AI in SharePoint, Formerly Knowledge Agent: What It Does and How To Roll It Out

SharePoint's AI layer now sits inside the same content system that already controls sites, pages, lists, document libraries, metadata, permissions, versions, retention, and search. A prompt can therefore trigger work against objects that have existed for years: a library with inherited permissions, a policy page last reviewed 18 months ago, a list with inconsistent columns, or a document set whose metadata was never completed. The quality of the AI experience is tied directly to the condition of that SharePoint estate.

The feature also moved through several names in less than a year. Microsoft introduced Knowledge Agent in September 2025. In March 2026, those capabilities became AI in SharePoint. Microsoft's August 2026 administration documentation now calls the preview Copilot in SharePoint. The underlying direction is consistent: natural-language assistance is being placed inside SharePoint so users can ask questions, create structures, enrich content, configure views and rules, and reuse site-specific instructions without leaving the content context.

For IT teams, this makes Copilot-Ready SharePoint a practical engineering target. The rollout begins with site scope, permissions, content quality, metadata, ownership, and user behavior. The AI button comes later in the sequence. A tenant with clean access boundaries and useful content structure gives Copilot better material to work with. A tenant with stale sites and broad sharing gives it the same material, only much easier to retrieve.

TL;DR

The Name Changed, and the Platform Scope Grew With It

Knowledge Agent began as an in-context SharePoint assistant focused on content cleanup, metadata enrichment, library organization, page work, and questions grounded in SharePoint content. Microsoft's March 2026 SharePoint product update moved that work into AI in SharePoint and described a broader building experience for sites, pages, libraries, lists, and other content objects.

The current Microsoft Learn pages use Copilot in SharePoint for the preview. That naming change matters for administrators because older PowerShell controls still contain KnowledgeAgent in parameter names. An admin can therefore see three labels across announcements, documentation, and command syntax while managing one evolving capability set.

The useful boundary is SharePoint itself. The AI works where content is stored, structured, shared, and governed. That means a rollout plan has to account for the surrounding Microsoft 365 architecture as well. Files reached through Teams may live in SharePoint. Site permissions can be backed by Microsoft 365 groups or Entra ID groups. Retention and sensitivity can come from Purview. Workflows can cross into Power Automate. Guidance on SharePoint and Microsoft 365 integration is relevant here because the content path often crosses several Microsoft services even when the user starts from one SharePoint page.

What can a Copilot in SharePoint actually do?

The preview is broader than a question-and-answer panel. It can read the context of the site or library, propose changes, create structures, and carry out supported content operations after the user reviews them.

Ask questions against the content in front of you

Users can ask questions grounded in SharePoint content they can access. This works best when the site contains a clear source of truth and when old or duplicate material has been removed or identified. A policy library containing four competing versions can still produce an answer. The harder problem is deciding which version deserves authority.

Build sites, pages, lists, and libraries from natural language

A user can describe the workspace or content structure they need. Copilot can plan the structure, propose components, and create supported SharePoint objects. For document libraries, Microsoft documents an AI-assisted flow that can generate the library name, description, metadata schema, and AI-powered columns from a plain-language description and an optional sample file.

Generate and apply metadata

AI-powered columns can extract information from supported documents and populate library fields. This turns SharePoint Metadata Management into an active content-processing step. A contracts library, for example, can use fields for counterparty, agreement type, effective date, renewal date, and owner. The business still needs to define which fields matter and how they will be used in search, views, retention, reporting, or workflows.

Create views and rules from a prompt

Copilot can build library views by showing, hiding, sorting, grouping, and filtering supported column types. It can also configure supported rules from natural language. During preview, Microsoft documents triggers for item creation, modification, and deletion, with actions such as email, move, copy, and set value. These capabilities make library design faster, but the rule set still needs an owner who understands what happens when documents change.

Save repeatable instructions as skills

Site-specific skills let users capture a repeatable multi-step instruction and save it for reuse. The skill is stored as a managed asset in the site and can inherit ordinary SharePoint governance such as permissions, retention, sensitivity, and auditing. This is useful when a team wants the AI to follow a consistent review checklist or document standard instead of rebuilding the instruction every time.

AI-Ready SharePoint Starts With the Content Layer

AI retrieval depends on the content it can reach. Box's 2026 enterprise AI content research surveyed 1,640 IT decision-makers and found that 96% said agents need access to company-specific content, while only 36% had connected agents to trusted internal content across many use cases. The gap is familiar to SharePoint teams: content exists, yet trust, structure, ownership, and access are uneven.

A readiness review should classify the condition of each target site before users receive broad AI access. SharePoint document management controls provide the same foundation from the document side: metadata, version history, retention, ownership, and lifecycle all affect whether a library behaves like a dependable knowledge source.

Permissions Still Decide What the AI can Surface

Copilot in SharePoint works inside the permissions already attached to the user and content. That preserves SharePoint's access model. It also increases the practical effect of old oversharing because a user can retrieve relevant material through natural language instead of knowing the exact site, library, folder, or file name.

Varonis analyzed nearly 10 billion files for its 2025 State of Data Security and reported that 99% of the organizations in the study had exposed sensitive data that AI could potentially surface. Its Microsoft 365 findings also point to broad employee access and weak labeling as recurring issues. The exact exposure level will differ by tenant, but the rollout lesson is direct: SharePoint Permissions for Copilot deserve an evidence-based review before large user groups receive the feature.

That review should cover organization-wide sharing, guest accounts, anonymous or "anyone" links, broken inheritance, direct user grants, old Microsoft 365 groups, inactive site owners, sensitive libraries, and sites with unusually high sharing activity. Access remediation should preserve legitimate work patterns. A finance site with broad readership can be valid. A payroll working library with the same audience needs a different decision.

At Calance, our SharePoint governance framework for 2026 treats permissions, ownership, lifecycle, labels, applications, and exceptions as connected controls. That same model fits SharePoint AI Governance because an AI rollout exposes the quality of those relationships very quickly.

Score Each Pilot Site Before it Enters the Rollout

A simple site score can stop a pilot from turning into an estate-wide cleanup project. Give each target site 0, 1, or 2 points for the checks below. A score of 2 means the control is documented and verified. A score of 1 means it exists with gaps. A score of 0 means the team cannot prove it.

1. Ownership: current business owner and technical owner are named.
2. Permission model: high-risk sharing paths and broken inheritance have been reviewed.
3. Source authority: users can identify which library, page set, or list is authoritative.
4. Metadata: the fields used for search, filtering, retention, or workflows are populated reliably.
5. Content age: stale pages and documents have a review or archive path.
6. Sensitivity: restricted content is classified and handled under known policy.
7. Lifecycle: the site has a purpose, review date, and closure rule.
8. Use case: the pilot group can name recurring tasks where SharePoint AI should save time or reduce search effort.
9. Support owner: somebody will review failed processing, poor answers, rule behavior, and user questions during the pilot.
10. Measurement: the team has a baseline for search success, task time, content quality, or another outcome.

A site scoring 16 to 20 is a strong pilot candidate. A site scoring 11 to 15 needs targeted cleanup before user testing. A site below 11 belongs in remediation first. The score is a triage tool, so every point should have evidence behind it.

Roll Out by Site Cohort, with a Reason for Every Cohort

McKinsey's 2025 State of AI survey found that nearly nine in ten respondents said their organizations were regularly using AI, yet almost two-thirds had not started enterprise-wide scaling. That difference between access and scaled operating practice is useful for SharePoint Copilot Rollout planning.

• Choose a bounded knowledge problem. Start with sites where content has clear business ownership and users already search or classify information frequently. Policy libraries, project knowledge sites, controlled document repositories, and well-managed intranet areas can produce useful pilot evidence.
• Audit the sites before licensing the cohort. Review permissions, stale content, metadata completeness, external sharing, owner status, and known search failures. SharePoint consulting and assessment services use this type of current-state review before changes are planned across a SharePoint environment.
• Define the allowed AI jobs. Name the tasks users should test, such as finding policy answers, creating a library, generating metadata, setting a view, or applying a repeatable skill. This keeps feedback tied to business work.
• Run a small user cohort across real content. Include site owners, frequent contributors, ordinary readers, and at least one person who understands the site's security or records requirements. Different roles expose different problems.
• Record failures as content or control defects. A weak answer may come from stale source material. A missing answer can reveal permissions or indexing issues. A bad generated column can reveal that the document type was never modeled properly.
• Expand only after the site has an operating owner. Each wave needs someone who can review content drift, permissions, usage, skills, library rules, and user feedback after the initial launch.

A Practical 60-Day Rollout Pattern

The preview can be introduced in a short cycle if the first scope is small. The schedule below treats deployment as a controlled site change with a bounded site cohort.

The calendar can stretch for regulated or very large environments. The sequence matters more than the number of days: inspect, repair, test, observe, adjust, then expand.

Metadata Becomes Operational Once AI can Create and Fill It

Metadata has often failed in SharePoint because humans see fields as extra form work. AI-powered columns change part of that equation by extracting values from documents and populating fields automatically. The business still has to decide which metadata is worth maintaining.

Proofpoint's 2025 data security research found that 46% of surveyed organizations cited cloud and SaaS data sprawl as a top security challenge, while 31% identified redundant or obsolete data as a significant risk. For SharePoint Metadata Management, the practical response is selective structure. A small set of fields that drive retrieval, lifecycle, reporting, or rules is more useful than a large taxonomy that lacks an owner.

Use AI-generated metadata where the field can be checked against a clear business meaning. Contract type, department, document status, project code, effective date, region, policy owner, or review date can all support useful downstream work. Free-form interpretation needs closer review when the field carries legal, financial, or compliance consequences.

Content placement matters too. A file stored in the wrong workspace can carry excellent metadata and still have the wrong owner or lifecycle. SharePoint versus Teams versus OneDrive model is useful when the rollout uncovers content that should move from personal or project storage into a durable SharePoint source.

Pick Early Use Cases where SharePoint Context Already has Value

Early use cases should involve work SharePoint already supports well. That makes it easier to tell whether the AI is improving the task or merely adding another interface.

• Policy and procedure retrieval: Ask questions against a controlled policy site where owners, review dates, and source documents are already known.
• Document intake: Create metadata-rich libraries for invoices, contracts, onboarding documents, project deliverables, or other repeatable document classes.
• Library cleanup: Generate useful columns, create views, and identify better grouping patterns for repositories that users currently browse manually.
• Page maintenance: Assist site owners with content updates where the page has a clear owner and approval path.
• Operational rules: Build simple notifications, moves, copies, or value-setting rules around defined document events.
• Reusable review skills: Save a site-specific checklist for recurring tasks such as document quality review, content preparation, or structured analysis.
• Structured document generation: Use approved Word templates and forms for documents that follow a repeatable structure and need consistent output.

High-risk repositories can enter later waves after the team understands how the feature behaves in its own tenant. Payroll, legal investigations, board material, merger work, regulated records, and highly restricted research often deserve deeper access and governance testing before AI-assisted workflows are introduced.

Preview Administration Needs Explicit Owners

The current preview has controls and limits that should appear in the runbook. Microsoft states that licensed users receive Copilot in SharePoint as an opt-out preview as the capability reaches their tenant. Administrators can manage availability at tenant or selected-site level through SharePoint Online PowerShell. The command parameters retain KnowledgeAgent naming during preview.

Site scope

KnowledgeAgentScope supports all sites, selected inclusions, selected exclusions, or no sites. The selected-sites list has a documented limit of 100 URLs. For a controlled pilot, an inclusion list gives the team a clear boundary while permissions and content are being tested.

Licensing and cloud availability

Users need an eligible Microsoft Copilot license for the preview. Microsoft currently excludes several government and sovereign cloud environments from support. Multi-geo tenants also need the administration step handled in each geo where applicable.

Usage and model behavior

Microsoft documents per-user daily and weekly usage limits during preview. The current August 2026 Learn page says Copilot in SharePoint uses a Microsoft-managed reasoning model from OpenAI and that Microsoft may change the model as the service changes. The organization therefore governs the service boundary and its own data controls, while Microsoft manages model selection inside the product.

Feature limits

Library views created by Copilot are public views during preview. Rule creation supports a defined set of triggers and actions, with a limit of 15 rules per list or library. Some file-processing features have language and encrypted-file restrictions. These details belong in user guidance because a feature can behave correctly and still fall outside what a user expected it to do.

Adoption Needs Workflow-Level Teaching

A feature can be technically available while daily behavior stays unchanged. WalkMe's 2026 digital adoption study surveyed 3,750 executives and workers across 14 countries. It reported that 54% of workers had bypassed AI tools and completed tasks manually at least once in the prior 30 days, and workers lost the equivalent of 51 working days a year to technology friction.

For SharePoint Copilot Readiness, training should use the library, page, or site the person already works in. Show a records coordinator how to test metadata extraction on the documents they receive every week. Show a communications owner how to update a page and verify the source content. Show a project team how a skill captures its review checklist. Then document where human approval remains required.

Adoption measurement can borrow from ordinary SharePoint behavior. SharePoint intranet adoption metrics focus on return behavior, search success, task completion, content trust, and participation. Those measures can be adapted to AI-assisted work without reducing success to prompt counts.

Measure Usefulness Before the Next Rollout Wave

Flexera's 2026 ITAM research findings reported that only 31% of organizations had visibility into AI software, only 29% measured AI software value, and 59% said wasted AI spend had increased. A SharePoint pilot already has a smaller boundary than the whole AI estate, so it is a good place to build measurement discipline early.

Use a small scorecard with baselines and post-pilot results:

• Retrieval quality = successful answers from approved sources / tested knowledge questions
• Metadata acceptance = AI-populated fields accepted without correction / fields reviewed
• Task time change = baseline task time - AI-assisted task time
• Return use = pilot users who come back to the feature in a defined period / enabled pilot users
• Content repair rate = content defects fixed after AI exposed them / defects logged
• Permission defect rate = access issues found during pilot / sites reviewed

A rising prompt count can accompany poor outcomes. Tie expansion to the measures that describe work quality, content quality, and control quality. Copilot readiness and rollout framework also connects rollout planning with governance and KPI tracking across Microsoft 365 Copilot programs.

Five Rollout Failures that Show Up Early

• Turning on broad site access before the content audit: The first wave then becomes a permission and stale-content discovery exercise. Run the discovery work first and use AI testing to find the remaining defects.

• Treating generated metadata as final truth: Extraction can reduce manual tagging effort, but fields still need business definitions, sample testing, and review rules. High-consequence fields deserve stronger checks.

• Letting every site invent its own skills and rules: Local flexibility can produce duplicate or conflicting instructions. Define which teams can create reusable skills, where ownership sits, and how old assets are retired.

• Measure completed work: Activity tells you the feature was opened. Outcome measures tell you whether search, classification, content creation, or document handling improved.

• Leaving pilot sites in permanent exception mode: Temporary inclusion lists, pilot permissions, experimental skills, and test rules need review dates. Every exception should either become part of the operating standard or be removed.

The Operating Model After Rollout Should Stay Small

SharePoint Site Lifecycle Management becomes more important after AI access grows because site quality affects retrieval quality. A light operating model can keep ownership visible without creating a new committee for every library.

Restricted Content Discovery SharePoint controls and other SharePoint Advanced Management features can also support higher-control estates where search and Copilot exposure need additional boundaries. The exact control set should follow licensing, regulatory needs, and the sensitivity of the site population.

Build the Content System First, then let AI Work Inside It

AI in SharePoint gives users a much more direct way to work with the platform. Natural language can create libraries, add structure, build views and rules, answer questions, and reuse site-specific skills. Those capabilities reduce several forms of manual SharePoint work, especially where the content model is already clear.

The rollout also makes existing SharePoint quality easier to see. Old permissions become retrieval paths. Weak metadata becomes vague context. Duplicate policies become competing answers. Unowned sites become maintenance problems. Copilot-Ready SharePoint therefore has a concrete definition: content has an owner, access has a reason, metadata supports real work, stale material has a lifecycle, and the team can measure whether AI-assisted tasks are producing better results.

Start with a small site cohort. Repair the content and access model. Test real tasks with real users. Keep the preview controls in the runbook. Expand when the evidence supports the next wave. That sequence gives Copilot in SharePoint a dependable knowledge layer to work against and gives IT a rollout it can explain months after the first pilot ends.

Frequently Asked Questions

1. Is AI in SharePoint the same product as Knowledge Agent?

Knowledge Agent was the earlier preview name. Microsoft renamed the capability AI in SharePoint in March 2026, and current Microsoft documentation now uses Copilot in SharePoint. Some PowerShell parameters still retain the KnowledgeAgent name during preview.

2. What license is required for Copilot in SharePoint?

During the current preview, Microsoft requires an eligible Microsoft Copilot license for users who access Copilot in SharePoint. Organizations should verify their tenant, cloud environment, license assignment, and current Microsoft terms before planning a rollout cohort.

3. Can Copilot in SharePoint read files a user cannot normally access?

Copilot in SharePoint follows the user's existing access rights. The bigger rollout concern is old oversharing, broad groups, guest access, and stale sharing links that already grant access and can make sensitive content easier to retrieve.

4. Should every SharePoint site be enabled at the same time?

A site-cohort rollout gives IT better control over permissions, content cleanup, support, and measurement. Start with well-owned sites and defined use cases, record defects, repair them, and expand after the pilot shows stable results.

5. What makes a SharePoint site Copilot-ready?

A strong site has current owners, intentional permissions, known authoritative sources, useful metadata, maintained content, sensitivity controls where needed, and a lifecycle rule. It also has a business use case and a measurable baseline for pilot testing.

6. Can Copilot create SharePoint document libraries?

Yes. Microsoft documents an AI-assisted library creation flow that can use a plain-language description and optional sample file to generate a library name, description, metadata schema, and AI-powered columns for supported document processing scenarios.

7. What are AI-powered columns in SharePoint?

AI-powered columns extract information from supported files and populate SharePoint metadata fields. They can reduce manual tagging work, but teams should define field meaning, test extraction quality, and review fields used for compliance or high-consequence decisions.

8. Can Copilot in SharePoint create views and workflow rules?

Yes. During preview, Copilot can create supported library views and rules from natural-language instructions. Microsoft documents limits on view behavior, rule triggers, rule actions, file processing, and the number of rules allowed per list or library.

9. What are skills in Copilot in SharePoint?

Skills are reusable site-level instructions for repeatable multi-step work. They can capture a review checklist or local document standard and are stored as SharePoint assets that can use ordinary permissions, retention, sensitivity, and auditing controls.

10. How long should a SharePoint AI pilot run?

A focused pilot can produce useful evidence in roughly 30 to 60 days when the sites are already reasonably governed. Larger or regulated environments may need longer for access review, content remediation, legal checks, and representative user testing.

11. What metrics should we track during rollout?

Track successful retrieval from approved sources, metadata acceptance, task-time change, returning users, content defects found and fixed, permission issues, support volume, and user confidence. These measures describe outcomes more clearly than raw prompt or license counts.

12. Does Copilot in SharePoint fix poor metadata automatically?

It can propose and populate metadata through AI-powered columns, which reduces manual work. The organization still owns the schema, field definitions, taxonomy, exceptions, quality checks, and downstream rules that depend on those metadata values.

13. Which SharePoint sites should enter the first rollout wave?

Choose sites with clear ownership, repeatable knowledge tasks, manageable permissions, maintained source content, and users willing to test real workflows. Avoid beginning with the most restricted repositories unless the pilot specifically targets those security and compliance controls.

14. How often should SharePoint AI governance be reviewed after launch?

Site owners can review content and AI behavior monthly, while administrators and security teams can review permissions, sharing, ownership, site scope, and usage quarterly. Review again before each rollout wave and whenever a high-value site changes ownership.