summarisation process can consider both incoming and sent messages so that the complete context of the conversation is preserved. 2. Contextual draft reply generation The assistant can prepare a suggested response based on the complete email thread and the user’s instructions. Users may select an appropriate tone, such as: Formal. Neutral. Technical. The generated response is placed in the user’s Exchange Draft folder or displayed for review within the OWA Add-In. A critical control is that the AI assistant does not independently send the email. The user remains responsible for reviewing, editing, approving, and sending the response. This human-review approach helps organisations adopt AI while retaining accountability over external and internal communications. 3. Natural-language mailbox search Traditional mailbox search generally depends on exact keywords, sender names, subjects, or date filters. An AI-enabled search experience allows users to search their mailbox using natural-language questions, such as: “Find emails received last month regarding the infrastructure renewal.” “Show unread emails from the finance team that require my approval.” “Find the discussion where the customer confirmed the delivery date.” “Summarise emails from a particular sender concerning the purchase order.” The solution can combine semantic search with traditional keyword-based search to improve the relevance of results. Users can also apply filters such as sender, subject, date range, read or unread status, and conversation thread. 4. Attachment analysis and OCR Important information is often contained within email attachments rather than the email body itself. An on-premises AI assistant can process supported documents, including: PDF files. Microsoft Word documents. Microsoft Excel files. Scanned documents. Image-based attachments. Optical Character Recognition, or OCR, can be used to extract information from scanned or image-based documents. The extracted content can then support email summarisation, contextual draft generation, and mailbox search. For example, a user may ask the assistant to prepare a response based on the commercial terms or delivery date mentioned in an attached document. Maintaining strict mailbox privacy Mailbox privacy is one of the most important architectural considerations for an enterprise AI email assistant. Each user must only be able to access and process emails from their own authorised mailbox. Information retrieved from one user’s mailbox must never become available to another user. This isolation should be enforced across every layer of the solution, including: User authentication. Exchange access tokens. Application sessions. Search indexes. Vector embeddings. AI prompts and responses. Cache and temporary data. Audit logs. The backend should validate the authenticated user and verify mailbox ownership for every request made to Exchange. Where vector search is used, mailbox data should be stored in user-isolated collections, partitions, or namespaces. This helps prevent cross-user retrieval at the data layer rather than relying only on controls within the user interface. Operating without public cloud AI services For organisations with strict security or data-sovereignty requirements, the AI models and supporting components can be deployed within the customer’s infrastructure. A fully on-premises architecture may include: A locally hosted Large Language Model. An on-premises embedding model. A re-ranking model. A vector database. A keyword-search engine. A backend API layer. An OCR or document-intelligence service. A secured OWA Add-In. Integration with Active Directory. Audit, monitoring, and administrative controls. The entire solution can operate within an internal network without sending email content, attachments, prompts, or generated responses to public AI platforms. This model can also support air-gapped environments where internet access is either restricted or completely unavailable. Centralised deployment and administration The OWA Add-In can be deployed centrally through Exchange administration controls. Administrators can assign the Add-In only to selected users or security groups. Employees who are not assigned the solution will not see the AI assistant within their Outlook Web interface. Administrative controls may include: User and group assignment. Enabling or disabling individual AI features. Summarisation-only mode. Draft-generation controls. Attachment-size restrictions. Session and retention policies. Model and performance settings. Usage monitoring. Audit logs. Resource utilisation monitoring. These controls allow organisations to begin with a limited group of users and gradually extend the solution after validating performance, security, and user adoption. Security and governance considerations Introducing AI into an enterprise email environment requires more than deploying a language model. The complete solution should be designed around security, privacy, governance, and responsible AI principles. Important considerations include: Authentication and authorisation The solution should integrate with the organisation’s existing identity infrastructure, such as Active Directory, and enforce role-based access controls. Encryption Email content, attachments, indexes, logs, and other sensitive information should be encrypted both in transit and at rest. Data retention Organisations should clearly define whether AI prompts, generated summaries, and draft content are retained. In privacy-sensitive environments, generated content may remain available only for the active session and should not be permanently stored by the AI application. Auditability The system should record relevant operational events, such as the user identity, requested function, timestamp, and processing status. Depending on the organisation’s policy, the actual email or generated response content may be excluded from audit logs. Human oversight AI-generated responses should be treated as suggestions. The user must review and approve them before sending. The assistant should also disclose its limitations and avoid presenting uncertain outputs as confirmed facts. Performance and capacity On-premises AI processing depends on available compute, memory, storage, and GPU resources. The architecture should account for: Number of concurrent users. Average mailbox size. Volume and size of attachments. Expected response time. Model size. Search-index growth. OCR processing requirements. High availability and backup requirements. A practical adoption approach Organisations do not need to introduce all capabilities at once. A controlled proof of concept can begin with a limited number of users and a defined set of features. A practical implementation sequence may include: Phase 1: Email summarisation Start with individual email and conversation-thread summaries. This provides immediate value while keeping the initial architecture relatively focused. Phase 2: Contextual draft replies Introduce draft generation with mandatory user review and clearly defined tone controls. Phase 3: Attachment processing Enable OCR and document parsing for supported attachments. Phase 4: Intelligent mailbox search Index authorised mailbox content and introduce natural-language search with strict per-user isolation. Phase 5: Wider deployment After security testing, performance validation, and user feedback, extend the solution to additional departments or user groups. Conclusion Organisations operating on-premises Microsoft Exchange do not necessarily need to choose between maintaining control over their email environment and adopting AI-assisted productivity capabilities. A carefully designed on-premises AI email assistant can support email summarisation, contextual draft preparation, attachment analysis, and natural-language mailbox search while keeping sensitive information within the organisation’s infrastructure. The most important success factors are not limited to the AI model itself. Identity controls, mailbox isolation, human approval, auditing, infrastructure sizing, and data-governance policies are equally important. For organisations evaluating private, sovereign, or air-gapped AI, on-premises Exchange can become an effective starting point for introducing practical AI capabilities in a secure and controlled manner. Connect with us to implement this solution in your organization. About the Author Vinay Prakash is the CEO of Glorious Insight and an Ex Microsoft Employee. He works with organisations on Data, Artificial Intelligence, Microsoft cloud technologies, enterprise applications, and secure on-premises AI solutions. For technical discussions related to private or on-premises AI adoption, he can be reached at: Email: email address removed for privacy reasonsUnauthorized: You don't have permission to access this page or perform this action
I frequently encounter an 'Unauthorized' error when attempting to manage Distribution Lists in the Exchange Admin Center, even while logged in with Global Administrator or Exchange Online Administrator privileges. Is this a known widespread issue, or is it specific to the configuration of these individual tenants?The Demise of the OWA Light Client
On July 8, Microsoft said that they will retire the OWA Light client for Exchange Server in August 2026. But what happened to the OWA Light client for Exchange Online? It seems like Microsoft announced the retirement of OWA Light for Exchange Online in June 2024, but didn’t really make the fact clear in a blog post about consumer accounts. In any case, you can’t run OWA Light for Exchange Online, even if you wanted to. https://office365itpros.com/2026/07/17/owa-light-retirement/
How to resolve Junk email validation error
I'm attempting to add an email address to all users mailboxes using Set-mailboxJunkEmailConfiguration. The command fails and returns an email address as the error (see below). I've dumped the three lists contained in the record and the address displayed in the error isn't in any of them. The lists are TrustedSendersAndDomains, TrustedRecipientsAndDomains, and BlockedSendersAndDomains. The command I'm running is: set-mailboxjunkemailconfiguration -identity email address removed for privacy reasons -TrustedSendersAndDomains @{Remove="email address removed for privacy reasons"} The error returns: Set-MailboxJunkEmailConfiguration: |Microsoft.Exchange.Data.Storage.JunkEmailValidationException|Junk email validation error. Value: mailto:email address removed for privacy reasons. Note: all email addresses have been changed. The error happens on a handful of accounts and the value is unique to each user. This error happens on approx. 20 of 5100 mailboxes. Any ideas?The 100 migration batches limit and how to not run into it
As you might or might not know, in Exchange Online we have an upper limit of 100 migration batches. You can see this limit for MaxNumberOfBatches in the Exchange Online PowerShell Get-MigrationConfig: (To understand the MaxConcurrentMigrations limit of 300, you can check this blog post written by one of our migration experts, Brad Hughes.) 100 batches created at the same time should be enough for all O365 multi-tenant customers. But, as we have seen in support, sometimes there are support tickets opened because customers would see the error that says: “The maximum number of migration batches is already running. Please remove a batch before you add another one”. The main reason why this number of batches limit is in place is that having more than 100 batches causes performance problems and could cause outages for other customers in the multi-tenant system. Additional reading on resourcing for mailbox moves can be found here. Let’s say that you want to migrate 50,000 mailboxes to Exchange Online: you could create 50 batches with 1000 users each, and you would still have 50 batches available. If you need to complete all the migrations of users from Batch_1 at the same time, you can easily set -CompleteAfter at the migration batch level. Once this batch is completed, you could delete it to create room for another one. In real world, though, we have seen that our customers rarely want to complete large amount of mailbox migrations at the same time. This is usually the reason why large migration batches are not created but rather smaller batches are preferred. That way, all the migrated users from one migration batch at a time can be completed together. In this scenario, our recommendation is to have large batches for initial sync and if you need to complete the migration for users at a separate date, then you would complete the migration for groups of users in smaller batches, as needed. So how would you create small completion batches for already synced users from larger synced batches? Suppose you want to complete migration of users A, B and C, who reached Synced status and they are all contained in a large batch of 1000 users called Batch_1. To do this, first ensure that you have fewer than 100 batches created at the moment: (Get-MigrationBatch).count Then, proceed with PowerShell commands to get the users from the larger batch Batch_1 into the new smaller batch CompletionABC. Then complete this smaller batch and optionally, remove the completed migration batch: New-MigrationBatch -Name CompletionABC -UserIds EmailAddressOfUserA, EmailAddressOfUserB, EmailAddressOfUserC -DisableOnCopy -AutoStart Complete-MigrationBatch CompletionABC Remove-MigrationBatch CompletionABC Another method we sometimes see people using is setting -CompleteAfter on the move requests directly using Set-MoveRequest. We don't recommend this method because the Migration Service may overwrite the per-request CompleteAfter setting with the one from the batch without notice. This method of individually completing the move requests should only be done when you’ve started the migration through PowerShell, using New-MoveRequest cmdlet, without creating migration batches to contain those move requests. I hope this post gives you a better idea of how to manage migration batches with many users being migrated and not have to run into the limit of number of batches. See you in the cloud! Special thanks to Brad Hughes and Nino Bilic who contributed to this blog post and their precious help in my daily support job. Mirela BuruianaCross-Tenant Message Recall and Duplicate DDG Detection Enhance Exchange Online Messaging
Last week, the Exchange development team announced two interesting developments. Stopping duplicate dynamic distribution groups and cross-tenant message recall. Preventing tenants from creating duplicate DDLs saves service resources and might make tenants easier to manage. The real value is in cross-tenant message recall, which solves a problem by allowing users to recall messages delivered outside their tenant. The facility only works if the receiving tenant allows, but nothing can be done for messages delivered outside Microsoft 365. https://office365itpros.com/2026/07/15/cross-tenant-message-recall/How to Check Distribution Lists for Activity Over the Last 90 Days
The new Get-MessageTraceV2 cmdlet can fetch 90 days of message trace data, which means that we can check for inactive distribution lists using data for the last 90 days instead of being constrained to the last 10 days. This is done by running fetches for nine batches of traffic, each covering ten days. Distribution lists are still important to Exchange Online and Microsoft 365, and it’s good to know which distribution lists are in active use (and how much traffic they get) and which are not. All done with PowerShell. https://office365itpros.com/2026/07/09/inactive-distribution-list-2026/
