AI & Automation
Microsoft 365 Copilot Cowork is a cloud workload, not a feature: here is what that means for your budget
July 9, 2026
5-minute read
A detailed sales proposal, over four iterations: 8,000 Ft + VAT. A daily executive summary: roughly 1,000 Ft. A thoroughly researched article written in two languages: 1,800–2,000 Ft. These aren’t list prices, but our own estimates based on real-world tasks over the past few days. And they show exactly what most onboarding guides leave out: Cowork isn’t a feature—it’s a workload. It needs to be treated as such.
Microsoft 365 Copilot Cowork is now generally available, and most organizations’ initial reaction is the same: it’s included in the package, so someone will give it a try. That approach won’t work here.
An M365 Copilot license is required to use Cowork, but! In addition, you’ll be billed based on usage, at $0.01 per credit. The actual cost of a task is determined by four factors combined: the model run, the amount of business context extracted from Microsoft Graph (now WORK IQ), the number of resources used, and the runtime. This leads to the model’s most important characteristic: it’s not the type of task that costs money, but the complexity of the task.
We can run an Excel formula as many times as we want; the bill doesn’t change. At Cowork, every iteration moves the meter. A vague initial prompt, followed by three rounds of revisions, is both a quality and a cost issue. The most important lesson from our measurements is this one sentence: the meter is moved by the iteration.
We carried out three typical types of tasks in real-world, client-focused work.
Daily executive summary from multiple sources: approximately 1,000 Ft. This is a reasonable price for an organization to pay if the task is repetitive, involves multiple systems (email, Teams, document repository), and would take 20–30 minutes to complete manually each day. Here, Cowork acts as a true agent: it gathers the context, structures it, and provides usable output.
Research-intensive article in two languages: 1,800–2,000 Ft. Includes source research , structure, and two language versions. Measured in man-hours, this is the best value of the three.
Comprehensive sales proposal with four rounds of revisions: 8,000 Ft + VAT. This amount is justifiable if the proposal targets a serious business opportunity and all four rounds resulted in genuine improvements to the content. It is questionable, however, if half of the iterations were due to the initial brief not being precise enough. Cowork rewards clear delegation and penalizes sloppiness—just like a good expert colleague would.
The first and most important decision regarding the rollout is not technical, but organizational: users must be taught which tasks warrant coworking and which do not. In practice, there are three levels.
Copilot Chat and Copilot built into the app: simple , one-step tasks. Rewriting emails, short summaries, quick Q&A. Cowork is too expensive for these tasks.
Cowork: repetitive , complex tasks that draw on multiple sources and result in some kind of “physical” end product. Executive summaries, preparing for client meetings, comparing documents, drafting proposals.
Copilot Studio: repeatable, integration-requiring use cases built into processes. Anything that runs on a weekly or event-based schedule and connects systems is not an ad-hoc agent task, but a process that should be automated.
If this routing is missing, the organization will pay an “agent fee” for administrative tasks. And in a few months, that invoice will be hard to justify to the CFO.
There’s one thing most organizations underestimate. Cowork operates through Microsoft Graph, which means it sees exactly what a given user’s permissions allow. If SharePoint permissions have grown organically over the years, if sharing links are more widely accessible than they should be, or if identity and group management is lax, Cowork brings these issues to light faster than any of our current processes.
This is both a warning and an opportunity. Anyone who completes the tenant review (permission structure, oversharing report, labeling of sensitive content) before the rollout reaps a twofold benefit: Cowork operates within a more accurate context, and the permission structure is restored. Those who don’t will have the review performed by the end of the first month—only in a more inconvenient way.
My position is clear: Coworköt should not be discussed as an abstract licensing cost, but rather evaluated as a business model—using the same logic we’ve already learned in cloud cost management. In practice, this means four things.
Designated users. Not everyone is granted access by default. Access is granted to those whose work involves at least one recurring, complex task that involves multiple sources.
Budget by task. Before going live , estimate the expected credit usage. A quote of 8,000 Ft is justifiable, but an email summary costing 8,000 Ft is not.
Reviewing permissions before deployment. Without tenant screening , Cowork will be the fastest—but least convenient—audit tool.
Business-results-based measurement. The question isn’t how many tasks were run, but how much manual work it replaced and what the quality of the output was. Work hours saved, measured per output, not credit statistics.
That's the line between maturity and immaturity. It's not the one who turns it on first who wins, but the one who knows why they're running it.
The implementation of Copilot Cowork is an organizational issue that we address with a technological tool. The technological side is a given: it’s pay-as-you-go, credit-based, and can be added to an M365 Copilot license. Everything is decided on the organizational side: routing logic, tenant hygiene, task-level budgets, and performance-based metrics.
Anyone who builds these four elements will see a real return on investment. Anyone who simply turns it on and waits will end up with a hard-to-understand bill and dissatisfied users a few months later. The difference isn’t in Cowork. The difference lies in whether the organization is ready to treat AI not as a feature, but as a workload.
László Földesi, one of Gloster's Microsoft practice leads, spent several weeks running Copilot Cowork on real client work before writing up his findings. The short version: organisations that treat Cowork as a smarter Office button will pay agent prices for assistant-grade tasks. The ones that manage it as a cloud workload will not.
Copilot Cowork became generally available alongside the existing Microsoft 365 Copilot licence, billed on a pay-as-you-go basis at $0.01 per credit. That sounds trivial until you understand what drives the meter: the model you run, the volume of business context pulled in, the number of tools invoked, and elapsed execution time. A single agentic task is not a fixed-price API call. It is a workload with variable cost, and it behaves like one. László's team tested this directly. A daily executive summary cost roughly £9 to run. A detailed sales proposal, taken through four iteration rounds, came to around £60 before VAT. Neither figure is alarming on its own. Both become alarming at scale if nobody owns the routing decision. That is the governance gap most organisations are walking into.
The cost structure of Cowork rewards precise delegation. A clear, well-scoped prompt that produces a usable output in one pass is cheap. A vague first prompt that requires three correction rounds is four times the cost and four times the latency, and the output quality is usually worse than a single well-formed request would have produced. This is simultaneously a cost problem and a quality problem, which means the fix is the same for both: prompt discipline. Organisations that invest in prompt standards early will see lower per-task costs and better outputs. Those that treat Cowork as a general-purpose chat interface will see neither. László's framing here is worth quoting directly: Cowork rewards clear delegation. Sloppy delegation it penalises.
Not every task belongs in an agentic workflow. László's testing identified a clear pattern for where Cowork justifies the credit spend: tasks that are repetitive, draw on multiple sources, and require synthesis rather than simple retrieval. The strongest use cases from the Gloster tests were executive briefing documents, client meeting preparation that pulls from CRM, email, and calendar simultaneously, document comparison across large sets, and multi-source proposal drafting. These are tasks where a human would spend 45 to 90 minutes gathering context before writing a single sentence. Cowork compresses that gathering phase. Simple tasks belong elsewhere. Rewriting a short email, summarising a single document, answering a factual question: these run faster and cheaper in Copilot Chat or in the application-embedded Copilot. Routing them through Cowork is the equivalent of booking a freight lorry to deliver a letter.
Cowork inherits your tenant's permission structure. If your Microsoft 365 tenant has accumulated years of over-permissioned service accounts, orphaned guest users, and SharePoint sites with inconsistent access controls, Cowork will surface those problems faster than your current processes do. This is not a Cowork bug. It is a signal. An agentic workflow that can traverse your tenant's data estate will find the gaps your manual processes have been quietly ignoring. The organisations that have maintained clean identity governance will find Cowork straightforward to deploy. Those that have not will find it educational, and not always in a comfortable way. László's assessment is direct: if your tenant hygiene is weak, Cowork will expose it faster than your existing workflows. The correct response is to fix the hygiene before you scale the workload, not after.
The practical implication of everything above is that Cowork requires the same governance discipline as any other cloud workload you run on Azure or AWS. That means named users with defined scope, cost budgets set at the department or project level, a regular review of what is actually running and what it costs, and measurement tied to business outcomes rather than feature adoption metrics. The question to ask is specific: did this agentic task produce an output that would have taken a person 60 minutes, and did it cost less than the hourly rate of the person who would have done it? If yes, the routing decision was correct. If the answer is unclear, the measurement framework is missing. Licence cost is the wrong unit of analysis for Cowork. Operating model is the right one.
Before any organisation scales Cowork beyond a pilot group, the routing question needs an answer that users can apply without asking IT. Which work earns an agent, and which does not? This is not a technical configuration. It is a policy decision, and it belongs with the people who understand the cost of the tasks being delegated. A well-designed routing guide, built from the kind of real-cost testing László's team ran, gives users a decision framework they can apply in under ten seconds. Without it, the default is to route everything through the most capable tool available, which is the most expensive option for the majority of tasks. The maturity line for Copilot Cowork adoption is not who switches it on first. It is who can articulate why a given task belongs there, and who can show the cost data to back that judgement.
If your Microsoft 365 Copilot renewal or initial deployment is coming up in the next two quarters, Cowork governance belongs in the conversation before you sign. The licence cost is predictable. The credit spend is not, unless you build the controls before you scale. Gloster's Microsoft practice has run this analysis across 400+ Microsoft 365 estates. The pattern is consistent: organisations that treat Cowork as a managed workload from day one spend less and get more usable outputs than those who retrofit governance after the fact. The difference is not the technology. It is the decision framework built around it.
László Földesi, Cloud-Berater bei Gloster, hat Copilot Cowork in den vergangenen Wochen an echten Aufgaben gemessen. Sein Befund: Wer den Dienst wie eine erweiterte Office-Funktion einführt, zahlt Agentenpreise für Assistentenarbeit. Wer ihn als verwalteten Cloud-Workload betreibt, gewinnt einen messbaren Produktivitätshebel.
Microsoft hat Copilot Cowork zur allgemeinen Verfügbarkeit freigegeben. Die Reaktion in vielen IT-Abteilungen: ein Eintrag im Changelog, vielleicht ein kurzes Pilotprojekt, dann Weiterbetrieb wie bisher. Das ist ein teurer Irrtum. Cowork rechnet nach dem Pay-as-you-go-Prinzip ab: 0,01 US-Dollar pro Credit, zusätzlich zur bestehenden Microsoft 365 Copilot-Lizenz. Klingt günstig. Wird es auch bleiben, wenn Sie wissen, was die Kosten treibt. Wird es teuer, wenn Sie es nicht wissen. Die Kostentreiber sind vier: das ausgeführte Modell, die Menge des eingebundenen Geschäftskontexts, die Anzahl genutzter Tools und die Laufzeit. Jede Iteration eines unklaren Prompts dreht an allen vier Reglern gleichzeitig.
László Földesi und sein Team haben Cowork an realen Aufgaben getestet, mit konkreten Kostenergebnissen. Eine tägliche Zusammenfassung aus mehreren Quellen: rund 1.000 HUF pro Durchlauf. Ein detailliertes Verkaufsangebot, über vier Iterationen entwickelt: rund 8.000 HUF zuzüglich Mehrwertsteuer. Diese Zahlen sind kein Urteil über Cowork. Sie sind ein Hinweis auf die richtige Frage: Welche Aufgabe rechtfertigt diesen Aufwand? Die Antwort ist präziser, als viele erwarten. Cowork lohnt sich bei Aufgaben, die wiederkehrend sind, mehrere Quellen zusammenführen und echte Komplexität tragen: Führungszusammenfassungen, Vorbereitung von Kundengesprächen, Dokumentenvergleiche, Angebotserstellung. Für einfache Aufgaben wie das Umformulieren einer E-Mail oder eine kurze Zusammenfassung ist Copilot Chat oder der in die Anwendung eingebettete Copilot die richtige Wahl. Wer das nicht unterscheidet, zahlt Agentenpreise für Assistentenarbeit.
Cowork erbt die Berechtigungsstruktur Ihres Tenants. Das ist keine Kleinigkeit. Wenn Ihre Tenant-Hygiene schwach ist, also veraltete Konten, zu weit gefasste Zugriffsrechte, unklare Datenklassifizierungen, dann bringt Cowork diese Probleme schneller an die Oberfläche als Ihre bisherigen Prozesse. Der Dienst greift auf Geschäftskontext zu, der im Tenant liegt. Er macht sichtbar, was dort bereits vorhanden ist. Das ist kein Argument gegen Cowork. Es ist ein Argument dafür, die Berechtigungsstruktur vor dem Rollout zu prüfen, nicht danach. Die erste Entscheidung bei der Einführung ist das Routing. Ihre Mitarbeitenden müssen verstehen, welche Aufgaben einen Agenten rechtfertigen und welche nicht. Ohne diese Unterscheidung entsteht unkontrollierter Verbrauch, und der ist in Pay-as-you-go-Modellen schwer rückgängig zu machen.
László Földesis Kernthese lautet: Behandeln Sie Cowork wie jeden anderen verwalteten Cloud-Workload. Das bedeutet konkret: benannte Nutzer mit definierten Anwendungsfällen, Kostenrahmen pro Team oder Abteilung, regelmäßige Überprüfung der Berechtigungen und Messung anhand von Geschäftsergebnissen, nicht anhand von Aktivierungszahlen. Diese Anforderungen sind aus dem Cloud-Betrieb bekannt. Sie gelten für Azure-Workloads, für Dynamics-Instanzen, für jede Ressource, die nutzungsabhängig abrechnet. Cowork ist keine Ausnahme. Wer diesen Schritt überspringt und Cowork als erweitertes Office-Feature behandelt, verliert die Kostenkontrolle. Wer ihn geht, gewinnt einen Dienst, der komplexe, quellenübergreifende Aufgaben strukturiert abnimmt und dabei nachvollziehbar bleibt.
Földesi benennt eine Grenze, die in der Einführungsdiskussion selten explizit gemacht wird: Der entscheidende Unterschied liegt nicht darin, wer Cowork zuerst aktiviert, sondern darin, wer weiß, warum. Organisationen, die diese Frage beantworten können, haben Routing-Regeln definiert, Kostenverantwortung zugewiesen und Tenant-Berechtigungen bereinigt. Organisationen, die sie nicht beantworten können, werden Cowork einschalten, Kosten akkumulieren und den Dienst nach drei Monaten als zu teuer einstufen, ohne je die richtigen Aufgaben damit bearbeitet zu haben. Das ist kein technisches Problem. Es ist ein Governance-Problem, und es lässt sich lösen, bevor der erste Agent läuft. Die Frage, die sich IT-Verantwortliche jetzt stellen sollten: Haben wir die Berechtigungsstruktur unseres Tenants im Griff, und wissen unsere Teams, welche Aufgaben einen Agenten rechtfertigen? Wenn eine der beiden Antworten unklar ist, ist das der Startpunkt.
