Copilot Cowork Is Now Generally Available: What Microsoft’s Latest AI Releases Mean for Your Business

Three Microsoft announcements landed in the last fortnight. Read individually they look like routine product news. Read together they mark a line in the sand.

Copilot Cowork went generally available worldwide. Copilot in Excel picked up the features finance teams have been waiting for. And Azure set a world record for AI training speed. The connective tissue is execution. AI on the Microsoft stack has stopped recommending work and started finishing it.

That distinction sounds small. It changes how you budget, who you hire, and which workflows you hand over first. Here’s what each release actually does, and the part that matters for businesses running on Microsoft.

Note: all dollar figures are in US dollars unless stated otherwise, in line with Microsoft’s reporting.

Key Takeaways

• Copilot Cowork is generally available worldwide. Over half the Fortune 500 used it in preview, and Microsoft calls it the fastest growing feature in its Frontier program’s history.
• Cowork completes long-running, multi-step tasks end to end and returns a finished result, not a draft. It runs in the cloud and grounds itself in your Microsoft 365 data.
• Pricing breaks from per-seat. Cowork is billed on usage in Copilot Credits at one US cent each, on top of a Microsoft 365 Copilot licence. Microsoft’s first major pricing change in nearly two decades.
• Governance is the real story in the pricing. Cowork is off by default, and admins cap spend at tenant, group and user level before anyone runs a task.
• Copilot in Excel now targets finance directly: reusable skills, live data connectors, and a plan-then-act mode that shows its working before it edits a cell.
• Traceability is the headline. Every Copilot edit in Excel is attributed and reviewable, which is the difference between a demo and something an auditor will accept.
• Azure trained Llama 3.1 405B in just over seven minutes, a world record. The point isn’t speed you’ll use; it’s proof of infrastructure depth, now backed by a A$25 billion local expansion across Australia and New Zealand.
• The pattern across all three: agentic AI moved from pilot to production. The advantage now goes to teams who pick a workflow and run it, not the ones still writing a strategy.

 

Copilot Cowork Goes Global: AI That Finishes the Job

Copilot Cowork reached worldwide general availability on 16 June, after three months in Microsoft’s Frontier preview. The adoption numbers are the tell: more than half the Fortune 500 were already using it, Accenture, Capital Group and Zurich Insurance among them. For procurement-cautious enterprises to commit during a preview says the value was obvious early.

The line that separates Cowork from regular Copilot: Copilot hands you a draft and you act on it; Cowork acts, then hands you the result. You describe an outcome, it builds a plan, runs the work across your emails, files, meetings and data, checks in at the points that need a decision, and returns something finished. Because it runs in the cloud, it keeps working after you close your laptop, and it stays inside your Microsoft 365 security boundary throughout.

Microsoft’s examples show why that matters. A sales lead pointed Cowork at a stalled pipeline and got back a ranked list of at-risk deals, each with the specific follow-up that had gone cold. A week of review became a morning. The shift isn’t that the AI is smarter. It’s that the human is no longer the one assembling the work.

What this looks like in practice

A collections operations manager needs the monthly arrears summary. Today that’s an export from three systems, a reconciliation, a round of emails chasing context, and a deck. Handed to Cowork, it pulls the data, compares it against last month, flags the accounts that moved, drafts the summary, and pauses for sign-off before finalising. The manager reviews and approves instead of building. The judgement stays human; the assembly doesn’t.

The pricing change you need to plan for

Cowork is billed on usage, not per seat. You still need a Microsoft 365 Copilot licence, but each Cowork task draws Copilot Credits priced at one US cent each, costed across four inputs: the model used, context retrieved, tools called, and runtime. Microsoft buckets tasks as light, medium or heavy. This is its first significant pricing change in close to twenty years, and the logic is hard to argue with. When one active user can run hundreds of agent tasks a week, a flat seat fee either bankrupts the vendor or throttles the user.

The consequence for you is that AI spend becomes variable and therefore needs managing like any other consumption cost. Microsoft built the controls to do it: Cowork is off until an admin switches it on, access is assigned deliberately, and spending caps and alerts sit at tenant, group and user level. The teams who do well out of usage pricing are the ones who know which workflows earn their credits and retire the ones that don’t. Turn the controls on before the feature, not after the first invoice.

On models, Cowork runs on Anthropic’s Claude Opus 4.8 and Sonnet 4.6 at launch, with Microsoft’s lower-cost Cowork 1 due shortly. The multi-model design isn’t a spec-sheet detail. It’s the mechanism that keeps cost down, matching a cheap model to a simple task instead of running everything on the most expensive option.

 

Copilot in Excel Grows Up for Finance Teams

The Copilot in Excel update on 25 June is the one aimed at the people who read this. Finance has always asked the same three things of any tool: show your working, use trusted data, trace every number to its source. Most AI tools fail that test on the first count. This release is built around passing it.

Skills are the structural change. A team defines a process once, building a DCF, closing the books, running a variance analysis, and saves it as a reusable instruction file in OneDrive. Next month, instead of re-prompting from a blank box, you invoke the named skill and Copilot follows your steps, structure and formatting. It turns spreadsheet expertise from something locked in one analyst’s head into something the whole team can run consistently. Microsoft ships sample finance skills now, with partner skills from LSEG, Vena and others arriving later in the year.

Connectors close the data-provenance gap. Live feeds from Moody’s, Morningstar, FactSet, S&P Global, PitchBook and CB Insights pull straight into the workbook, so analysis starts from current figures rather than a copy-paste from a filing that may already be stale. Some connectors carry their own subscription.

Traceability is where finance teams should focus. Plan with Copilot makes it state which cells, formulas and assumptions it intends to touch, and lets you approve or change that plan before anything happens. Afterwards, every edit Copilot made is attributed in the Show Changes pane next to the human ones. That answers the question that has kept AI out of serious finance work: if a regulator asks how this number was reached, can you show them? Now you can.

What this looks like in practice

An analyst closing the month invokes a variance analysis skill. Copilot lays out its plan, the sheets it will read, the assumptions it will apply, and waits for approval. It compares actuals to plan, surfaces the largest movements, and drafts the summary in the format the business review always uses. Every cell it touched is tagged in Show Changes, so the reviewer sees exactly what the AI did and can challenge any of it. The reconciliation that ate a morning is done before the first coffee, and the audit trail is built in rather than reconstructed later.

The features are live across Excel for web, Windows and Mac for Microsoft 365 Copilot customers, with custom skills broadening through July. The sensible sequence is internal-facing work first, reporting packs, reconciliations, variance reviews where a human signs off, before any AI-assisted figure goes into a regulated external deliverable.

 

The Infrastructure Underneath: What Azure’s Record Proves

In mid-June Azure set a world record on the MLPerf benchmark, training Meta’s Llama 3.1 405B in just over seven minutes across more than 8,000 NVIDIA GB200 GPUs. You will never run a job like that. Almost no one does; training a frontier model from scratch is the work of a handful of labs. So why does it matter to a lender in Sydney or an insurer in Auckland?

Because a benchmark like this isn’t really about speed. At this scale the bottleneck stops being raw compute and becomes coordination, keeping thousands of chips perfectly in sync, since the whole job runs only as fast as its slowest member. Hitting a record means the networking, the orchestration and the stability all held together without a single stall. That is the same engineering that decides whether your Copilot query returns in two seconds or twenty when half the country is online at 9am. The flex is the headline; the proof of operational depth is the point.

And that depth is increasingly local. Microsoft runs 29 data centre sites across three Australian regions in Sydney, Melbourne and Canberra, opened its first New Zealand region in Auckland in late 2024, and in April committed a further A$25 billion to grow its Australian footprint by more than 140% by 2029. The record-setting infrastructure and the capacity landing on Australian and New Zealand soil are the same engineering programme. One proves what the platform can do; the other puts it within reach of workloads that have to stay onshore.

That distinction is what should interest a regulated ANZ business. The questions that decide a cloud commitment in financial services aren’t about benchmark times. They are about data residency, latency, resilience and certification. A deeper local footprint answers all four: workloads that stay in-country, real-time AI that behaves like real-time, fewer capacity bottlenecks, and a region carrying the compliance certifications regulated workloads require. The MLPerf result tells you the ceiling is high. The local build tells you how much of that ceiling you can actually use.

Capability Is No Longer the Question

A year ago the open question was whether agents would deliver. These releases close it. Cowork puts task completion in everyone’s hands, Excel puts it inside the tool finance has trusted for decades with the audit trail to match, and the infrastructure underneath, increasingly built on Australian and New Zealand soil, is being expanded to carry whatever comes next.

The constraint has moved. It used to be the technology. Now it’s how fast you can point it at real work and govern what it costs.

If you want to talk through what any of this means for your business, your AI roadmap, or your Microsoft investment, get in touch with the 365 Mechanix team. We work with organisations across Australia and New Zealand to turn announcements like these into something you can act on.

 

FAQs

What is Copilot Cowork?

Microsoft’s agentic AI inside Microsoft 365. Rather than answering one prompt, it completes multi-step tasks end to end, gathering context, doing the work and returning a finished result while you do something else. Generally available worldwide since 16 June 2026.

How much does Copilot Cowork cost?

You need a Microsoft 365 Copilot licence first. Cowork is then billed on usage in Copilot Credits, one US cent each under pay-as-you-go, with a discounted pre-purchase option. Each task’s cost reflects the model used, context retrieved, tools called and runtime, grouped as light, medium or heavy.

Is Copilot Cowork turned on automatically?

No. It’s off by default. An admin enables it, decides who gets access, and sets spending limits and alerts at tenant, group and user level. Worth configuring before anyone runs a task, because the pricing is variable.

What changed in Copilot in Excel?

Three things, announced 25 June: reusable skills for repeatable processes like variance analysis and DCF models; live data connectors that pull market data into the workbook; and stronger traceability through a plan-then-act mode and attributed edits in Show Changes.

Why does the Excel traceability matter for regulated businesses?

Finance and compliance teams have to show how a number was reached. Plan with Copilot lets you review the AI’s intended actions before it acts; Show Changes records what it changed afterwards. That review trail is what auditors look for. Start with internal-facing work before putting AI-assisted figures into regulated external deliverables.

What was the Azure training record, and does it affect us?

Azure trained Meta’s Llama 3.1 405B in just over seven minutes, a world record on MLPerf. You won’t run a job like it, but it demonstrates the coordination and stability of the infrastructure that also runs your everyday Copilot and Azure AI Foundry. For ANZ businesses it lands alongside a growing local footprint, 29 sites across three Australian regions, a New Zealand region in Auckland, and a A$25 billion Australian expansion by 2029, which is what makes that capability usable for workloads that must stay onshore.

How does 365 Mechanix help?

We turn Microsoft’s announcements into working implementations for organisations across ANZ. Dynamics 365, Power Platform, Copilot, agents, the lot. Whether that’s planning a Cowork rollout or governing usage-based spend, get in touch.