← All news
Press · November 18, 2026 · 8 min read

When your AI agents join your team workspaces, do they know what was decided?

When your AI agents join your team workspaces, do they know what was decided?

A team workspace mixes decisions, drafts and proposals. An agent can only tell what was adopted if the document says so.

On 17 November at Ignite, Microsoft introduced Copilot Workspaces: a space where people and agents work together “with shared context”¹. An agent drafts a section, a person reviews it, a comment from one is picked up by the other.

It is a good idea. It also has a blind spot that applies to every company bringing agents into team workspaces in 2027, whatever the vendor: a workspace holds decisions, but also everything that comes before them. Drafts, proposals, working group memos, annotated versions, and now text written by agents. A person usually knows what was adopted. An agent only knows if it is written down.

That is the question of AI agents and working drafts: before opening a workspace to an agent, knowing what it can access is not enough. You need to know whether it can tell what was decided from what was merely proposed. Our position: that distinction is document work, done once, upstream, and every subsequent agent benefits from it.

What changes when an agent joins a team workspace?

An internal assistant connected to a published library (intranet, procedure base, quality repository) reads content that has its flaws, outdated versions and duplicates first among them, but that someone at least published.

A team workspace is different in kind. It is work in progress. You find the purchasing procedure in force, and next to it the working group memo proposing to relax it next year. The minutes of the meeting where the idea was discussed. The draft review sheet an earlier agent wrote based on that memo.

Take an agent preparing the annual supplier review. To assess a supplier’s dependency, it looks for the applicable criterion. It finds two: the procedure’s, which requires a second source for any critical purchase, and the memo’s, more recent, more detailed, written in the present tense. Nothing in the memo says it was never adopted. Depending on its configuration, the agent may retain either. If it retains the memo, a single-source supplier can drop out of the review without anyone noticing.

This is a constructed example, not an observed incident. It illustrates a simple mechanism: in a shared workspace, proximity stands in for legitimacy.

Why can’t an agent tell a draft from a rule?

A colleague relies on signals nobody wrote down: they know who drafted the memo, they were in the meeting, they know the committee never decided. An agent has only the text, and what it has been told about that text.

Existing tools help without settling the question:

  • Versioning in the document management system tells you which version is the most recent. It does not tell you which one was adopted. The latest version of a working memo is still a working memo.
  • Naming conventions (“DRAFT”, a “work in progress” folder) hold as long as everyone applies them, and provided the agent takes the file name or location into account.
  • Relevance engines select what matters for a task. Microsoft highlights this capability for Work IQ, its context layer¹. But relevance and adoption are different things: the most relevant document may be a proposal.
  • Human sign-off before key actions, which Microsoft provides for its autonomous agent¹, covers what the agent is about to do. To spot that a cited criterion comes from a proposal, the person signing off must know the document’s history. Without a written status, every sign-off reconstructs what nobody recorded.

A document’s status, in force, proposed, superseded, is a decision by its owner. If that decision is written nowhere, no tool can read it.

What we see in the field

Our document audits do not focus specifically on team workspaces. They do show how much material in a real corpus should not serve as a reference. On a repository audited by K-AI (anonymised case), more than 1,300 anomalies were detected, around 30% of document volume was removed, and precision exceeded 90% after cleaning.

This measure does not say how many removed documents were working documents; the public source does not specify. It shows that before giving a set of documents to an AI, a significant share may need to leave it, or to carry a status. A team workspace, by construction, holds more work in progress than a published library.

In this work, K-AI counts documents and identifies the applicable version and obsolescence. Linking a rule to the document that carries it requires judgement: the tool proposes candidate documents, business experts decide. Processing covers document content only, within a contractual scope validated with the business owner and the CISO or DPO, with no reuse for model training.

How do you prepare a workspace before an agent joins it?

You do not need to clean the whole workspace. Handle what the agent’s task depends on, before it arrives.

  1. List the criteria that drive the outcome. For the supplier review agent: second source, criticality threshold, compliance. Three to five criteria, taken from its instructions and configuration, are enough.
  2. Link each criterion to a document in force and an owner. For example: second source for a critical purchase → supplier evaluation procedure, version 5, owner: procurement process owner.
  3. Give working documents on the same criteria a written status. A statement in the text itself, “proposal, not adopted”, stays readable whatever tool reads it.
  4. Set the rule for unclear cases. A criterion that exists only in working documents leaves the agent’s scope: it flags the case, a person decides on the sheet, the owner settles the rule.

This list belongs in an annex to the procedure the agent executes, held by the process owner. It may reveal an awkward disagreement: the team may already apply, in practice, a proposal that was never adopted. That is exactly the kind of gap better found before an agent applies it at scale.

The expected effect: an agent that reads the rule rather than the debate, faster reviews, and work that serves the next agent. We have not measured it on team workspaces; it can be checked from the second agent reading the same workspace.

Where does a Document Knowledge Platform fit?

A Document Knowledge Platform (DKP) is the document quality layer that runs upstream of agents, on the documents they consume. It governs (ownership, authority, lifecycle), cleans (duplicates, obsolescence, competing versions), and activates the corpus for AI only once those two steps hold.

It does not replace the collaborative workspace, the agent platform or the search engine. It works on what they read. Workspace models will change; the question of what was decided will remain.

In short

Bringing agents into team workspaces is likely the next step for enterprise agentic AI. Before opening one, audit what the agent will read, clean up the status of the documents its task depends on, then monitor at every revision and every new agent.

Frequently Asked Questions

Does an AI agent read the drafts in a shared workspace?

It depends on its access and configuration. An agent working in a workspace shared with a team can read the documents stored there, drafts included. The real question is whether those documents state their own status.

Isn’t document versioning enough?

It shows the most recent version, not the adopted one. A recent proposal can look more “up to date” than a procedure in force without ever having been approved.

Who should decide the status of a working document?

Its owner, or the owner of the process concerned. The CDO arbitrates when a criterion has no owner. The AI team should not settle a business rule.

Do we need to clean the whole workspace before an agent joins it?

No. Handle the criteria its task depends on. The rest can wait, provided the agent flags cases it cannot link to a document in force.

Who validates the document scope analysed?

The business owner of the documents and the CISO or DPO, jointly, never IT alone. The analysis covers document content only.

Notes

¹ Microsoft, “Copilot at Microsoft Ignite: Expanding what’s possible with intelligence built for work” (Jared Spataro), 17/11/2026, and Satya Nadella’s keynote “Ignite 2026: Building the frontier ecosystem”, 17/11/2026. Workspaces is available in the Frontier early-access programme.

Sources

  • Microsoft Copilot Blog, “Copilot at Microsoft Ignite: Expanding what’s possible with intelligence built for work”, 17/11/2026: microsoft.com
  • The Official Microsoft Blog, “Ignite 2026: Building the frontier ecosystem”, 17/11/2026: blogs.microsoft.com
  • K-AI, Customers page: k-ai.ai/en/clients

Where to Go From Here

K-AI Corpus Diagnostic — 10 business days on your document estate, full report of the 20 most critical anomalies, money-back guarantee if no meaningful anomaly is found. A one-hour conversation helps pin down, for the workspace your first agent will read, which criteria rest on a document in force and which live only in working documents. Reach the K-AI team: contact@k-ai.ai. The scope of every diagnostic is validated jointly by the business Document Owner and the CISO/DPO, never by IT alone.

K-AI already works with CMA CGM, Veolia, PwC, BNP Paribas, TotalEnergies and CEVA Logistics. Partners: AWS, Snowflake, Microsoft, Wavestone, Devoteam.

And in your organization, what does your document estate look like?

30 minutes with a founder. We audit a sample of your documents for free and show you exactly what K-AI detects.

Book a demo → Read other articles