The Knowledge Agent in SharePoint – for KM & innovation leaders | Atlas by ClearPeople

The Knowledge Agent for SharePoint is an exciting new tool, so should you consider this as part of your content and knowledge management strategy?

Microsoft has recently introduced the Knowledge Agent in SharePoint (in preview as of September 2025), a new Microsoft 365 Copilot for SharePoint capability that brings some structured information management capabilities directly into end-users’ hands. The feature makes it possible to automatically generate columns, views, and rules that help organize content in document libraries.

For Knowledge Management and Innovation leaders, this shift is significant. Capabilities that were once hidden behind specialist expertise or IT governance are now moving closer to business teams and content owners. This democratization of metadata and structure is exciting, but it also raises questions about governance, consistency, and the role of KM leaders in guiding how these tools are adopted and scaled. KM and Innovation leaders are right to carefully consider all these aspects, the associated risks, and not to forget, the associated costs.

(If you want to learn about the initial set-up and configuration, please also see: Unpacking the Knowledge Agent in SharePoint – notes for technical and compliance teams)

NOTE: towards the end, I explore the cost of running the Microsoft 365 Copilot Knowledge Agent in SharePoint. The answer is not as clear as you might hope.

Also note: As of September 2025, Knowledge Agent is available in public preview for tenants that opt in and for users with Microsoft 365 Copilot licenses. On 1 November 2025, Microsoft is enabling site-level opt-ins / opt-outs to allow more selective deployment. General Availability is currently planned for early 2026, at which time Microsoft has indicated more comprehensive licensing, governance, and metering models will be published.

Why is Microsoft investing in the Knowledge Agent in SharePoint?

The Knowledge Agent in SharePoint is part of a broader push by Microsoft toward “human-agent teams,” where multiple AI agents (e.g. in Teams meetings, Viva communities) operate together and share context.Microsoft. For example, the Knowledge Agent enriches content in project workspaces, while a Teams agent could surface meeting summaries and assign follow-ups, and a Viva community agent may manage Q&A. These agents interoperate via the Microsoft Graph and Model Context Protocol.

For knowledge leaders, this means your taxonomy, governance, and design decisions in SharePoint will ripple across multiple agent surfaces. Ensuring being able to set context through coherence in metadata, prompt design, and cross–agent consistency will be increasingly critical.

In fact, context is everything in the world of Gen AI. This realisation is hitting most of the big tech providers. Without grounding the AI in relevant and authoritative content, the outcomes are going to be poor.

Microsoft’s goal for the Knowledge Agent is designed to make SharePoint content AI-ready by automatically enriching and organizing files. This – if done right – is a crucial foundational step for AI to deliver accurate and context-aware responses. It also helps maintain a clean and structured environment, reducing the need for manual effort and improving the overall efficiency of content management processes.

Microsoft is hoping that its investment in the Microsoft Copilot Knowledge Agent in SharePoint, will help organizations to ensure that their SharePoint environment is future-proofed for AI, allowing for better collaboration and productivity. Said in other words, Microsoft is betting on this as an enabler to show real value with Microsoft 365 Copilot, which in return provides organizations with a return on their investment.

With this mind, we will take a deeper dive at the Knowledge Agent in SharePoint. In particular, we I will look at how it may address common KM challenges – that if addressed, will drive better AI outcomes across the enterprise.

But first, for those that don’t work with or have little experience with knowledge management and its challenges, let me do a quick recap to set the context from a KM and KM challenges point of view, before we assess the Knowledge Agent’s capabilities.

Defining knowledge management

At ClearPeople, the team behind Atlas, we often hear comments on how we are reimagining knowledge management. This is a great compliment and great for marketing purposes. In this article however, I think it is worth just to quickly clarify what I mean when I refer to knowledge management (KM) in its traditional form.

Simply speaking, KM is a process, which has the objective of capturing content that has or creates value when reused. Typically the KM process includes some variation of the following steps: Capture, Curate, Approve, Distribute, Review, Dispose. There are a number of slight variations of this process and the steps/step-names within it. In reality, they all amount to about the same thing and typically, they are shown as a circular process diagram of some sort.

Irrespective of the exact steps or shape of diagram, the takeaway is that KM is not a tool, a solution or a platform. It’s a process. You don’t buy “KM”, you buy tools, solutions or platforms that support the process of KM.

I venture that most real KMers out there, to lesser or greater degree, will agree with the above.

The common knowledge management challenges

While I am conveniently going to ignore the perennial problem of “tacit knowledge” in KM, effective KM processing is hampered by a number of other challenges. Especially when working at scale in most organizations. Even small organizations will recognize a lot of these challenges.

I have revisited these challenges here, as they are important to consider in order to assess the potential impact, benefits and value of the Knowledge Agent in SharePoint as a solution or a tool for KM.

 

 

Step Issue
Capture Capturing information at the source or in the flow of work often does not happen.
Capture End users forget where to correctly load content, resulting in files ending up in OneDrive or other incorrect libraries and folders.
Curate People are often too busy, or find it too cumbersome and boring, to tag content properly.
Curate When tagging is done, it is often incorrect or inconsistent.
Distribute End users face information overload, mainly with irrelevant content.
Distribute Permissions are often “over shared” to allow easier sharing.
Dispose Multiple copies of content are rarely disposed of.
Dispose Approved content becomes stale and outdated, and thus implicitly incorrect.

Has the Knowledge Agent turned SharePoint into a KM solution?

When testing and reviewing the Knowledge Agent’s capabilities, it is clear that this is a very useful tool in the making.

But as with all technology, there is a risk that marketing messaging and Microsoft-fan-posts lead to setting the wrong end-user expectations and that the shiny new tool’s effect quickly wears off. Some articles, for instance, have boldly concluded that the Knowledge Agent turns SharePoint into a Knowledge Management solution.

I thought Microsoft’s own “getting started” article sets the expectations quite nicely: “Knowledge Agent is a built-in SharePoint capability designed to help your organization prepare content for AI.”

The subsequent sentence unfortunately does quite the opposite: “It enriches, organizes, and maintains SharePoint content in a structured, authoritative format—optimized for Microsoft 365 Copilot agents.”

Clarilis

To me, this sentence is not clear, and it is open to interpretation and counter arguments.

First, it implies that the agent looks across SharePoint to enrich content. It does not. It looks at a library at a time, which means that the enrichment does not take into account a) your global IA or taxonomy or b) how similar content stored elsewhere may have been tagged.

Second, it assumes that the enriched output is an authoritative result. It is not. Unless you are willing to fully trust the suggested metadata as authoritative without validation or approval, which very few will.

Thirdly, it implies that *it* organizes and maintains content. It does not. It suggests – when correctly prompted – actions to users (Owners and Members in the Site, List or Library) visiting that site, list or library who each – and independently of each other – decide whether they agree with the Knowledge Agent’s suggestions.

I get it and it is a very exciting time to work in the crosshairs of tech and knowledge management, but I am super-keen to make sure we leverage the new tools for the job(s) they are meant for.

I know that experienced knowledge management leaders will quickly see through the marketing hype, but like me and many others, will also get excited about the actual opportunities that the Knowledge Agent does in fact bring to the knowledge management toolbox.

So, turning on Knowledge Agent for SharePoint, does not turn your SharePoint into a Knowledge Management solution. It does however offer tooling to solve discreet problems within information silos, which in turn can help the overall knowledge management process outcomes.

Evaluating Knowledge Agent in SharePoint

We can break down the high-level features and preview actions in the Knowledge Agent as follows below.

REMINDER: The Knowledge Agent works within the silo of the current Site or Library the user is in.

Site Pages Fix broken links

Retire inactive pages

Find content gaps

Create and publish pages from site content

Create an FAQ (based on static page content)

Summarize this page (based on static page content)

Libraries “Organize library “

  • Create columns
  • Create a rule
  • Extract key actions
  • Summarize documents

“Create rules”

“Create new view”

Lists Not currently supported in preview.

I am not going to describe the features above in any further detail here, as Microsoft’s own materials do a decent job of that. Also, keep in mind that each of these features above in reality are “prompts” to the Knowledge Agent, meaning you can of course ask the agent to assist with something else.

Instead, let’s consider each of these features/actions above from a knowledge management process perspective.

Step Issue How can Knowledge Agent help Score
Capture Capturing information at the source or in the flow of work often does not happen. Knowledge Agent does not specifically address this. 0/5
Capture End users forget where to correctly load content, resulting in files ending up in OneDrive or other incorrect libraries and folders. Knowledge Agent does not address this and non-SharePoint locations (including OneDrive) obviously do not support SharePoint metadata. 0/5
Curate People are often too busy, or find it too cumbersome and boring, to tag content properly. Here, the Knowledge Agent can really shine, with auto-filling of metadata within document libraries. 3/5
Curate When tagging is done, it is often incorrect or inconsistent. Knowledge Agent will attempt to tag in many cases, but naturally you need to expect that the tags may be incorrect and so will require review. Especially for taxonomy based Managed Metadata data with current limitations. 2/5
Distribute End users face information overload, mainly with irrelevant content. In combination with search interfaces, the Knowledge Agent’s enhancement of content through tagging, will offer more opportunities to deliver relevant and search results. Subject to the metadata being reasonably correct. 2/5
Distribute Permissions are often “over shared” to allow easier sharing. Knowledge Agent does not address this. 0/5
Dispose Multiple copies of content are rarely disposed of. Knowledge Agent can potentially assist in identifying duplicates within a single library. It cannot do this across multiple libraries (or multiple sites). 1/5
Dispose Approved content becomes stale and outdated, and thus implicitly incorrect. Page reviews: Currently Knowledge Agent appears to be handling this decently for pages in a site (possibly due to available analytics). 3/5
Document review: prompting the agent to identify outdated or stale documents (in various variations of the prompt) only resulted in a suggestion to create a (new) column to track modified date (which is already a column). 1/5

My scores are obviously completely arbitrary, based on limited testing and I recognize that Knowledge Agent for SharePoint is in preview. So naturally, I am happy to take feedback for further review.

From the table, I want to call out the two excellent features that both score 3/5:

  • Auto-filling: despite lots of teething issues (and nasty defaults being set for auto-generated columns) this is where the real power of the Knowledge Agent comes to the fore and shows a great potential. The score of 3/5 may be a bit generous at this stage, but the opportunity here is fantastic as the capability matures. It will however be vital to appreciate how to manage and deploy this capability at scale in enterprise scenarios. (Note that the quality of the tags themselves score 2/5.)
  • Page reviews: I feel like this capability might stretch to a 4/5 as the initial testing looked really good and pretty well baked. It will need proper testing on larger sets of real site pages with good content, a decent user base and comparable usage statistics to validate degree of false (positives/negatives).

I also want to highlight the most disappointing outcome: Why has Microsoft not doubled down on one of the most common admin problems for content managers: Duplication of content. Considering that the scope of the Knowledge Agent is restricted to the silo of a library, having a simple pre-canned and optimized tool the agent can call to check for duplicates – within that library – seems like a huge miss.

Risks associated with leveraging the Knowledge Agent

Ignoring any risks related to leveraging preview or maturing technology, let me highlight some key concerns I raised in my earlier blog here. Naturally, we expect that some of these may improve as the Knowledge Agent matures to General Availability and beyond. So, as of today:

  • There is no strong control over column naming conventions. Site members can accept suggestions; there are few (if any) constraints at the moment over enforcing naming standards.
  • Column creation via Knowledge Agent appears to create columns within the default content type. This may not align with custom content types or existing taxonomy.
  • Versioning: when metadata autofill is applied or updated, changes occur in current item version. There is no rollback option. So, if metadata is “over-enriched” or mis-applied, it may not be an easy undo.
  • Auditability is limited: it will be hard to trace who accepted what suggestion, when a column is added, or when metadata is filled in, especially if many site owners/members do this independently of each other.
  • There are a number of limits to metadata features: e.g. managed metadata term sets are only partially supported (only the first 100 terms are considered in a term set) in the current version. That can limit accuracy and consistency if term sets are large.
  • There is a lack of cost transparency and lack of cost controls, which will cause concern for budget holders.
  • Because Knowledge Agent will generate suggestions based on patterns in existing content, there is risk of reinforcing incorrect or inconsistent metadata decisions, especially in low-data or low-structure sites. Human oversight is essential. With preview tooling, there is no built-in audit trail or rollback capability, so organizations should pilot carefully and impose review checks for agent suggestions.

As a consequence, a number of risks materialize. I have listed out the most obvious, but in reality, in addition to risk related to costs, most of them boil down to lack of control, inconsistency and quality. We are even starting to talk about a new type of information silo, The “Metadata Silo”, being introduced due to clusters of metadata being created within the silos of the document libraries.

Risk Description Consequences
Divergent column names (naming inconsistency) If different site teams accept suggestions or create new metadata columns in their libraries, those columns may have different names, spellings, casing, punctuation etc. E.g. “ClientName”, “Client Name”, “Customer Name”, “client_name” etc. Also, display name vs internal name mismatches. Problems for search and discoverability across sites; users may not find content if they search under different names. Difficult to build reports, dashboards, policies that assume standard metadata fields. Higher support / maintenance cost. Potential for duplicate or redundant metadata fields.
Inconsistent data types, formats, or column settings Even where naming is similar, different libraries might have columns of different types (single line text vs choice vs managed metadata vs date etc.), different default values or different validation. Inconsistent filtering, sorting, grouping in views; broken or non-uniform behavior; poor user experience; difficulty in applying enterprise rules; risk of data integrity issues (e.g. date formats or locale issues).
Overlapping / redundant metadata fields Multiple fields might cover overlapping concepts, but people don’t know which to use; e.g. “Department” vs “Dept”; “Project Code” vs “Proj Code”. With the Knowledge Agent suggesting new columns “on the fly”, redundant fields are likely. Confusion for content creators/users; metadata fragmentation; inconsistent tagging; drift in how fields are used; difficulty aggregating or integrating metadata; duplication of effort.
Inconsistent tagging / classification Similar content might be tagged differently across sites, or tags might use free text vs controlled vocabulary vs managed metadata vs no standard taxonomy. Some content may not get tagged properly or gets mis-tagged due to AI misclassification. Poor search or Copilot responses (which rely on metadata for “grounding”); difficulty in running enterprise reporting; compliance risks (e.g. regulatory or records management) if classification or retention policies rely on metadata; risk of data silos.
Unexpected overwrites / “drift” over time Autofill or metadata autofill may overwrite existing metadata (if run automatically) or change behavior when content changes. Also, when column suggestions are revised, new versions of content may diverge. Loss of previously entered metadata; inconsistencies between content versions; confusion when the same document looks different in different contexts; potential audit issues; user distrust if metadata changes without clear visibility.
Lack of governance / oversight As noted in early commentary, Knowledge Agent gives site owners and members power to create columns, accept suggestions, set up views / rules etc., often with limited controls (e.g. little or no control over column naming, limited audit trails) in preview. Potential for inconsistent practices, accidental non-alignment with enterprise architecture or taxonomy, difficulty enforcing compliance, higher risk of information chaos rather than order.
Scalability complexity and coordination issues In large organizations you will have many sites, many authors, many types of content. Without central coordination, what works in one site may not in another, and the number of metadata fields and variations may explode. Search, discoverability, knowledge management degrade; user confusion; technical debt in metadata; burden on governance teams later; risk of rework or migration costs.
Localization / multilingual inconsistency If metadata display names or column labels are in different languages, or formats differ by locale, this might further complicate consistency across global sites. Users in different regions may misinterpret, search fails; duplication of metadata fields; inconsistent filtering or translation issues.
Unclear costs and lack of cost control While the use of the Knowledge Agent in SharePoint currently is included in the M365 Copilot user license, the processing of any Autofill columns is a metered PAYG service. With very limited ability to control where and how many Autofill columns get configured in libraries, the risk of significant spend is real.

Read more about this in the cost section below.

Leveraging the Knowledge Agent for SharePoint in enterprise scenarios

Irrespective of all the concerns and the risks (and ignoring costs which I touch on below), the fact remains that the Knowledge Agent is on track to become a very powerful tool in the right hands and when employed sensibly.

Whether you are a local, regional or global law firm, a small or large professional services business or a global multinational multidisciplinary firm of professionals, you will be dealing with “enterprise scenarios”.

Real world business is not confined to single SharePoint sites or libraries, and as knowledge management leader it is vital that your information architecture, your taxonomy and your filing experiences are as easy and consistent as possible. This is what gives you a fighting chance to deliver great output.

So, how might you consider leveraging the Knowledge Agent within your content or knowledge management platform design?

This is clearly a topic which is rich enough for a separate article, but let me highlight a few key pointers:

The more autonomous / free-will your front-of-house is, the more you need rigor and discipline in the backend to avoid tagging chaos, metadata silos and unintended loss of context.

In other words, you will need a solid enterprise-wide approach that includes:

  • Taxonomy management, offering clarity while being flexible to change
  • Provisioning of Teams and Sites, while being business centric
  • Template based governance (from site down to lists and library/folder level)
  • Self-service approach to “bring people in”, rather than going rogue with various agents.
  • Implement simple and speedy approaches to approvals where possible
  • Create an accountability framework and “safe spaces” to share mistakes or mishaps.

Understanding the cost

Today, in preview, the usage of the Knowledge Agent in SharePoint is included with Microsoft 365 Copilot license. Whether this shifts to a usage-based model later is not clear. (Please see Roadmap ID: 503145 which indicates support for Pay-As-You-Go for SharePoint Agents, which is slated for General Availability in October 2025.)

However, if my understanding is correct, the actual Autofill processing (i.e. the extraction of metadata) is based on usage per Autofill column and per document page and is therefore an additional cost.

Let us look at a specific use case:

Alice has a Microsoft 365 Copilot license and is an owner of a SharePoint collaboration area with a library that contains 1,000 contract agreements. The contracts are PDF documents with an average of 20 pages each.

Alice uses the Knowledge Agent to organise the library with three new columns (a Summary, a Contract Renewal Date, and Country of Use). Each of these are automatically suggested by the Knowledge Agent with a suggested Autofill prompt. Alice now makes a few minor adjustments and submits, thereby modifying the library settings to include three three new columns.

Alice then gets the help from the Knowledge Agent to set up a rule (to get updated when new contracts are uploaded)

Finally Alice uses the Knowledge Agent to suggest a new View for her colleagues to easier locate contracts.

Alice’s use of the Knowledge Agent to organise the library, to create a rule and to create a view is included within Alice’s M365 Copilot license (although see note above re future PAYG agents).

However, as soon as Alice submits the suggested columns (which use Autofill) the actual Autofill processing of each document is now running as a metered Autofill service on a PAYG basis.

The cost of Autofill running once against the 3 columns for the 1,000 documents in this one single library can be calculated as:

3 Cols * $0.005/page * 20 pages/document * 1,000 documents

Total cost: USD 300 to run once in the library

Naturally, if Alice or someone else manually selects the same documents and chooses to run the Autofill again, then this is additional usage and cost.

When new documents are added, it is naturally also additional cost.

Looking at a firm-wide use case with a 1000 people, each saving only one new 20 page document per working day, into similar libraries, we can estimate a monthly Autofill processing cost as follows:

3 Cols * $0.005/page * 20 pages/document * 1000 users * 1 document/user/day * 20 days/month

Total cost: USD 6,000 per month

This is of course in addition to the M365 Copilot cost per user.

I won’t be the judge of whether this is a reasonable cost or not. I will however say that when organisations don’t feel they are in control over the costs and the costs lack predictability and transparency, then any cost outside of what’s expected will be unwelcome.

Closing thoughts

As with most other tools, the Knowledge Agent for SharePoint is not a silver bullet for knowledge management, but it does represent an important evolution in how AI can support KM processes.

For Knowledge Management and Innovation leaders, the real opportunity lies in shaping how these emerging tools are governed and integrated into enterprise knowledge strategies.

Used wisely, the Knowledge Agent can likely – as it matures – play a role to reduce friction in curation, improve findability, and support better content hygiene. Left unmanaged, it risks creating metadata silos and increasing complexity. This will however be subject to Microsoft being clear about the cost profile and allowing firms to set and predict their spend.

KM leaders who start experimenting now, with governance guardrails in place and the right enterprise-wise platform architecture in place, will be better positioned to shape, rather than chase, how AI transforms enterprise knowledge practices.

The challenge, and the opportunity, for KM leaders is to strike the right balance between enabling business users and safeguarding enterprise-wide consistency, while keeping a keen eye on cost-creep across a plethora of AI services. Those who do will not only harness the potential of the Knowledge Agent, but will also strengthen the role of KM as a driver of innovation and productivity.

Atlas sets the standard for instant knowledge and precise AI in legal, transforming Microsoft 365 into an intelligent knowledge hub.