A software company
What does one customer cost us in AI each month?
$1.84a customer, a month
Three AI agents in the product, each call labelled with the customer it served. The finance lead owns the number, and the next price review starts from it.
Your AI agents pay big-model prices only for the work that needs a big model.
Much of what a growing business asks its AI agents to do is routine: sort an email, read an invoice, draft a reply. First we put a cost on every task, team and outcome. Then we test every model on your own cases and send each job to the smallest one that passes, so the bill comes down without losing quality.
A one-day workshop with the engineers who’d do the work. AWS funding typically covers the workshop. An AWS Premier Tier partner, one of about 300 among more than 130,000, and a Databricks consulting partner.
Most of the fruit hangs low. In this orchard small AI agents pick it, one on a ladder picks the middle branches, and a hired crane reaches only the top.
The demo ran on the biggest model.
So has every call since.
In most growing businesses AI agents arrive one at a time, each wired to the biggest model because that’s what worked in the demo. One request to an AI agent becomes a plan, a few tool calls, a retry and a summary, and each is a model call carrying the same instructions. The biggest model is priced by the word, and every new AI agent adds calls of its own, so the bill grows faster than the business. Most of the fruit hangs low, and every apple still gets the crane.
- The biggest models cost 10 to 100 times more per token, the unit AI is billed in, than the smallest from the same maker, at October 2026’s list prices.
- Sorting, tagging, pulling a field or writing a short summary doesn’t need them.
- The bill here is an example. The workshop reads your own.
AI model billOctober, one total
this month, every call on the biggest model
A year ago $4,200. Customers since then: up 38%.
Example
Every other cost has a unit and an owner.
The AI line arrived with neither.
Every AI agent draws from one tower of AI models, through one valve house: a meter on each team’s line, and a valve that holds.
In most growing businesses the AI bill has no forecast either. It sits between an engineering card and a cloud bill nobody reconciles, and it grows with use, not with headcount.
One customer asks where their order is. The AI agent plans a reply, reads the order, checks the courier, retries, drafts, checks its draft and writes a summary for the CRM. That’s a dozen model calls or more, and none of them says who asked. It’s a whole street on one water meter: everyone draws from it, nobody owns it, and the bill gets split by guesswork.
- Shared API keys make the bill impossible to split, because every team’s calls arrive as one.
- Finance can’t allocate it. Engineering can’t defend it. So it gets argued about instead of managed.
- That’s why we measure before we cut, giving every call a team, a job and the outcome it served.
Cost per token is engineering’s number.
Cost per outcome is the board’s.
Once every model call carries a team, a job and the request it belongs to, the questions finance and the board ask get a number. Six examples from six kinds of business, each a question its owner couldn’t answer before.
An online shop
What does a ticket resolved by an AI agent cost?
31¢a ticket resolved
Every call of every request is counted, retries too, against the tickets the help desk marks resolved. The support lead owns it.
A wholesaler
Is the AI agent that drafts our quotes worth it?
$2.90a quote won (38¢ a quote sent)
The CRM counts a quote as won, not the AI. Against an average order of $2,300, the sales lead keeps it without a meeting.
An accounting firm
Which client’s month-end costs us most in AI?
$212one client, in October
Its receipts arrive as phone photos, so the AI agent reads each one twice. A partner decides whether to fix the photos or price it into the fee.
A property manager
What will AI cost next quarter with 40 more buildings?
$6,900a month, give or take 8%
Cost per repair request, times the requests 40 more buildings bring. It goes in the budget before the contracts are signed.
Any business
What’s still running from March?
$380a month, two experiments
Two AI agents from a trial nobody finished were still calling a model every hour. Their owner switched them off the day they showed up on her line.
Example businesses and figures.
Easy jobs go to small models.
Hard ones stay on the biggest.
We test each job on your own cases, and it runs on the smallest model that passes. Here are six examples from six kinds of business, each priced as one request at October’s list prices.
A property manager
Tenant messages sorted
$18.80 $0.90 per 1,000
The small model sorts 97 of 100 the way the team would. The biggest scored 98, at 21 times the price.
A software company
Support replies drafted
$18.80 $1.96 per 1,000
The router reads each ticket first and sends up anything the small model isn’t sure of. A person sends every reply.
A marketing agency
First drafts in each client’s voice
$18.80 $5.10 per 1,000
We cache each client’s style guide, so after the first read the agency pays about a tenth of the price for it. A writer finishes every draft.
A wholesaler
Orders read from emails
$18.80 $0.90 per 1,000
The AI agent only reads the email. Prices, discounts and totals are ordinary code, so the sums never touch a model.
An insurance broker
Policy questions answered
$18.80 $9.70 per 1,000
The smaller models got 19 of 100 of the broker’s own questions wrong. So this job stays on the biggest, and a broker signs off every answer. Caching the policy wording still halved its cost.
An online shop
Returns handled, under a ceiling
Never more than 3¢ a case
Each case has a hard spend limit. A thread going round in circles stops at it and goes to a person, instead of running up the bill overnight.
A cheaper model is one line of code.
Keeping it right takes six decisions.
Anyone can point every request at a cheaper model. Switch without testing, and it gets more wrong without anyone noticing. The first decision measures what each job costs, and the other five bring that cost down without changing what it gets right. Our engineers make them with you, the way AI Inference Engineering builds every system, starting from our Cost Control Tower, which is already built and runs in your own AWS account.
Support repliesA software company
per 1,000 replies
97 of 100 right, before and after
Example figures
- 1
Measure first: a cost on every task, team and outcome.
Your software calls one gateway in your own account instead of each provider, and an AWS Service Control Policy refuses any call that doesn’t. We retire the shared API keys, so each AI agent and each app calls as itself. Depending on your models, the gateway is Amazon Bedrock AgentCore Gateway or LiteLLM, an open gateway, with Amazon Bedrock behind it, so you’re never tied to one model provider. It tags every request with its team, product, use case and AI agent, taken from the identity of whoever is calling, so no developer has to remember to add them, and the bill gets a line for every job.
- The trace records which model answered, the tokens each way, what it cost and how it ended.
- Inside an AI agent one request id follows every call: the plan, the model’s turn after each tool’s result, the retry and the summary, each with its model and its cost. Those calls add up to the one request, so a request that became twelve calls shows as twelve.
- With finance in the room we write down what a task and an outcome are for each AI agent: a ticket resolved, an invoice matched, a quote won, each an event your own systems already record. Cost per outcome is every call of every request that led to it, the failed and retried ones counted in. Finance signs the definitions, and engineering builds to them, so the board gets a number finance trusts, not a range engineering defends. The scorecard puts the cost of each finished task beside its result.
- The labels land where finance already looks: AWS Cost Explorer, Budgets and the Cost and Usage Reports, with each call’s detail in Amazon CloudWatch. AWS CloudTrail ties every call to the person who asked.
- Next month is each team’s cost per outcome times the volume it expects, so a new customer or a busy season shows up in the forecast before it shows up on the bill. When a team’s cost per task jumps, after a prompt change, a new model or a loop, its owner hears that day. No month-end surprise, and a better answer for the board than “we’re working on it”.
- Model provider invoices, AI subscriptions on company cards and the cloud bill go into one ledger, with any AWS commitment going unspent beside it. We count the share of AI spend that still arrives some other way, and drive it down as teams move onto the gateway.
- Nothing is hosted by us. The gateway inspects prompts and answers in your account, they never leave it, and only the metadata goes to telemetry, in your own tools.
WhySpend that shows up on nobody’s line stops, and so does the AI agent someone wired up on a personal card. You can’t cut what a job costs until you can see what it costs.
- 2
Every model sits the same test, on your own cases.
Before anything changes, we agree three numbers for each job: the accuracy bar, the slowest reply you’ll accept and the cost per 1,000.
- The test is a golden set from your own work: real requests, each with the answer your people signed off.
- The rare and costly cases are in it on purpose.
- Big and small models, rented and open, are scored on what they got right, what each right answer cost and how fast they replied.
WhyA cheaper model can’t quietly get more wrong.
- 3
Each request goes to the smallest model that passed.
A router sends it there and keeps the hard ones on the larger model. It decides each step inside an AI agent separately, so the plan can stay on the larger model while pulling a field or writing the summary goes to a small one.
- Before the call, a small classifier trained on your cases reads the request and spots the job and how hard it is.
- After the call, the router checks the answer against your rules and your format before it goes out, and a sample is scored every day.
- If the model wasn’t sure, the request goes up, and the case joins the next round of tests.
WhyThe biggest model stops doing easy work, and a small one isn’t left guessing at hard work.
- 4
Send less, and pay a fraction for what repeats.
We cache the instructions and documents every request repeats, so they’re re-read at a fraction of the price. Each step of an AI agent re-reads everything before it, so we cache the shared start and pass each step only the results it needs.
- Known rules and sums stay as ordinary code, because a model only adds cost and variation they don’t need.
- One model call makes one judgment, such as tagging a request or pulling a field, with only the context that judgment needs.
- Where the steps are known, code owns the order and the model works inside a step. We keep an AI agent that plans its own steps for work whose path changes every time, because every step it plans is another call.
- Answers are only as long as the job needs, often one label or a few lines. At the big makers’ list prices, a token out costs five times a token in.
WhyNobody pays full price, on every request, to re-read the same 2,400 words.
- 5
A ceiling on every task, a budget on every team.
Time and spend per task have hard limits, enforced at the gateway. Hitting one stops the task and hands it to a person, because an alert afterwards would be too late. Each team and each AI agent also gets a budget in dollars, set by its owner. The gateway checks it on every call, before the model runs, along with the rate limits and the list of approved models. Near the line, the owner hears first. At the line, the work waits for their yes, or moves to a cheaper approved model where they’ve said it may.
- Every call takes one more hop, and we measure what it adds on your own traffic in the weeks the gateway only watches.
- Every rule watches and logs before it blocks, and we switch it on one account at a time, once the counts are trusted.
- If the gateway can’t check a call, the call waits rather than spending unchecked, and a named owner has a documented way to break the glass.
WhyOne AI agent in a loop can’t run up a month’s bill overnight, and the experiment still running from March gets noticed.
- 6
Your own model, only when the volume pays.
Your own model has a fixed cost: the GPUs, and the work of keeping it current. Below a certain volume renting is cheaper, and the benchmark says which side you’re on, with your numbers. In the illustrative example on our AI Inference Engineering page, it pays for itself at around 56,000 routine requests a day, against a large model with caching on.
- Where a job runs at volume, we fine-tune the open model that won your benchmark. We start with LoRA adapters, which cost far less than full fine-tuning.
- It ships only if it beats the simpler fix on cases it has never seen, at a lower cost per task.
- Prices fall fast, so the tests run again when a new model comes out, and the router’s choices change with them. A cheaper biggest model is still the biggest model doing work a small one does as well, so routing and caching save on top of every price cut.
WhyYou don’t rent forever what would be cheaper to own, or buy GPUs too early.
Seen in the field, not our work. Coinbase fine-tuned a small open model, Qwen3.5-9B, on its own data to prevent fraud. Its CEO, on how it did:
…it out performed frontier models, at much lower cost and speed…
One day to a plan, two weeks to the meter.
About 90 days to the new price.
We start with a one-day workshop, How to cut your AI bill without losing quality, with your finance lead, an engineer who knows your AI features and whoever runs your AWS account, in one room with the engineers who’d do the work. Bring every AI invoice you can find, the card statements with AI subscriptions on them, last month’s cloud bill and your usage exports, and 50 typical requests with personal data taken out.
- In the morning we put a number on today. Your finance lead brings the invoices and card statements, your engineers bring the usage exports, and by lunch every AI invoice, every AI subscription on a company card and every cloud line is one figure, with any AWS commitment going unspent beside it, and a first split by key, AI agent and team as far as today’s data allows. The exports also show how many calls each request makes, how long the context runs and which model takes each one.
- In the afternoon we run your 50 requests on several models side by side, quality beside cost per task. Then we design the target with finance in the room: what a task and an outcome are for each AI agent, the labels, the budgets and who owns each, the guardrails and the order the teams move onto the gateway; and, from what the 50 requests show, which model takes each job, what we cache, what becomes plain code and how the router decides.
- You leave with a findings report with your baseline, a target architecture (the Cost Control Tower’s plan for your accounts, and the router’s) and a costed roadmap: what each job costs today, what it should cost, and an estimate from your own 50 requests of what you’d save at the same quality. It’s yours to act on with us or without us.
AWS funding typically covers the workshop. AWS has funded part of the work after it for several of our clients, and we’ll tell you whether yours could qualify.
After it, the meter comes first: about a two-week build in your account, once the definitions exist, built with your engineers. Budgets switch on team by team, watching first, and from the next month-end the AI line arrives already split by team. Your team runs it after, with the runbooks, or AK Run, our managed service, does, where the same gateway shows cost by team every day.
Beside it, a benchmark on your own requests typically takes two to three weeks and turns that estimate into the numbers on a scorecard you sign. In this page’s example the bill falls by about three quarters, because most of that business’s requests are routine and the hard ones stay where they are. Yours could be more or less. There’s a go-or-stop decision in week 4, on that scorecard, and about 90 days to go live. If it’s stop, you keep what we found: the baseline, the benchmark’s results and the scorecard.
What we’ll tell you, even when it’s not what you hoped.
- Keep the big model where it earns its price. Sometimes the cost is the price of a product that works, and then we say so, and you leave with a forecast you can budget for instead of a cut.
- Don’t start with a dashboard. On a shared key it only shows what’s already spent.
- Don’t switch every budget on at once. Watch first, then enforce one team at a time.
- Don’t move providers to save money before the work is routed. A badly routed workload costs more anywhere.
- Don’t buy GPUs yet if renting is cheaper at your volume. We’ll show you the sums.
- The gateway won’t see the AI tools your people use in the browser. Those subscriptions need a different control, and we won’t pretend otherwise. The workshop still counts them in the one number.
For example, your support replies: cost per 1,000At week 13
less a reply,
at week 13
Example
Costs we’ve already taken apart.
Most of them weren’t AI bills.
Only Ryno’s was about an AI bill. The others were two cloud bills, a shipping bill and a board’s own numbers, and the work started the same way each time: put every cost on a name, then pay only for what it needs.
Ryno Strategic Solutions
Marketing for home services
Under 2 weeks
More than 150 people analyzed tens of thousands of inbound calls by hand, every day. Our AI call analysis was production-ready in under two weeks, inside a four-week proof of concept. We benchmarked and fine-tuned Amazon Nova models, and wrote a total-cost analysis with a go or no-go framework: fine-tuning opens a path to a lower inference cost per call at the same accuracy. AWS funded $25,000 of the work.
Pixis
An AI marketing platform
25% less
Its infrastructure cost kept rising, with nothing showing which part of the platform it came from, and nine AWS accounts blurring billing. We made cost visible by workload with AWS’s cost dashboards, right-sized its workloads and cut the nine accounts to four. Cloud spend fell 25%. Compute cost fell 48%, and uptime held at 99.996% through the work.
DigitalStrike
A marketing platform for car dealerships
20% less
Hundreds of dealerships share one platform. Lifecycle policies, auto scaling and right-sized compute brought monthly AWS spend down 20%. Uptime went from 95% to 99.9%.
Hyundai Glovis
Automotive logistics, North America
15% less
Its shipping costs were driven by weekly retrospective reports, and dealers phoned operations to chase delays. We built its first dealer-facing real-time vehicle tracking portal, with automatic exception alerts. Shipping costs fell 15%, and on-time vehicle deliveries rose 45%.
An executive search firm
C-suite and board-level search
Board answers
Its vendor’s platform shut down in December 2025. Every workload moved onto AWS systems the firm owns, with zero disruption, and now the CEO and board answer their own questions on revenue, margin and pipeline. All business data stays inside AWS. AWS fully funded the discovery.
Checked by AWS. An AWS Premier Tier Services Partner with more than 500 customer launches on AWS. 1,300+ engineers, and a Databricks consulting partner.
What owners and finance leads say to put it off.
And what we tell them. Got another? Bring it to the workshop.
Our AI spend is still small.
Then it’s the cheapest time to fix it. Each new AI agent tends to copy the last one’s defaults: the biggest model, and the full instructions on every call. Set them right while the bill is small, and every AI agent after it starts right.
It’s also the cheapest time to put a unit on it. The plumbing is the same at ten thousand users and at a million. What changes is how much of it you redo under pressure, and how many months of the bill nobody can split by then.
The workshop takes a day, and you leave knowing what each job costs.
There’s no budget for it this year.
AWS funding typically covers the workshop, and AWS has funded part of the work itself for several of our clients. At the workshop we’ll tell you whether yours could qualify.
The money may already be spent, too. If your AI runs on another provider’s account while an AWS commitment goes unused, you may be paying twice: once to that provider, and again for commitment you don’t use. Models on Amazon Bedrock bill by the token, on the AWS bill you already have. The workshop checks with AWS whether that spend would count toward your commitment. If it would, we route the work before anything moves.
And the meter finds the AI agents nobody owns, like the two still running from March.
Finance owns that.
Then finance is in the room, which is how the workshop runs. Engineering can’t define cost per outcome alone, and finance can’t meter a model call. Your engineers could build the plumbing, and with the accelerator already built it’s about two weeks. The day is for the definitions.
We already do FinOps.
That usually means cloud FinOps, which doesn’t touch per-token, per-team, per-use-case attribution of model calls. The cloud bill shows an AI line, and your model provider’s dashboard shows usage per key, after the fact. Neither can say which team’s AI agent made which call, or what a resolved ticket is, unless the call carried that name when it was made. The gateway puts it there, and your FinOps tools read it.
Our engineers can switch to a cheaper model themselves.
They can. It’s one line of code. The work is knowing it still gets your answers right, on your own cases, every time a model changes.
We build the tests, the router and the traces with your engineers, inside your business, so they run it afterwards and try every new model the same way.
We’re happy with the model we use now.
Keep it for the hard work. The router sends routine requests to smaller models and the rest to the model you trust, so you pay top prices only where the work earns them.
A job moves to a smaller model only if it passed on your own cases, against the bar you set. When the smaller model isn’t sure, the request goes up to yours.
Our data can’t leave our cloud.
It doesn’t have to. The gateway, the router, the traces and the tests run in your own AWS account or Databricks workspace, and so does your own model, where the volume pays for one.
Requests go to a hosted model only where you choose. Through Amazon Bedrock the model’s maker never sees them, though on some top models AWS keeps them up to 30 days for review.
A client of ours, a Chicago-area media archive and licensing company, put it in its technical leadership’s own words: “Armakuni helped us build agentic AI capabilities that work inside our content pipeline. The orchestration layer sits in our AWS account, governed by our IAM, audited by our team. We own every piece of it.”
Our AI runs on another provider, not on AWS.
The meter sits in front of Amazon Bedrock, in your account. Moving your model calls there takes about two weeks, because the code your engineers wrote for the model makers’ own SDKs keeps working. The work is the platform around the calls. Until then, the workshop still puts every invoice in one number.
Bring every AI invoice you can find.
We start with a one-day workshop with the engineers who’d do the work, and your finance lead in the room. AWS funding typically covers the workshop. You leave with your baseline, a findings report, a target architecture and a costed roadmap: what each job costs per task, team and outcome, which model should do it, what it should cost and an estimate of what you’d save at the same quality. If your AI bill turns out to be the price of a product that works, we’ll say so, and you leave with a forecast instead of a cut.