Get notified when Mark publishes

Receive a notification each time a new blog post is published. You will receive a confirmation email to approve your email. Unsubscribe anytime.

Mark Smith

Mark Smith

· 7 min read

Microsoft’s Copilot Strategy: Turning Shared Context and Persistent Agents into Business Outcomes

Excel followed Lotus 1-2-3. Windows Server followed Novell NetWare. Xbox entered behind PlayStation but stayed in the fight long enough to earn a permanent seat at...

Share

Microsoft has a habit of getting there.

It is rarely first into a market. It is often the last one standing.

Excel followed Lotus 1-2-3. Windows Server followed Novell NetWare. Xbox entered behind PlayStation but stayed in the fight long enough to earn a permanent seat at the table.

Microsoft is not always first. But once it commits to a category, it tends to get there in the end. That is the lens I am using for its latest Copilot announcement.

What Microsoft announced#

On 25 September, Microsoft introduced a new Copilot organised around three capabilities.

Home brings Chat and Cowork together. Chat is for quick questions and drafts. Cowork is for work you hand off and get back finished. With Office in Copilot, Word, Excel, and PowerPoint files can be created and edited there, live for the team.

Code lets people describe an app, tracker, dashboard, or automation in plain language, and Copilot builds it. Code uses the same underlying technology as GitHub Copilot. A new Copilot Managed Runtime, currently in preview, hosts and governs those apps inside the Microsoft 365 environment.

Autopilot, previously called Scout, is a persistent agent with its own identity, memory, and workspace. You give it a role and a goal, and it keeps working when you are not.

Home and Code are scheduled to roll out through the Frontier programme in the coming weeks. Autopilot is scheduled to enter private preview at the end of September.

The individual features are not the story. The story is how they fit together.

Three people building the same thing#

Last week I was talking with someone from a New Zealand surveying firm. About 100 people, spread across the country, surveying land to exacting standards.

He, his boss, and another colleague had started building their own internal software. Not because they wanted to be developers, but because what they needed could not be bought off the shelf. The tools made their work more consistent, lifted quality, and gave back hours that a slow, repetitive task had been eating.

Here is the catch. All three were often building the same thing. Different parts of the country, different laptops, no shared view. Nobody knew the other two were solving the identical problem.

That is the gap I see this announcement closing. When the people who understand the work can build solutions inside the same environment their colleagues already use, with shared documents, shared context, and a place to host what they build, three people duplicating effort becomes three people improving one tool.

For a firm that size, that can change the result.

Why Microsoft 365 matters: Work IQ#

The reason this works in Microsoft 365 is context.

For about a decade, Microsoft Graph has been the plumbing underneath Microsoft 365. It connects email, meetings, chats, files, people, and calendars. Graph tells an AI what exists.

Microsoft introduced Work IQ as an intelligence layer built on top of Graph. Work IQ is intended to help Copilot understand how an organisation works: which people collaborate, which email thread, Teams meeting, and SharePoint document belong to the same project, what deadlines are coming, and what is stalled.

The business-data capability, drawing on Dynamics 365 and Power Platform, is scheduled to begin preview on 30 September and roll through October. That timing matters: keep it separate from capabilities already available today.

Work IQ respects existing access controls. Copilot can access organisational content only in line with the permissions of the person using it.

That is what the surveying firm was missing. The ability to build, plus a shared understanding of what everyone else was building.

The harness story#

The concept I want every leader to understand is the harness.

A model is the engine. It does the reasoning and writing. The harness is everything around the engine. It decides when to call the model, what context to provide, which tools to use, how to recover when a step fails, and what gets logged.

Microsoft already runs several harnesses in Copilot Studio, including one that uses GitHub Copilot technology for complex, multi-step work. That does not mean the whole Copilot Studio harness is built on GitHub Copilot.

What Microsoft is really offering here is an enterprise harness: identity, permissions, audit, data, workflows, and cost controls around a choice of models.

Copilot uses models from OpenAI and Anthropic alongside Microsoft’s own models. The strategic advantage is not that one model will always win. Models will keep leapfrogging one another. The advantage is that context, skills, workflows, evaluations, permissions, and governance can remain in the enterprise layer while model choice changes underneath.

That is my analysis, not a promise that switching models will be frictionless. But it is the strategic possibility Microsoft is positioning itself to provide.

Where your data goes#

My biggest concern in the market right now is this.

The more of a company’s context, memory, and workflows it builds directly on one lab’s platform, the more operating knowledge can become dependent on that platform. Staff pasting company data into personal accounts creates a separate risk, because those accounts may have weaker protections.

The Microsoft model is designed around Microsoft’s service boundary, existing permissions, and contractual data protections. It is not accurate to say that all data stays inside the Microsoft 365 tenant in every scenario.

Model, subprocessor, and regional-processing terms vary. When Copilot uses Anthropic models, Anthropic operates as a Microsoft subprocessor under Microsoft’s applicable terms; Anthropic processing is outside the EU Data Boundary. Organisations should therefore review the relevant Product Terms, Data Protection Addendum, service documentation, and regional commitments rather than relying on a blanket "data stays in the tenant" assumption.

The principle remains important: the enterprise harness should keep identity, permissions, governance, and control around the use of models.

Building capability as a team#

This moves AI from something individuals dabble with to something teams build together. That is where the value is, because AI is only worth the effort if it moves one of three drivers:

  1. Revenue: faster proposals, better-prepared customer meetings, fewer dropped follow-ups.
  2. Cost reduction: less time lost to coordination, chasing, and rebuilding the same thing twice or three times.
  3. Innovation: people who understand the business building solutions for it.

If a use case does not move one of those, park it.

My one big caveat: who is watching the spend?#

Everyday Copilot Chat and Copilot use in Office sit within fixed subscription arrangements. Cowork, Code, Autopilot, long-running agentic work, and some frontier-model use introduce usage-based billing through Copilot Credits.

That is a new way of working, and I do not think most finance departments are ready for it. Finance teams know how to budget for licences. They are far less practised at forecasting variable usage, especially from agents that keep running in the background.

Microsoft provides controls including spending policies, organisation and user limits, alerts, and approval routing. But the tool is not the policy. Someone has to decide what an individual can spend, what a team can spend, what the company will spend in total, and who approves more.

One detail to check: spending policies can automatically include new services as Microsoft adds them, and that setting is on by default.

Get finance, IT, and the business in a room before the features arrive, not after the first invoice.

What comes next#

Microsoft has committed to this category. Your organisation needs to be ready when it lands.

Pick one team. Pick one outcome tied to revenue, cost, or innovation. Put your spending policy in place first. Then let the people who understand the work start building together.

Evaluate the result, not the novelty of the feature. Measure time saved, revenue enabled, cost removed, quality improved, or new capability created.

Microsoft has a habit of getting there. I would like you to be there waiting when it does.

So, over to you. Is anyone in your organisation quietly building the same thing as a colleague down the road?


About Mark Smith

Mark Smith is Principal AI Counsel at Cloverbase, where he advises New Zealand and Australian organisations on AI strategy, governance, and skills. He is a mentor, author, and speaker, and a Microsoft AI MVP. Cloverbase Ltd.

Mark Smith

Mark Smith

Principal AI Counsel · Microsoft MVP

Helping people build practical AI skill in the Intelligence Age.

Discussion

Comments

Loading the discussion for this post.

Loading

Leave a comment

Your email stays private. If it matches a Gravatar account, your public avatar can appear after the comment is approved.

Used only for reply notifications and optional Gravatar matching. Never displayed publicly.

Max 2000 characters0/2000

More from nz365guy