Ask a business agent a question about the business and you quickly learn where its knowledge stops. It answers well about anything that lives in a system it can read, and it goes vague the moment the answer depends on a spreadsheet, a thread, or a decision that was never written down anywhere.
That boundary is not a model limitation. It is a data and permissions boundary, and Microsoft's documentation for the layer that feeds its agents is unusually explicit about both halves.
The Work IQ overview describes it as "a workplace intelligence layer that enables agents to access and reason over organizational and business data, context, and tools", building "a semantic understanding across Microsoft 365 and external systems with built-in, permission-aware governance".
Two phrases carry the weight. The first is organizational and business data: not just files and mail, but, in the overview's own words, "business and workflow data from Dynamics 365 and Power Platform". The second is permission-aware: the layer does not grant the agent new sight. It inherits the sight the signed-in person already has, with their labels, compliance controls and audit trail attached.
That distinction decides what kind of project this is. It is not an AI project with a data appendix. It is a data and permissions project with an AI interface.
The MCP server documentation says the server exposes tools that let agents "read, create, update, and delete Microsoft 365 entities", invoke Copilot for reasoning, and discover schemas through one endpoint. Then it states the control philosophy in four words worth remembering: "Policy over scopes. Four broad OAuth permissions gate capability; fine-grained access control is enforced per path, method, and tenant policy."
Read the right-hand column again. Every one of those failures looks like an AI problem and is actually a filing problem, which is the same reason deployment rather than budget is the real barrier for small business AI.
On cost and packaging the documentation is specific: "Usage is independent of Microsoft 365 Copilot licensing and available through usage-based billing", with "enterprise governance and cost management controls ... available in the Microsoft 365 admin center", and the API overview adds that usage is licensed under the Microsoft Product Terms.
One caution worth carrying into any planning meeting: several summaries circulating this month describe this layer as newly arrived in preview. The current Microsoft Learn overviews carry no preview label, and they describe billing and admin controls as live. For a business, the practical question is not the label anyway. It is whether the two inputs the layer depends on, clean data and reviewed permissions, exist in your company.
Nothing in this approach fixes an undocumented process. An agent reading a system of record reports what the system says, including the parts that are wrong because two people have been doing the same job differently for a year. It also cannot create knowledge that was never captured: if a decision was made in a corridor, no semantic layer will find it.
That is why the honest sequencing for most small businesses is unglamorous. Fix the process, put its state where software can read it, review the permissions, and only then point an agent at it. Projects that run in the other order produce a demo that impresses and a rollout that stalls, which is the pattern behind why small business agent adoption lags enterprise.
The controls that go on top of this layer are a separate and equally necessary piece of work, covered in Google ships agent controls and in the Microsoft material on persistent agents with audit and governance behind them, where the objective and the boundaries are set by the business. Permissions decide what the agent may touch. Rules decide what it may do.
Not according to the documentation. The overview states that usage is independent of Microsoft 365 Copilot licensing and runs through usage-based billing, with cost controls in the admin center.
It inherits permissions rather than bypassing them, which is the point of the design. That also means every over-permissioned folder in your tenant becomes an agent decision nobody consciously made.
Technically, and not usefully. One process with a known answer set will teach you more in two weeks than a tenant-wide connection will teach you in two months.
Then the first project is a data project. That is a real answer and a common one, and it is better discovered at scoping than after a rollout, as described in what AI operations actually requires.
Maxpertise builds and runs the AI team that operates your business, in a managed lane where we host and run the team, or an embedded lane where AI engineers work inside your existing team. Both lanes keep the two human gates: a person approves the plan before the work starts, and a person reviews the result before it ships.
The first deliverable is usually not an agent. It is a written map of one process, its system of record, its exception paths and its permission boundary, because that map is what makes an agent's answers dependable rather than plausible. Once it exists, the measurement side looks like AI operations for business numbers, and the work is scoped the way spec-driven development with AI agents describes. If you want the method, start at how it works; the engagement options are on pricing, and contact with the one process you would most like to stop running yourself is enough to scope it.
[1] Microsoft Learn, Work IQ overview, last updated September 25, 2026. Read directly: "a workplace intelligence layer that enables agents to access and reason over organizational and business data, context, and tools", "built-in, permission-aware governance", business and workflow data from Dynamics 365 and Power Platform, endpoints including a remote MCP server, usage independent of Microsoft 365 Copilot licensing with usage-based billing, and governance and cost controls in the Microsoft 365 admin center.
[2] Microsoft Learn, Work IQ MCP overview, last updated August 21, 2026. Read directly: tools to "read, create, update, and delete Microsoft 365 entities", and "Policy over scopes. Four broad OAuth permissions gate capability; fine-grained access control is enforced per path, method, and tenant policy."
[3] Microsoft Learn, Microsoft Work IQ API, last updated August 21, 2026. Read directly: licensing under the Microsoft Product Terms and usage-based payment.
[4] Microsoft, Introducing the new Copilot with Home, Code and Autopilot, September 25, 2026. Read directly on October 10, 2026: a persistent agent that appears where people already work with "permissions, audit and governance behind it", and the instruction that you set the objective and boundaries.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.