Quickfire: Your 6 “Must Have” Key takeaways

• A chat is not a unit of cost. Credits are metered from four things: model use, context retrieval, tool calls and runtime.
• Pay-as-you-go is around a penny a credit. Cost rises with how heavy the task is, so light jobs cost very little and big multi-output jobs cost the most.
• Your base Copilot licence already does a lot. Cowork is an add-on for complex, multi-step work, not a replacement for everyday Copilot.
• Do not always reach for the most powerful model. Leave the picker on Auto and keep Opus for work that genuinely needs it.
• You control access and spend. Cowork is off by default, with limits at tenant, group and user level.
• Going outside Microsoft 365 swaps containment for more predictable pricing. Decide on data sensitivity first, cost second.

Find out where Cowork could save your business the most time with our free AI Readiness conversation:

Book My Free AI Readiness Conversation With labdesk

For a while, Cowork came bundled into the wider Copilot story. Microsoft has now changed the access model: since Cowork became generally available in June 2026, it has been billed by usage, in Copilot Credits. You pay for what you actually use, rather than a flat add-on.

That is a sensible model, but it raises a fair question for any business leader: what is a credit, what drives the cost, and how do you stop it drifting? The good news is that you stay in control, and a lot of everyday work does not need Cowork at all.

Here is our plain-English guide: what a credit is, what is already included, how to pick the right model, how to keep a lid on spend, and when it is worth looking outside Microsoft 365.

So what actually is a credit?

Cowork does not charge a flat price per question. Each task uses a variable number of credits, worked out from four things:

• Model use. Which model runs the task and how much reasoning it does. Heavier models cost more.
• Context retrieval. How much of your data it pulls in: emails, files, Teams, SharePoint, calendars.
• Tool calls. Each action it takes: searching, reading files, building documents, running analysis.
• Runtime. How long the task runs from start to finished result.

So a quick lookup touches all four lightly and costs very little, while a long task that reads lots of sources, reasons deeply and produces several outputs touches all four heavily and costs the most. Cost tracks how heavy the task is, not the number of chats.

What your Copilot licence already covers

Cowork is an add-on. A user has to hold a Microsoft 365 Copilot licence before Cowork is even possible. That base licence is a flat, predictable fee, and it already covers the large majority of day-to-day AI work:

• Copilot Chat for questions, summaries, drafting and ideas, grounded in your work data.
• Copilot in your Office apps - Word, Excel, PowerPoint, Outlook and Teams - for help where you already work.
• The Work IQ context engine that grounds answers in your real files, emails, meetings and chats.
• Frontier-grade models for everyday tasks, included in the flat fee.
• Pre-built agents such as Researcher and Analyst, with no usage-based billing.
• Custom agents you build yourself with Agent Builder, included with the licence.

The practical point

• Most everyday AI work by summarising, drafting, quick analysis and answering questions. All of which is already covered by the flat fee. You only spend credits when you step up to Cowork’s more complex, multi-step, agentic work.

Cowork vs Copilot at a glance

When to use Cowork, and when you do not 

A simple rule of thumb: if you want help in the moment, use Copilot; if you want to hand over a whole task and get a finished piece of work back, use Cowork. A good share of the time, you will not need Cowork at all. It earns its keep on jobs like these:

• Building a deck or a report from several input documents.
• Comparing hundreds of files across two versions.
• Turning a messy pipeline into a ranked list of actions.
• In-depth research pulled together into a single, structured output.
• Building a comprehensive cost model or workbook from raw figures.

For the everyday bits like summarising a thread, drafting an email or a short post, a quick question over a single document, Copilot is the right tool, and it is already paid for.

Pick the right model, do not reach for the most powerful

The model you run is one of the four cost drivers, so choosing a model is a cost decision. The most common, and most expensive, mistake is reaching for the most powerful model for everything.The "Auto" option exists precisely so you do not have to.

• Auto (the default). Leave it here for most work. It routes lighter tasks to cheaper models automatically, which is the single biggest cost saver.

• Claude Sonnet 4.6. Fast and efficient for everyday drafting and quick checks. Cheaper than Opus.

• Claude Opus 4.8. The heavy hitter for complex, deep-reasoning work across several sources. This is the expensive one, so save it for jobs that genuinely need it.

• Sonnet + Opus Advisor. Sonnet does the work and Opus reviews it for accuracy. A second pair of eyes on important deliverables, without paying for Opus throughout.

• GPT 5.5 (Frontier). Versatile, and good for longer writing and citations, where your organisation has it switched on.

• Cowork 1 (coming soon). Microsoft's own model, built to handle everyday tasks at a substantially lower cost. One to watch for cost-sensitive work.

You can change the model in the same conversation without starting again, so you can begin on a light model and step up to Opus only if a task turns out to need it.

What makes one task cost more than another

Two tasks show why some jobs cost more than others.

Ask Cowork to summarise a single email thread and the cost stays low. It uses a light model, reads one source that’s already open, runs a couple of simple actions, and finishes in a few seconds.

Ask it to build a full cost spreadsheet from several reports and everything ramps up at once. It has to reason more deeply about the numbers, pull in more context from each report, call tools to open files, create the workbook, write formulas, check them, and run for several minutes.

Same tool, very different cost. What you ask it to do matters more than the format of the output. A long, text‑based analysis with lots of retrieval and reasoning can cost more than generating a single quick image.

You can see the exact credits for any task in Cowork’s per‑task cost readout as it runs.

How to read a cost estimate

Microsoft's estimator works out cost as the number of users in each profile, multiplied by their mix of light, medium and heavy tasks, multiplied by a price per task. It assumes the most expensive model by default, so it gives you a cautious upper figure.

• Light. A few sources, limited reasoning, one output or fewer.
• Medium. Several sources, structured reasoning, two or more outputs.
• Heavy. Broad aggregation, deep reasoning, many outputs.

How to read a scary number

Around 5,000 credits a month is a lot of heavy, all-Opus work. That kind of usage is what sits behind the really high estimates. If a number looks worrying, look at three things:

• Profile: Most people don’t run heavy tasks all day.
• Task mix: Keep heavy tasks to a few a week, not several a day.
• Model: Real usage on Auto, and soon Cowork 1, is usually well below an all-Opus assumption. Committing to a volume upfront can bring it down further.

Staying in control of spend

Cowork is off by default, and you manage it all from the Microsoft 365 admin centre:

• Decide who gets it. Turn it on per tenant and choose exactly who can use it.
• Set hard limits. Spending caps and budgets at tenant, group and user level, including per-user caps inside a group.
• Get alerted. Set the thresholds that matter and choose who is notified when spend crosses them.
• Let people ask for more. Users can request extra credits in the app rather than overspending quietly.
• See where it goes. Usage reporting by user, group and feature, plus the per-task cost shown to each user.
• Choose how you pay. Pay-as-you-go for flexibility, or commit volume up front for a lower rate.

What about using Claude directly?

A fair question is whether a flat-fee subscription such as Claude Pro is cheaper or more scalable than per-task credits, especially for self-contained document and report work. Sometimes it is. But if you advise clients to keep data inside Microsoft 365, the deciding factor is usually where the data lives, not just the price.

For something like an audit report, which usually carries sensitive client detail, doing it outside Microsoft 365 is exactly what a keep-it-contained policy is there to prevent. The flat fee might look cheaper, but it is the wrong trade for sensitive work. A quick way to decide:

• Is the data non-sensitive and fine to leave Microsoft 365? If not, stay in Cowork.
• Is the work self-contained, so you can hand over the inputs yourself? If it needs live grounding in your tenant, use Cowork.
• Is your volume high and steady enough that a flat seat beats per-task credits? If yes, and the first two are clear, an external tool may be cheaper and more scalable.
• Do you need 365-native governance such as audit trail, eDiscovery and data residency? If yes, keep it in Cowork.
  

Worth knowing: Microsoft's own June 2026 testing found Cowork averaged 30–40% cheaper per prompt than running Claude with its Microsoft 365 connector (both on the same top model), so outside is not automatically cheaper. For team-wide external use, Claude also offers Team and Enterprise plans with admin controls; check current per-seat pricing before you quote it.

The labdesk take

Credits do not change the real lesson: the value comes from knowing what to use, and when. Most everyday work is already covered by the licence you pay for; Cowork earns its credits on the heavy, multi-source jobs; and anything sensitive stays inside your tenant. Knowing which is which is a readiness question, and it is exactly what our Copilot Workshop is built to answer.

Getting the most from Cowork, the labdesk way

Cowork is a powerful tool. But like any AI, it is only as safe as the foundations underneath it. Access, data, security, and everyday ways of working decide whether Cowork becomes a trusted teammate across the business or something people feel they need to tiptoe around.

That foundation is what we help you put in place.

A simple first step is a free, tailored AI Readiness Conversation with a member of the labdesk crew. Together we look at how your team actually works today, where tools like Copilot and Cowork can save the most time, and what needs to be in place so they do it safely.

If that feels useful and you want to go further, our Copilot Enablement and Readiness Workshop takes things end to end.

Over several online sessions, we:

• Survey how your business really runs (tools, tasks, meetings, focus time, ownership).
• Review your Microsoft 365 data and setup so Copilot and Cowork sit on solid, governed ground.
• Run a security and risk assessment across 17 core items, including identity, devices, sharing, sensitivity labels, governance, and Secure Score.

You leave with clear findings, a simple picture of your current risk, and a phased roadmap from today’s foundations to being genuinely ready for AI. It is written in plain English, not tech-speak, and it is yours to keep and use however you like: run it in-house, hand it to your IT team, or have labdesk deliver it alongside you.

For most founders, the business case is already there. In a 100-person company, a 20-seat Copilot pilot can unlock around £110,000 of time value per year on roughly £6,000 of licences, provided the foundations are right. The workshop makes sure the work you are already doing on Microsoft 365 also moves you toward that AI readiness.

For qualifying projects, there is up to £1,000 of Microsoft funding available if you choose to move ahead.

If you would like your AI foundations in place before you roll out tools like Cowork to everyone, you can book your free readiness conversation here:

→ Book My Free AI Readiness Conversation With labdesk