Skip to content
Refine

Product · Built by HabileLabs, an AWS Advanced Tier Partner

Find out what AI costs you before finance does

Bedrock tokens by model, AI developer seats by person, GPU compute by instance family — read out of the AWS billing export you already have. No agent, no provider API keys, no second cloud.

Why AI spend hides

Your AI bill is already here. It is just filed somewhere else

27 of the businesses Refine monitors carry AI spend on their AWS bill today. Almost none of them could have told you the figure from Cost Explorer, because three things conspire to keep it off the page.

It is not filed under "AI"

Third-party models on Bedrock bill as Marketplace transactions, so AWS files them under Other — not under AI and Machine Learning. A report that trusts the service category alone under-reports Bedrock several times over. We found exactly that when we measured it, which is why the classifier uses two signals instead of one.

The seats outspend the models

Kiro and Amazon Q Developer bill per user, per month, against an identity-store ARN rather than a resource — so AI developer tooling lands on the bill as a long list of anonymous user ids. On the accounts we monitor, those seats are the larger half of AI spend, and they recur whether or not anyone signs in.

Inference profiles hide the owner

A cross-region inference profile routes to whichever region has capacity and carries no team tag, so the spend lands with no owner and a region you did not pick. Application inference profiles can be tagged; system ones cannot, and knowing which you are looking at is the whole allocation problem.

What counts as AI spend, stated rather than implied

A FinOps number is only useful if you can say what is inside it. This is the whole rule, and it is one definition shared by the page, the budgets, the anomalies and the alerts — not a different answer per screen.

Counted

  • Everything AWS itself files under the AI and Machine Learning service category — Bedrock, SageMaker, Rekognition, Textract, Comprehend, Transcribe and the rest, including services AWS has not launched yet
  • Third-party foundation models on Bedrock, billed through Marketplace — Claude, GPT, Cohere and their siblings, which the category files elsewhere
  • Amazon Bedrock AgentCore, the agent runtime, and AWSDevOpsAgent — both billed under Other rather than under the AI category
  • Kiro, AWS’s agentic development tool — a product call, because at the time it was the second-largest AI line on the book and leaving it out understated AI share by about a third
  • GPU and accelerator EC2 — kept as its own figure, never folded into the one above

What the figure does not count

Spend billed to AWS AI services. GPU and accelerator compute is billed to EC2 and counted separately as AI workload compute; storage and networking supporting AI workloads are billed to S3 and networking and are counted in neither figure.

How fresh it is

AI figures refresh daily, on the same cadence as the rest of your bill, because they come from the same AWS billing export — which AWS itself lands with hours of lag. Anybody selling you real-time AWS cost is selling you the lag with a different name on it.

Tokens, by model, at your own prices

Growth and up

Bedrock rows already carry the token count and the meter that separates input from output. No new pipeline, no agent, no change to the permissions you granted — the data was in the export the whole time.

Input and output, separated

Bedrock bills input and output tokens on different meters, at different prices. Refine reads the meter rather than inferring it, so a model whose answers are getting longer shows up as a changing ratio instead of a vaguely larger bill.

Cost per million tokens

The one figure that compares two models honestly. Computed per model from your own usage and your own rates — not a price list we keep, which would be stale within a quarter and wrong silently.

Cache reads and batch

Prompt-cache reads and batch inference bill on their own meters at their own discounts. Both are broken out, so you can see what share of your traffic is already getting the cheaper rate before anyone proposes a change.

Model mix over twelve months

Which models you actually run, how the mix moved, and which one quietly became the majority. A migration from an expensive model to a cheap one is visible as a shift in share, not just a smaller total.

Allocating model spend to a team

Growth and up

An application inference profile is the only Bedrock resource you can put cost-allocation tags on. Refine tells you which of your model spend is routed through one, which is routed through an untaggable system profile or a bare foundation model, and what the untagged share costs you per month — then that spend joins the same showback report as the rest of the bill, split by the dimension your org organises around.

The recommendation that comes out of it is a configuration change in your account, not a button here: Refine onboards read-only and does not hold the permission to create a profile for you.

The AI line nobody budgets for

Growth and up

Kiro and Amazon Q Developer bill per user per month against an identity-store ARN. On the accounts Refine monitors, those seats are the larger half of all AI spend — and unlike a model call, a seat costs the same whether the person used it forty times or not at all.

  • Seat count and cost per seat, per month, for Kiro and Amazon Q Developer
  • The roster behind the count, resolved from your identity store rather than left as user ids
  • Seats added since last month, as an alert — each one recurs until somebody removes it
  • Seats billing for people who have left, flagged against the identity store

Who holds a seat is identifying information about your staff, so the roster is shown to admins and owners only — a teammate on the User role sees the counts and the trend, never the names.

AI developer seats

Illustration
Kiro seats24 · $46.00 ea
Amazon Q Developer4 · $19.00 ea
Added since last month+6
Billing for leavers2

What it adds up to

Two seats still billing for people who have left is $1,104 a year that nobody will ever notice on an invoice line that reads the same every month.

Five things worth hearing about, and nothing else

AI spend is small and spiky, so a detector tuned for a $26k bill either never fires on it or fires constantly. The two rate-based types — token spike and ratio drift — therefore need both a multiple of your own trailing baseline and an absolute token floor; seat growth needs a minimum number of seats added. The two awareness types, a new model and a new region, have no threshold at all: they are either on or off, because a model you have never run before is worth a line whatever it cost. Every control is yours to change.

Token spike, per model
One model’s tokens per day running at a multiple of its own trailing rate — with an absolute token floor, so a $3 experiment that triples does not page anyone.
New model first seen
A model or inference profile with usage this month and none in the three months of history the snapshot keeps. One notification per month per model — and a model you stopped using for a whole quarter and picked up again counts as new, because by then the history no longer holds it.
Bedrock in a new region
Cross-region and global inference profiles route silently. This is the alert that tells you your prompts started leaving the region you chose.
Output / input ratio drift
The prompts stay the same size and the answers balloon — the signature of a retry loop or an agent loop, which a dollar figure alone will not show you until the invoice arrives.
AI developer seats added
Headcount, not dollars. Five new seats is something to know about before the renewal, not something to wake up for.

Budgets that scope to AI

Growth and up

Set a budget on all AI services, on one Bedrock model, or on one inference profile — then get the same threshold, pace and forecast alerts the rest of your budgets give you. The scope means exactly what the AI Spend page means, because both ask the same classifier; a budget that quietly counted different rows than the page would be worse than no budget.

Alerts route through the channels you already configured — Email, Slack, Microsoft Teams — with the same quiet hours, grouping and per-event mutes.

What it is costing you to leave things on

Every figure here is reproducible from the billing row it came out of. Where a number is a projection rather than a measurement, the recommendation says so on its face.

Idle provisioned throughput and endpoints

Bedrock provisioned throughput and SageMaker endpoints bill by the hour whether or not anything calls them. What your bill cannot say is whether they were busy — capacity hours look identical either way — so provisioned throughput is reported with its commitment cost as a CEILING rather than as a saving, the endpoint signal was a billing proxy for invocations until model health started measuring it — a measured zero now confirms the finding and measured traffic withdraws it — and the idle-notebook finding carries no dollar figure at all. Each card says which of the three it is.

Batch and prompt caching, priced

Batch inference is half the on-demand token price across every model that offers it. Refine prices what your current traffic would cost at that rate, and labels the prompt-cache estimate as the projection it is — caching only pays when prompts genuinely repeat, and billing data cannot say whether yours do.

Bedrock configuration findings

Model invocation logging off, no guardrails configured, no application inference profile to allocate spend with. Raised only in regions where your bill shows Bedrock actually being used — an absence is a finding only where the service is in use.

GPU and accelerator compute

Every accelerated instance family — g and p for GPUs, inf and trn for AWS’s own inference and training silicon, dl for the Deep Learning instances — identified and totalled as its own figure. Deliberately never added to the AI-services number: what you bought from an AI service and what you ran AI on are different claims, and one table would invite somebody to total the column.

Whether it is working, not just what it costs

Growth and up

Billing data has no idea what a retry is. CloudWatch does — so model health reads it directly: per Bedrock model and per SageMaker endpoint, per region, over the last seven days.

Errors and throttles, per model

Client errors, server errors and throttled calls as rates against your own invocation count, with an alert on each — tuned so a model called three times never pages anyone. A throttled call bills nothing, which is exactly why the bill cannot show you this.

Latency at p50 and p99

What a typical call feels like and what the worst one percent feels like. A mean would hide both, and the gap between the two is usually the thing worth acting on.

Cost per successful invocation

Your tokens for the window, priced at this month’s cost per million from your own bill, over the calls that neither errored nor were refused. A model failing a quarter of its calls shows a third more cost per success — which is where that waste actually lands, so there is no separate waste figure to invent.

SageMaker endpoints, variant by variant

Invocations, 4XX and 5XX rates, and model latency kept apart from SageMaker’s own overhead — a slow model is your code, slow overhead is instance sizing or cold starts. Production and shadow variants stay separate rows, because averaging a canary into its baseline hides the exact thing you deployed it to watch.

It makes the idle-endpoint finding honest

The idle-endpoint recommendation infers idleness from hosting hours billed with no inference-request charges, which is a proxy and wrong in both directions. Measured against CloudWatch it stops being one: a measured zero confirms it, and measured traffic WITHDRAWS it — a false "delete this endpoint" on something serving requests is the most expensive mistake that check can make.

What it cannot see, named

CloudWatch is read in the AWS account your connector stack was created in; your billing export covers the whole organisation. Where the two differ, the page says so — "4 of 9 billed models have metrics" — rather than leaving a shortfall that reads like an outage.

The one thing on this page that needs a permission change

Everything else here reads your billing export. This reads CloudWatch — both the Bedrock and the SageMaker namespaces, through one extra action, cloudwatch:ListMetrics, on the read-only role you already granted. A connector created from today’s template has it; an older one takes a single CloudFormation stack update, which Refine offers you in Settings with the change set to expect. It adds that one read and touches nothing else — not your export, not your bucket, not your cost data — and until you run it, this section says so rather than showing you an empty table.

CloudWatch metrics are counts, sums and percentiles. Bedrock invocation logs are the only place a prompt or a response exists, and nothing Refine holds can read them.

Seeing the problem is free

On every plan

The AI Spend page itself: what AI costs, its share of your bill, and the breakdown by service, model, account and region, across twelve months. Free under $2,000/month of AWS spend.

On Growth and above

Growth and up

Token analytics, the inference-profile table, the seat roster, model health, AI budget scopes and the AI alert settings — the parts you use to act on what the free view shows you.

What Refine does not do

Four things you will find elsewhere and will not find here. If one of them is what you actually need, the honest answer is that another tool fits better.

No direct OpenAI or Anthropic API spend
If you call those providers outside AWS, that spend is on their invoice, not in your AWS billing export — and bringing it in would mean asking you for provider API keys or wiring a second ingestion path. We decided against it. Tools that do offer it are being honest about a real gap in ours; we would rather name the gap than ask for the keys.
No second cloud
Refine is an AWS product. Azure AI Foundry and Vertex spend are not here and are not planned. If your AI runs across three clouds, a multi-cloud platform is the right tool and we will say so.
No agent, and no write access
Nothing is installed in your cluster and nothing is deployed in your account beyond a read-only role and a billing export. The same boundary that makes Refine easy to approve is the reason it cannot see what only an in-cluster agent could.
No prompt or response content, ever
Refine reads billing rows, plus CloudWatch metric numbers — counts and percentiles — where a check needs them. Bedrock invocation logs are the only place a prompt or a response exists, Refine does not read them, and the read-only role you grant carries no permission that would let it.

Frequently Asked Questions

  • No. Every figure on the AI Spend page comes from the FOCUS billing export and the read-only role you already granted — the same connector that powers the rest of Refine. Connect an account and the AI view is populated on the next report run, with twelve months of history behind it.

See the AI line on your own bill

Connect an AWS account and the AI view fills in from twelve months of your own export. Free under $2,000/month of AWS spend.

Refine is built and supported by HabileLabs, an AWS Advanced Tier Services Partner.