Savings
Read the savings findings your providers recommend and the ones Costfluent reads from your bill, record your own, move each one from open to accepted, implemented or declined, and read what a measurement found.
Savings sits under Optimization in the product's sidebar. It is a queue of savings findings: each one names cost that can be removed, the evidence behind it, and who owns it. Findings come from four places: once a day Costfluent reads the recommendations AWS Cost Optimization Hub, Azure Advisor and Google Cloud Recommender already compute for your connections, and it reads your AWS and Azure bills itself for savings no provider lists; people in your organization author their own; and on a plan with curated savings Costfluent adds findings from its own analysis of your cost.
Before you start
- Any team role can view the queue. Authoring, editing, accepting, implementing, declining and retiring a finding need the team role Editor with edit access to the workspace, as does assigning its owner.
- The queue is available on every plan. Measuring a finding's saving is not; see Measured savings below.
- Findings from AWS need Cost Optimization Hub turned on for the account and
cost-optimization-hub:ListRecommendationson the connection's role (AWS); while the hub is off, the empty queue names the connection and the step that turns it on (AWS); findings from Azure need nothing beyond Cost Management Reader (Azure), and for a connection collecting a billing scope they need Cost Management Reader on each subscription the scope covers, because Advisor reads per subscription; a subscription the principal cannot read is reported as on any Azure connection (Azure); findings from Google Cloud need the Compute, project committed use discount and Cloud SQL Recommender viewer roles for Costfluent's account on each project to read, which the Google Cloud module grants on the projects listed in itsrecommender_project_idsinput (Google Cloud). - Findings read from your bill need nothing beyond the AWS or Azure connection that collects the cost.

Findings from your providers
Every day at 05:00 UTC Costfluent lists the recommendations of each AWS, Azure and Google Cloud connection a workspace reads, for the accounts, subscriptions and projects that workspace reads, and keeps the queue in step with them:
- Each recommendation becomes an Open finding showing its source (AWS Cost Optimization Hub, Azure Advisor or Google Cloud Recommender), what it asks for (Rightsizing, Idle resource, Newer generation, Storage, Commitment purchase or Other), the resource it is about, and the provider's monthly estimate in the provider's currency. The account, subscription or project it is in shows on a line of its own, and its Owner reads Unassigned until you assign one.
- A recommendation estimating less than 5 a month in its currency is left out.
- While a finding is open it takes the provider's latest estimate and description each day, and it cannot be edited. Once you accept it, its estimate stays what you accepted.
- When the provider stops recommending an open finding, its status becomes Withdrawn; it reopens if the provider recommends it again. A finding you declined or retired is never recreated.
- If a connection cannot be read, nothing changes in the queue, and the connection's Savings recommendations check names the missing permission. Savings names the connection too, with links to it and to how to fix it: in place of the empty queue, or under Savings recommendations could not be read above a queue that holds findings.
- While your organization's collection is paused (Plans and billing), or a connection is paused over the connection limit set for your organization, its recommendations are not listed: its findings stay as they are, none is withdrawn, and the daily listing reads the connection again once collection resumes.
For Google Cloud, Costfluent reads the projects, regions and zones the connection's cost came from in the last 30 days: virtual machine rightsizing, idle virtual machines, disks, IP addresses and images, committed use discounts, and idle or overprovisioned Cloud SQL instances. A project not granted to Costfluent is skipped, and no connection check names it; while one is skipped, no Google Cloud finding of that connection is withdrawn. Savings of a workspace that reads a Google Cloud connection says so, in place of the empty queue or above its findings, and links the Recommender roles.
Findings read from your bill
Every day at 05:15 UTC Costfluent also reads the bill of each AWS and Azure connection a workspace reads, up to the day the connection's data reaches, yesterday at the latest because today's cost is still arriving, for five savings AWS Cost Optimization Hub lists only with Compute Optimizer turned on, or no provider lists at all: four from the last 14 days of an AWS bill, and extended-support fees from the last 7 days of an AWS or Azure bill. Each becomes an Open finding with the source Bill analysis by Costfluent, the resource it is about, and a monthly figure worked out from what the bill charged. The account or subscription it is in shows on a line of its own, and its Owner reads Unassigned until you assign one. Google Cloud bills are not read for these findings.
- Idle public IPv4 addresses, kind Idle resource: addresses AWS billed as idle. One address costs too little to list alone, so the idle addresses of one account and region make one finding, which names up to 50 of them, the most expensive first, and counts them all. Its figure is their idle charges over the last 7 days, scaled to a month. Declining it also hides every later idle address in that account and region.
- Idle load balancers, kind Idle resource: an application, network or gateway load balancer billed on at least 13 of the 14 days whose capacity-unit hours stay under 1% of its billed hours, or a classic one that processed under 1 GB. Its figure is every charge on the balancer in those days, scaled to a month.
- Idle NAT gateways, kind Idle resource: a NAT gateway billed on at least 13 of the 14 days that processed under 1 GB. Its figure is its hourly and processing charges in those days, scaled to a month.
- gp2 volumes, kind Storage: a gp2 volume of at most 1,000 GiB billed on at least 13 of the 14 days. gp3 costs 20% less per GB-month and includes 3,000 IOPS, the most such a volume has as its baseline, and 125 MiB/s, so its figure is 20% of its gp2 storage charges in those days, scaled to a month.
- Extended-support fees, kind Extended support: what AWS charges for an EKS cluster, an RDS or Aurora database, an ElastiCache cluster or an OpenSearch domain, and Azure for an Azure Database for MySQL or PostgreSQL server, while it runs a major version past the end of standard support. Its figure is the fee's daily average from the first of the 7 days through the last day the bill charged it, scaled to a month, because upgrading to a major version in standard support ends it; the days after the last day the bill charged the fee are left out of the average rather than counted as days without it. The upgrade itself can change the resource's other costs, which the figure leaves out. Its evidence names the charge as the bill words it and the fee's daily average.
An idle resource or gp2 volume must be billed on each of the last three of those days, and a resource paying an extended-support fee must have paid it on at least one of them, so one you delete, convert or upgrade leaves the queue within three days: its open finding becomes Withdrawn. A gp2 volume or NAT gateway that already has a finding from AWS Cost Optimization Hub in the workspace, in any status but Withdrawn, declined and retired ones included, gets no second finding from the bill, and an open one from the bill is withdrawn once the hub lists the resource. As with your providers' findings, a figure under 5 a month in its currency is left out, an open finding takes the latest figures each day, and a declined or retired one is never recreated. While the connection's data is more than four days old, or a reading of the bill fails, the findings are refreshed but none is withdrawn.
The bill does not show whether an address or volume is attached, or how much throughput a volume uses. Check that a gp2 volume needs no more than 125 MiB/s before converting it, and that an idle load balancer is not a standby you keep on purpose.
Findings from your Kubernetes clusters
Every day at 06:00 UTC Costfluent also reads the last 30 days of each Kubernetes cluster a workspace reads, from the CPU and memory its agent measured for each container, and suggests lower requests for a container that asks for more than it uses. Each suggestion becomes an Open finding with the source Kubernetes analysis by Costfluent and the kind Rightsizing, about one container of one workload: its namespace, controller kind, controller and container. Its cluster shows on a line of its own, and its Owner reads Unassigned until you assign one.
- A container needs rows on at least 7 of the 30 days, one of them in the last 2. A bare pod and a Job no CronJob owns have no template to change, so they get no finding.
- Of a cluster's containers with a row in the last 2 days, the 5,000 that cost the most over the 30 days are assessed. The rest get no finding, and an open finding of one of them becomes Withdrawn until it is among the costliest again.
- The suggested CPU request is the 95th percentile of the container's daily CPU peaks plus 15%, rounded up to 10m. The suggested memory request is its highest daily memory peak plus 15%, rounded up to 16Mi, because a memory request set too low ends in an eviction or an out-of-memory kill. The current request is the one of the container's latest day, and a request the rule does not lower stays as it is.
- A finding opens when a suggestion is below 80% of the current request, and stays open while one is below 90%. Its title and evidence name the current and the suggested request in Kubernetes notation, CPU in millicores and memory in mebibytes, such as
1000mto580mor1024Mito448Mi. - Its figure is what allocation would have charged the container less over the 30 days at the suggested requests than at its current ones, scaled to a month, in the cluster's currency. The freed requests become idle node capacity, so the saving is realised once the cluster runs fewer or smaller nodes.
As with your providers' findings, a figure under 5 a month in its currency is left out, an open finding takes the latest figures each day, and a declined or retired one is never recreated. An open finding whose container stopped running, or whose requests no longer fall far enough, becomes Withdrawn. While a cluster has no row in the last 2 days, none of its findings changes.
Costfluent changes nothing in your cluster. Apply a suggested request yourself, and check it against bursts inside a day, which a daily peak cannot show.
Author a finding
- Open Savings and select Author finding.
- Fill in Finding, Evidence, Estimated monthly saving with its Currency, and Owner.
- Rate Confidence, Risk and Effort as Low, Medium or High.
- Select Author finding. The finding starts as Open.
Evidence is required: an estimated saving nobody can check is not a finding.
Assign an owner
The owner is who acts on a finding, a person or a team. A finding you author has the owner you wrote, a curated one starts with the owner Costfluent wrote, and a finding from your providers, your bill or a Kubernetes cluster has none until you assign one.
- In the row's menu, "Actions for" and the finding's title, select Assign owner. An open finding you authored takes its owner in Edit instead.
- In Assign an owner, enter Owner, at most 200 characters, and select Assign.
Any finding that is not declined or withdrawn takes an owner, whatever its status. An owner changes none of the finding's figures. A finding from a provider, your bill or a cluster keeps its account, subscription, project or cluster on its own line, and while it is open each daily read still updates its figures but keeps the owner you assigned. The Public API's owner is the owner assigned, or the account while nobody is.
Move a finding through its states
A finding's row shows How to apply and its next step, Accept or Mark implemented. Edit or Assign owner, Decline and Retire are in the row's menu, "Actions for" and the finding's title.
- Accept an open finding when you decide to act on it.
- Mark implemented an accepted finding once the change is made, after Mark this finding implemented?: an implemented finding cannot go back to Accepted or be declined.
- Decline an open or accepted finding from the row's menu, with a required Reason. The dialog says a decline is final. The reason stays visible under the finding, and its status reads Declined.
- Edit a finding you authored while it is still open, from the row's menu, in Revise finding.
- Retire a finding in any state from the row's menu to remove it from the queue, after Retire this finding?.
Pointing at Accept or Mark implemented says what the step does. A step the finding's current status no longer allows is refused with the reason, which names the status it can move from; a step that fails for any other reason says "Could not update the finding".
To answer several findings at once:
- On To review or In progress in Savings, tick the box of each finding, or the box in the table's header to select every finding on the page. Verified and Closed offer no selection, since no finding there can be accepted or declined.
- In the bar above the table, which shows how many are selected, select Accept or Decline; In progress holds no open finding, so its bar offers Decline only. Clear selection unticks them all.
- For Decline, check the number of findings and their estimated monthly saving per currency, enter one Reason for all of them, and select Decline.
Accept applies to the selected open findings and Decline to the selected open and accepted ones; a selected finding in any other status is left as it is. Accepting asks for no confirmation. The selection covers the page shown, and changing the tab, a filter, the order or the page clears it.
A decline is final. The one reason is recorded on every finding it declines, each moves to Closed with the status Declined, and a declined finding is never recreated, even when its source recommends it again.
When a finding cannot take the step, because its status changed after the page was read (a colleague answered it first, for example) or because it is no longer in the workspace, the other findings are still answered, and a dialog lists each finding left unchanged by its title, with the reason.
Apply a recommendation
Costfluent changes nothing in your accounts and asks for no permission to. For a finding that is not declined or withdrawn, How to apply on its row opens How to apply this finding: the commands to run, the change to make in your code, and a prompt for a coding agent, each with its copy button.
- Open Savings and select How to apply on the finding's row. The panel says when the provider last recommended the finding.
- Read the note above the steps: "Review each step before you run it. Costfluent does not run these commands and cannot see your code."
- On Steps, copy each command in order. Every AWS, Azure and Google Cloud command names its region, subscription or project. Open in the provider’s console opens the resource, or the provider's own list of recommendations.
- On Terraform, or Kubernetes manifest for a container, copy the attribute and value to set where your code declares the resource.
- On Agent prompt, copy the prompt into a coding agent. It asks the agent to confirm the account first, find the code that declares the resource, change only that attribute, never run
terraform apply,terraform destroyor the provider's command line, and open a pull request for review. - Once the change is deployed, Accept the finding if it is still open, then select Mark implemented and confirm it. Both steps are on the finding's row and at the foot of the panel. Costfluent then measures the saving (Measured savings below).

What a finding's steps hold depends on what it recommends:
- EC2 rightsizing: stop the instance, change its instance type, start it again, and
instance_typein Terraform. A move to a Graviton (arm64) type gets a check of the instance's architecture and the code change instead, because an x86 image cannot start on arm64. An idle instance gets the command that stops it. - EBS volumes: a gp2 volume read from your bill gets
aws ec2 modify-volumeto gp3 andtypein Terraform; an unused volume gets a snapshot, then its deletion, and the resource to remove from Terraform. - Lambda: the new memory size and
memory_sizein Terraform. - Azure virtual machines that Advisor right-sizes or finds idle:
az vm resizeandsizein Terraform, oraz vm deallocate. - Compute Engine: a new machine type with
machine_typein Terraform, and idle instances, disks, addresses and images to stop or delete. Cloud SQL: a smaller tier, or the activation policyNEVERfor an idle instance, with thesettingsblock in Terraform. - Kubernetes containers: the
resources.requeststo set in the manifest or chart values, andkubectl set resources, which changes the running workload until the manifest follows. - Commitments: the term and the other parameters the provider gives, and its purchase page, never a command: buying one is a decision for a person.
- Every other finding, the ones you author and curated ones included: the provider's console where Costfluent can link the resource, and the prompt.
Only identifiers in their provider's exact format, such as an instance ID, a region or a machine type, enter a command or the prompt. A finding's title, evidence and the provider's own descriptions never do, because a coding agent acts on what the prompt says. A finding whose values are in any other format gets the console link and the prompt without commands. The prompt is in English in every language of the console, since coding agents work in English.
Read the queue
- The tabs split the queue by stage and show how many findings each holds: To review holds open findings, In progress accepted and implemented ones, Verified verified ones, and Closed declined and withdrawn ones. The page opens on To review.
- All sources filters the queue to one source, such as Authored, AWS Cost Optimization Hub or Bill analysis by Costfluent, All kinds to what a finding asks for, and All connections to the findings one connection's recommendations, bill or cluster generated. The tab counts follow these filters.
- Minimum monthly saving leaves out findings that estimate less a month. It applies when you press Enter or leave the field, and an empty field or 0 shows every finding again. A negative value, or text that is not a number, is not applied: the field goes back to the minimum in use and says so. The help beside the field says each finding is compared in its own currency.
- Newest first or Largest saving first orders the whole queue. It shows 50 findings a page; Previous and Next move between pages.
- The tab, the filters, the order and the page are kept in the page's address, so reloading or bookmarking the page keeps them.
- The totals, one card per currency, cover every finding of the selected tab that the filters keep, not only the page shown: the Open estimate on To review, the Accepted and Implemented estimates on In progress, and Verified (measured) on Verified, each with the number of Findings. Closed shows no totals: a declined or withdrawn finding has no saving to sum.
- The table lists each finding with its source, its risk and effort, each rated low, medium or high, and, for a finding from a provider, read from your bill or from a Kubernetes cluster, its kind, resource and account, then its Owner (Unassigned until one is assigned), Estimated monthly saving, Measured saving and Status. While an issue tracker is connected, Issue shows the finding's issue, or Create issue on an open or accepted finding.
Saved views
A saved view keeps a tab of the queue and its filters, All sources, All kinds, All connections and Minimum monthly saving, under a title. Views belong to the workspace: everyone who can view the queue lists and opens them, and saving, renaming and deleting one need the team role Editor with edit access to the workspace, as authoring a finding does. The order and the page are not part of a view.
- Set the tab and the filters you want to keep.
- Open the More menu beside the list of views and select Save as new.
- Enter a Title of at most 200 characters and select Save. The queue now shows the view, and the page's address names it.
- Open a view from the list of views beside the filters, or choose All findings for the queue without a filter. A view's link opens it too: Copy link copies one that also names the workspace, so it opens there for everyone who can view the queue.
- Changing the tab, a filter or the order leaves the view: the list of views shows its title followed by “(changed)”, and Save writes the queue as it now stands back to the view. Rename changes only the title. Delete removes the view for the whole workspace, after Delete this view?, and shows every finding again.
- A view that names a connection removed since keeps it and lists no finding; the connection filter then reads Removed connection. A link to a deleted view opens the queue with every finding.
Curated findings
Planned: curated findings
Included in the Business and Custom plans. Compare plans
Costfluent reviews your cost and writes what it finds into a workspace's queue. A curated finding carries the source Curated by Costfluent and its own evidence, estimate and owner, and moves through the same states as any other finding: you accept, implement, decline or retire it. You cannot edit it, because its estimate and evidence are Costfluent's, but you can assign it an owner of your own.
On a plan without curated savings the queue shows no curated findings, including ones written before the plan changed. Your own findings and your providers' stay in the queue on every plan.
Measured savings
Planned: measured savings
Included in the Custom plan. Compare plans
An estimate is never reported as a saving. Costfluent measures an implemented finding by comparing its committed cost before and after the change; the product has no action that does this or that marks a finding verified. Measured saving reads Not measured until a measurement exists, and then shows its outcome:
| Outcome | What the measurement found | The finding |
|---|---|---|
| Realized | A reduction that meets or exceeds the estimate | Becomes Verified, with the measured saving |
| Partially realized | A real reduction, smaller than the estimate | Becomes Verified, with the measured saving |
| Not realized | No reduction, or cost rose | Stays Implemented |
| Inconclusive | The movement cannot be attributed to the change | Stays Implemented |
A finding from a provider is measured on its resource, or on its account for a commitment purchase. A finding from a Kubernetes cluster is measured on the whole cluster: resizing a container frees part of a node, which the cluster then shows as idle, so only a smaller bill for the cluster, from fewer or smaller nodes, counts as saved. A finding can be measured again; each measurement is kept.
Each finding is measured on the cost basis its estimate was made on, so the saving compares like with like (billed or effective cost). A finding you author is measured on billed cost. A finding from your providers or read from your bill is measured on amortized cost, and a curated finding on the basis Costfluent estimated it on. On amortized cost, a commitment's saving is its discount on the usage it covers. A finding from your providers that stops an idle resource is measured on billed cost instead: usage a commitment covers costs nothing before and after, so stopping such a resource saves nothing until the commitment ends, while amortized cost would count its share as saved.
A commitment finding you author is measured on billed cost, which shows usage a commitment covers as free and an upfront commitment in full on the day it was bought. Its measured saving is then the whole covered spend rather than the discount, or, when the commitment was bought inside the measured window, it can read Not realized. A commitment's saving is measured as its discount only on a finding from your providers, or on one Costfluent curated on amortized cost.
A provider's finding that stops or deletes a resource is measured even when the resource has little or no cost after the change, as long as the cost of those days is complete and the resource's account still reports cost on at least four in five of them; a removed resource then counts its whole earlier cost as saved. If the account no longer reports cost, the measurement is Inconclusive, because the cost may have left what Costfluent collects rather than stopped. For any other finding the measurement is Inconclusive when what it measures has no cost on more than one day in five after the change, for example because its resource was replaced under a new ID.
How it behaves
- Only a measurement sets Verified, and Verified (measured) totals measured savings only, never estimates.
- Declined, verified and withdrawn findings take no further step except Retire.
- Retiring hides a finding from the queue; it does not change any cost figure.
- Minimum monthly saving and Largest saving first compare each finding's estimate in its own currency, without converting it, so with findings in two currencies 100 USD and 100 EUR count as equal.
- Sorting by a column header orders only the findings on the page shown.
- A finding's steps are read from the provider's latest recommendation while the finding is open, and kept as they were once you accept or decline it. A finding answered before Costfluent read these details gets the console link and the prompt without commands.
- AWS Cost Optimization Hub does not name an EBS volume's type in the recommendations Costfluent reads, so its EBS upgrades get the console link and the prompt.