
Every month the Azure bill is a little higher than the month before. Nobody decided it. Finance receives a document many pages long, full of meters, regions and resource names, and cannot tell which part is the business running and which part is waste. IT says everything on the list is needed, and most of it probably is. The conversation ends in one of two ways: nobody touches anything and the bill keeps growing, or someone switches something off to see what happens.
There is a better order. The bill can be read, and the reasons it grows are few and well known: resources left running that nobody uses, resources sized for a peak that never came back down, things that keep being billed after the machine they belonged to has gone, and data that only ever grows. Microsoft gives you two tools for finding them, Cost Management and Azure Advisor, and Microsoft states that there is no charge for using Cost Management on your Azure subscriptions.
This article goes through them in the order we use: first read the bill, then find out who owns what, then look where savings usually hide, and only at the end commit to anything. Each saving is written down with its amount and its risk, because the risk is what makes a saving safe to apply. At the end there is a list of what you can do this week.
A word before starting. Everything in the first half is reading: looking at Cost Management and Advisor changes nothing in your subscription. The changes come later, and each one is checked first with whoever knows what the resource is for.
Why a cloud bill grows on its own
In a data centre of your own, a server that nobody uses costs the same as one that works: it was paid for when it was bought. In Azure, most resources are billed for as long as they exist or as long as they run. That is the point of the cloud — you pay for what you use — and it is also why the bill grows when no one is looking. Four patterns explain most of it.
Added for a peak, never scaled back
A season of high sales, a migration, a month-end close that ran slowly once: a server is resized up, a database is given more capacity, a second instance is added. The peak passes. Scaling back down is nobody’s job, and it does not happen. The resource now costs the peak’s size every month of the year.
Created for a test, never deleted
A test environment for a new version, a proof of concept, a copy of production for a training session. When the work is over, the environment stays, because deleting it feels riskier than keeping it. Often it runs at night and at weekends too, when no one could possibly be using it.
Left behind by something deleted
When a virtual machine is deleted, its disks are not always deleted with it — it depends on how the machine was created — and Microsoft is plain about the consequence: after a virtual machine is deleted, you continue to pay for its unattached disks. A snapshot taken before an upgrade years ago, a public IP address reserved for a server that no longer exists: each keeps appearing on the bill on its own.
Data that only grows
Storage, backups and logs accumulate. Nobody deletes a file that might be needed, nobody moves old data to a cheaper tier, and every application writes logs that someone, someday, might want to read. Unlike a server, this cost never goes down by itself.
None of these is a mistake in the usual sense. Each was a reasonable decision at the time. The problem is that nothing in the subscription remembers why a resource was created, and nothing reminds anyone to look again.
Read the bill before cutting anything
The invoice is a summary. The detail is in Cost Management, in the Azure portal, and it can be read without changing anything. Microsoft calls its cost analysis the first stop when you need to explore your costs.
Who can read it
To read costs without being able to change resources, Azure has two roles. Reader gives view-only access; Cost Management Reader gives view-only access to cost data and recommendations, and cannot create budgets. Neither gives access to invoices. Assigned on a subscription, they show the costs of that subscription; Reader also shows the resources themselves, which Cost Management Reader does not, and a review that looks at disks and addresses needs them. Depending on how your company buys Azure, the costs of the whole billing account need a billing role instead, such as the Reader role of a Microsoft Customer Agreement or the read-only user of an Enterprise Agreement. Under an Enterprise Agreement there is one more switch: for cost data to be visible at subscription level, the setting that lets account owners view charges must be on. Ask whoever manages your Azure contract which case is yours.
What to look at first
Cost analysis has ready-made views by resource, by resource group, by service and by subscription, and three of them answer most of the first questions.
The cost by service tells you what kind of thing you pay for: virtual machines, storage, databases, networking, monitoring. The cost by resource group and by resource tells you which ones. And the cost per day tells you when the growth started: a step on a given day usually has a cause on that day — a deployment, a resize, a new environment — and someone may remember it. Cost analysis shows daily costs for up to 92 consecutive days; beyond that it switches to months. The portal keeps the last 13 months.
What the numbers are
Three things to know when reading. First, cost data is not live. Under an Enterprise Agreement or a Microsoft Customer Agreement it typically arrives within 8 to 24 hours, and on pay-as-you-go it can take up to 72 hours, so the last day shown is usually incomplete. Second, until the month is invoiced the figures are estimates: Microsoft says the open month’s data should be considered estimated only, and the month is typically closed up to 72 hours after it ends.
Third, there are two ways of showing a reservation or a savings plan. Actual cost shows the purchase as it appears on your bill, so a virtual machine covered by a reservation shows a cost of zero for the month. Amortised cost spreads the purchase over its term and assigns it to the resources that used it; the part nobody used is assigned to no subscription. To see what each resource really costs per month, use amortised cost. One limit: cost analysis does not show amortised reservation costs for a pay-as-you-go subscription.
The first list
At the end of this reading you have a short list: the twenty resources that cost the most, the services that grew the most, and the days on which the growth happened. It is not a list of savings yet. It is the list of questions to ask.
Find out who owns what
A resource without an owner is a resource nobody will dare to touch. Before looking for savings, each of the expensive resources needs a name next to it: who created it, what it is for, and whether it is still needed.
Tags
Azure lets you attach tags to resources: name and value pairs such as an owner, a cost centre, an environment (production, test) or a project. Cost analysis can group costs by tag, which turns a list of resource names into a list of owners and projects. Costs without the tag appear together as «Untagged», always at the bottom of the list, even when they are the largest group — unless there are ten or more tag values, when they fall into «Others».
Two things to know. Tags are not retroactive: they appear on the cost of usage recorded while the tag is on the resource, not on the months before. And they must be applied directly to each resource: a resource does not inherit the tags of its resource group. Cost Management has a setting for this, tag inheritance, available under an Enterprise Agreement, a Microsoft Customer Agreement, or a Microsoft Partner Agreement with Azure plan subscriptions. It applies the tags of the subscription and the resource group — under a Microsoft Customer Agreement, also those of the billing profile and the invoice section — to the usage records of the resources inside them: the records, not the resources. It reaches back only to the start of the current month: switched on on the twentieth, it updates the month’s records from the first, using the tags as they are on the twentieth.
Resource groups
Even without tags, resource groups are often already organised by application or by environment. Grouping cost analysis by resource group, and asking the technical team what each group is, is often enough for the first pass.
Where to look first
With the bill read and the owners known, the search for savings can start. These are the places where it usually pays, in the order we look at them.
Virtual machines that run when nobody uses them
Azure Advisor, also in the portal, analyses the use of your virtual machines and recommends shutting down those that look idle and resizing those that look too big. For a shutdown it reads processor and outbound network use, not memory: a machine is a candidate when the 95th percentile of its maximum processor use is under 3 %, its average processor use over the last three days never went above 2 %, and its outbound network stayed under 2 % over seven days.
Its recommendations rest on a lookback period of seven days by default, which Advisor’s configuration lets you set to 7, 14, 21, 30, 60 or 90 days; after a change, the recommendations can take up to 48 hours to update. A machine used only at month-end can look idle over seven days and busy over thirty. Set the period to cover at least one month before reading the list.
There is a detail that costs money every month in many subscriptions. A virtual machine shut down from inside its operating system goes to the state «Stopped», and in that state it is still billed for its compute. The Azure command line says the same of its own stop command: the machine continues to be billed, and to avoid that you deallocate it. Only a deallocated machine — «Stopped (deallocated)», or «Hibernated (deallocated)» if it was hibernated — has stopped being billed for compute — and even then, Microsoft adds, disks and networking continue to incur charges.
For test and development machines used only in office hours, Azure has an automatic shutdown at a set time, configured on each machine under Operations, with an optional notice before it happens. Its time zone is UTC unless you change it, which matters when the office is in Madrid or Milan. The morning after the first automatic shutdown, check in the portal which state the machine is in.
Machines sized for a peak
Advisor also recommends a smaller size for a machine whose use stays low. For resizing it reads processor, memory and outbound network, and it picks a size on which, for a workload that users face, processor and network would stay at or under 40 % and memory at or under 60 %. A resize restarts the machine, and the smaller size leaves less room for the next peak. Both are part of the risk, and the owner is the person who knows whether a peak is coming.
One caution on the amounts. Microsoft notes that these recommendations do not take into account reservations or savings plans you already have, so the savings listed might be higher than actually possible. The amount you write next to each saving comes from cost analysis, not from Advisor.
Disks, snapshots and addresses that belong to nothing
A managed disk that is not attached to any machine is still billed. In the portal, each disk’s page shows its state, and an unattached one says so. Snapshots are billed for the data they actually use: an incremental snapshot keeps only the changes since the previous one, but the first one is a full copy of the disk. And a static public IPv4 address — static is the only kind the Standard SKU offers — is charged whether or not it is associated with anything, until the address resource is deleted.
Each of these is small on its own, and they add up over the years. The risk here is different from a resize: Advisor’s own recommendation on unattached disks says it in five words: deleting a disk is irreversible. Before deleting, find out why it was kept. A snapshot taken before an upgrade may be the only way back from it.
Databases with more capacity than they use
An Azure SQL database in the provisioned tiers is billed for the capacity it is given, not for the work it does. A database sized for a peak, or a test copy given the same capacity as production, pays for that capacity all month, and the database’s own metrics show how much of it is used. For databases used only at certain times, Azure SQL Database has a serverless tier that bills compute per second and can pause the database after a period with no activity, set in the General Purpose tier between 15 minutes and seven days. While the database is paused, compute costs nothing and only storage is billed.
Storage in the wrong tier
Blob storage has access tiers — hot, cool, cold and archive — that trade a lower price for keeping data against a higher price for reading it. There is also a smart tier, which moves data among hot, cool and cold by itself and charges no early deletion fees. Data written once and rarely read, such as old documents, backups or exports, often sits in the hot tier. Lifecycle management policies can move blobs to a cooler tier, or delete them, according to rules based on when a blob was created, last modified or, if access tracking is on, last read.
The tiers have rules that turn a saving into a cost if they are ignored. In a general-purpose v2 account, cool, cold and archive have a minimum stay of 30, 90 and 180 days, and a blob deleted, overwritten or moved to another tier earlier is charged an early deletion fee for the rest of the period: Microsoft’s example is a blob deleted 120 days after it went to archive, charged as if it had stayed 180. Data in archive cannot be read directly. It has to be rehydrated first, which with standard priority can take up to 15 hours for objects under 10 GB. A rule that sends data to archive needs to know that nobody will want it tomorrow.
Logs that nobody reads
Microsoft writes that for most Azure Monitor implementations the most significant charges are the ingestion and retention of data in Log Analytics workspaces. An application or a resource that sends detailed logs nobody reads can become one of the larger lines of the bill, and it grows with the traffic. Each workspace’s usage shows which tables and which resources send the most data. On the analytics plan, the first 31 days of retention are included; keeping data longer costs more, and deleting data does not lower that cost — shortening the retention of the workspace, or of a table, does.
A workspace can be given a daily cap, and Microsoft is clear about what it does: once the cap is reached, data collection stops, which affects your ability to monitor resources and receive alerts. Data collected above the cap is still billed. Microsoft presents it as a safeguard, not a routine way to control costs. The saving comes from collecting less: fewer diagnostic categories, less verbose application logs, shorter retention where your own rules allow it.
An example: the first afternoon
Here is how the first reading of a subscription might go. It is an example, not a client: the names are invented and there are no amounts.
The daily view of the last three months shows one step, in the middle of the second month, and the cost stays at the new level from then on. Grouped by resource group on that day, the step belongs to a group called after a migration project. The project’s lead explains that the migration ended two weeks later. Inside the group are two virtual machines in the state «Stopped» — shut down from inside the operating system, still billed for compute — three disks, one of them unattached, and a storage account holding the exports of the old system, in the hot tier.
Advisor, with its lookback set to thirty days, lists one more machine, outside the project, as a candidate for a smaller size. Its owner says it is the machine of the month-end reports.
So the first list has five lines. Deallocating the two stopped machines: large amount, low risk, once the project lead confirms that nobody needs them running. Deleting the unattached disk: small amount, but it cannot be undone, so it waits until someone confirms it holds nothing to keep. Moving the exports to a cooler tier: the rules say how long they must be kept and how often they are read, and the minimum stay of the tier is checked against both. The report machine: tested after a close, not before one. And the question that remains is why the migration group was never deleted — the answer becomes the habit of asking for an end date.
Every saving with its amount and its risk
At the end of the search you have a list. Before applying anything, each line gets two things written next to it: the estimated monthly amount, taken from cost analysis, and the risk — what could stop working if the saving is applied, and how to check it beforehand.
The amount comes from the bill. The risk comes from the owner. Switching off a test environment at night carries little risk. Resizing the database of the online store does not: it is tested first. Deleting a disk cannot be undone, so it waits until someone has confirmed why it was kept. A saving with a large amount and an unknown risk is not applied: its risk is found out first.
Sorted this way, the list decides itself. The large savings with low risk go first. The large ones with real risk are tested. The small ones with high risk are often left alone. And decisions stop being made by switching something off to see what happens.
Commit last: reservations and savings plans
Azure offers lower prices in exchange for a commitment of one or three years. A reservation commits to a specific resource, size and region; it is a billing discount and does not change how the resource runs. A savings plan commits to a fixed amount of spend per hour across eligible services and regions, and the commitment not used in an hour expires and does not roll over; the savings plan for compute does not cover software, networking or storage. Advisor recommends both from your recent use, and gives no reservation recommendation for resources that are regularly shut down.
The order matters, and Microsoft’s advice opens with it: right-size first, remove unused or oversized resources before buying any discount, because discounts reduce rates, not waste. A commitment calculated on today’s use, when today’s use includes machines that should be smaller or switched off, locks in the waste for one or three years.
There is a third lever that has nothing to do with use: licences. With Azure Hybrid Benefit, Windows Server licences your company already owns with active Software Assurance, or qualifying subscription licences, can be used for virtual machines in Azure, with at least eight core licences per machine. SQL Server licences with Software Assurance work the same way for Azure SQL, except in the DTU purchasing model and in the serverless tier. Whether your licences qualify is a question for whoever manages them, and it is worth asking.
Keep it from growing again
A review fixes the bill once. What keeps it fixed is a few habits and two features of Cost Management.
Budgets that warn
A budget in Cost Management sets an amount for a scope, such as a subscription or a resource group, with up to five thresholds, each on actual or on forecast cost. Budgets are evaluated every 24 hours, and the emails normally go out within an hour of the evaluation. Microsoft is explicit that a budget does not stop anything: when its thresholds are exceeded, notifications are sent, resources are not affected and consumption is not stopped. Acting automatically on a budget takes an action group, possible only at the scope of a subscription or a resource group. A budget is a warning, and it works only if the alert reaches someone who will act on it.
Anomaly alerts
Cost Management also looks for anomalies on each subscription, once a cost analysis smart view has been opened on it. Every day it compares the day’s usage with a forecast based on the last 60 days, and it runs 36 hours after the day ends. When it finds one, an alert can send an email, once. A cost that doubles overnight because of a forgotten test or a talkative log is found in a day or two, not at the end of the month.
A weekly look, and tags from day one
The rest is habit. Microsoft recommends checking costs weekly, and daily for the first week or two after you deploy or change resources. A short look at the daily cost and the top resources, with the owners. A tag with the owner on every new resource, and an end date for anything temporary. And the question «when will this be deleted?» asked when a resource is created, not two years later.
Where to start this week
None of this needs a project. These steps take a few hours and, apart from Advisor’s lookback and the budget, change nothing in your subscription.
Get read access to the costs
Ask for the Reader role on the subscription, which shows both the costs and the resources, or the billing role that shows your billing account, for whoever will do the review.
Look at the daily cost of the last three months
Find the days when the cost stepped up, and ask what happened on those days. Then the last year, month by month.
List the twenty most expensive resources
Grouped by resource, with amortised cost where your agreement shows it. Next to each, write who owns it and what it is for. The ones nobody can explain go first.
Open Advisor’s cost recommendations
Ask whoever administers the subscription to set the lookback to thirty days or more, wait for the list to update, and read it without applying anything. For each recommendation, ask the owner what it would break.
Look for what belongs to nothing
Unattached disks, old snapshots and public IP addresses associated with nothing. Write down why each one was kept before anyone deletes it.
Set a budget
On the main subscription, with an alert that reaches finance and IT. It needs a role that can create budgets, and it changes no resource — but the next surprise arrives as an email instead of an invoice.
How we do it
This is our Azure cost review. With read-only access to your Azure subscription, we read the billing of recent months and compare it with what is really used: virtual machines left running that no one uses, disks and snapshots that are no longer needed, databases with more capacity than they use, and data kept in a more expensive storage tier than it needs. We note each possible saving with its estimated amount and its risk: what could stop working if it is applied, and how to check that beforehand.
We go through the list with your technical team, who know which resources are critical, and hand it to you sorted by amount and by risk. Nothing is changed in your subscription during the review: which savings are applied, and when, is your decision, with the list in hand. The formula and the price are written down before we start.
If your bill keeps growing and nobody can say why, start with a free one-hour video call: tell us how your company uses Azure and what worries you about the bill, and we will tell you where we would look first.
The sources we cite
- Microsoft Learn: Quickstart: Start using Cost analysis, Customize views in Cost analysis, Understand Cost Management data, Understand and work with scopes, Assign access to Cost Management data, Azure built-in roles, Azure Advisor permissions and View amortized benefit costs.
- Microsoft Learn: Group and allocate costs using tag inheritance, Tutorial: Create and manage budgets and Identify anomalies and unexpected changes in cost.
- Microsoft Learn: Optimize virtual machine (VM) or virtual machine scale set (VMSS) spend by resizing or shutting down underutilized instances and Cost recommendations.
- Microsoft Learn: States and billing status of Azure Virtual Machines, Delete a VM and its attached resources, az vm, Auto-shutdown a virtual machine, Find and delete unattached disks in the portal, Create an incremental snapshot and Understand Azure Disk Storage billing.
- Microsoft Learn: Public IP addresses in Azure; Azure, IP Addresses pricing, its questions and answers.
- Microsoft Learn: Access tiers for blob data, smart tier, Blob rehydration from the archive tier and Azure Blob Storage lifecycle management overview.
- Microsoft Learn: Azure Monitor Logs cost calculations and options and Set daily cap on Log Analytics workspace.
- Microsoft Learn: Serverless compute tier for Azure SQL Database, What are Azure Reservations?, Reservation recommendations, What are savings plans? and Decide between a savings plan and a reservation.
- Microsoft Learn: Azure Hybrid Benefit for Windows Server and Azure Hybrid Benefit for Azure SQL.
What we do about it: Azure cost review