An assistant produces an excellent report in a minute during a demo. The next day, an employee opens the chat again, attaches files, explains the assignment and transfers the answer into a working document. The result is convincing, but preparation and handover still depend on a person.
Enterprise AI updates in 2026 are tackling these steps: starting work without a new prompt, completing several actions and turning meeting notes into tasks. Here are five Microsoft and Notion releases announced by 27 September, with examples of where they may help. This is a selection of workflow updates, not a comprehensive review of the market or new models.
Three key points
- 01
Agents are gaining event triggers, multi-step assignments and shared team instructions.
- 02
Automatic task creation requires checking the original agreement.
- 03
After a model update, compare results on the same working examples.
Agents can start from an event or a schedule
On 24 February 2026, Notion introduced Custom Agents, which can perform configured work on a trigger or schedule, such as handling requests and preparing reports. Employees do not have to reopen a chat and repeat the assignment every time.
Convenience can quickly become noise without clear delivery rules. A daily digest covering every project may be another unread message. A more useful report might cover only changes requiring a decision: a blocker, a moved deadline or an assignment without an owner.
Set an owner, schedule, sources and stopping rule before launch. Test weekends, unavailable sources and changes to employee permissions. A missed run must be distinguishable from a run that found no changes.
Office assistants take on multi-step assignments
On 16 June, Microsoft announced general availability of Copilot Cowork. Its description centres on assignments requiring several actions using Microsoft 365 work context.
Consider preparing for a management meeting. The assistant must find reports, identify discrepancies, prepare a document and submit it for review. There are several opportunities for error between the request and the finished briefing. A manager needs to see where work stopped, which sources were unavailable and what remains an assumption.
An intermediate plan gives the reviewer a chance to stop a mistaken approach before it produces dozens of pages. Internal documents and external correspondence also need different approval rules.
Meeting notes can trigger the next action
On 31 July, Notion announced that AI Meeting Notes can trigger Custom Agents. Examples include updating project trackers, sending recaps and turning feedback into tasks.
A project manager may no longer need to transfer every item into a tracker after a call, provided that connection is available and configured. But a mistake in a summary can now become an actual assignment. Reviewing the wording becomes part of assigning work.
“We can probably finish on Friday” should not automatically become an approved commitment. Proposals and accepted decisions need separate statuses. Processing a meeting again should update or preserve an existing task, not duplicate it.
Teams can save reusable instructions
On 15 September, Notion 3.7 introduced a shared skills library. Teams can store reusable instructions, such as the structure of a management memo or the steps for reviewing a product proposal.
This helps when results depend on an employee who knows how to phrase the request. Shared instructions make expectations explicit: required fields, how to mark unknown information and which actions need approval.
An instruction still needs an owner, a version and examples of acceptable output. Otherwise, the organisation replaces contradictory chats with contradictory instructions.
Model choice becomes an administrative decision
On 9 September, Notion added model availability controls, letting workspace owners manage choices separately for Notion Agent and Custom Agents.
Extracting a date from a short message and preparing a complex analytical briefing place different demands on a model. An expensive model may be excessive for the first; a cheaper one may create too many corrections for the second. Both costs matter.
Keep the test set consistent: the same sources, constraints and expected results. After an update, compare errors, cost per accepted output and the need for human intervention. Better results on a vendor's general benchmark do not guarantee improvement in your workflow.
What these changes mean for Executive AI
These releases show a shared direction: vendors are connecting model output to meetings, tasks, documents and recurring reports. That does not mean any agent can handle any company's processes. It does change the questions worth asking at a demo.
Executive AI's approach addresses the same management problem: preserving the path from a meeting to a decision, task, owner and result. Test it through a single project history: another meeting takes place, the deadline changes and a new employee takes over. Can the manager identify the current agreements without rereading every message?
An individual feature is unlikely to remain exclusive. Several vendors are developing search, minutes, memory and automated actions. Selection depends on how well the specific workflow performs, how practical deployment is and what ongoing support costs.
Which updates are worth adopting?
Start with work the team already does each week: a project report, preparation for a client call or a review of open assignments. Identify which new capability removes a manual step, then test one workflow on a limited dataset.
A useful update lets an employee finish faster or more reliably without review consuming the entire gain. If the output reads better but requires the same manual assembly, it is too early to expand deployment.
Use the overview of AI assistants for leaders to compare categories. If the main difficulty is repeatedly reconstructing project history, read about corporate memory. Executive AI's approach is presented on the product website.
