Contact us
If your AI agents call a model maker’s own API, we move their calls to the same models on Amazon Bedrock, where Bedrock offers them, so company data stops going to a model maker’s servers on every call. Your code keeps working, so repointing the calls takes about two weeks, and each AI agent then works under its own identity, on the AWS bill you already have.
The workshop is one day, 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, validated by AWS for agentic AI consulting.
Your AWS account is the keep. The same models now stand inside its walls, the drawbridge stays up, and every AI agent carries the name of the person who sent it.
In most growing businesses the first AI agent was wired straight to a model maker’s own API, because that’s where the demo ran. Everything else the business runs sits in an AWS account that is locked, logged and checked. Every call an AI agent makes leaves it. A customer’s email, a contract and last month’s numbers go to another company’s servers, under one key that every AI agent shares and nobody owns. The bill goes on a card, beside any AWS commitment the business has already signed.
Last month’s AI callsToday
calls went to a model maker’s own servers
AWS commitment$11,600 of $40,000 unspent
Example
Six examples, from six kinds of business. Each AI agent is already at work on a model maker’s own API, and a person manages it and now gets through far more. The workshop tells you which kind of move each one needs, before anything moves.
A recruitment agency
CVs read, shortlists drafted
Repoint
The code moves as it is. The CVs stay in the AWS account where the applicant records already live, and go to the model only through Bedrock.
Amara reviews each shortlist and calls the best eight the same afternoon.
An accountancy firm
Bank statements matched to the books
Build around it
It ran under the firm’s one key. Now it acts as the accountant who asked, so it opens only her clients’ files.
The partner signs each month’s books with the five flagged lines in front of her.
A software company
Answers inside its own product
Rebuild part of it
It was built on a model maker’s hosted assistant, with its files kept there. The files move into the company’s own account, and we rebuild the conversation on AgentCore.
Its CTO answers the enterprise buyer’s question in one line, and the deal moves on.
A clinic group
Patient messages sorted for the front desk
Repoint
The same code, now behind guardrails that mask the identifiers you choose, so it never repeats them in a reply.
The front desk confirms the day’s changes by 8 and rings back anyone the AI agent flagged.
A property manager
Leases read, renewal terms pulled
Repoint
The leases stay in the account they’re filed in, and each one goes to the model through Bedrock, where the model’s maker never sees it.
The portfolio manager opens each renewal with its dates and terms already pulled.
An online shop
Customer emails answered from the order history
Build around it
A refund over $100 now waits for a person. That rule sits at the gateway, where security keeps it, rather than in the AI agent’s prompt.
Theo in support clears the refunds that wait for him, with the order and the email in front of him.
Code written for the model makers’ own SDKs keeps working against Amazon Bedrock. Repointing a call is mostly a new address and new credentials, about two weeks for the lot. The six decisions below are what make the move worth doing. Without them, the old failures follow your AI agents inside. Our engineers make them with your engineering, security and finance leads, the way AI Inference Engineering builds every system.
One reply, order 6182An online shop
times it reached a model maker’s own servers
97 of 100 right after the move
Example
We read the code and the usage exports for every model call: which model, which SDK and endpoint, which key, which AI agent, and what kind of data it sends. We find the keys wherever they are, in code, in config files, on laptops, or in the no-code tool someone connected last spring.
Spend by provider goes beside your AWS commitment, so the business case is made in your own numbers.
WhyA call nobody knew about would still leave after the move.
Each call moves to Amazon Bedrock, made from your own account over a private connection (a VPC endpoint), so it never crosses the public internet. The code keeps its SDK. What changes is the address and the credentials, which now come from the AI agent’s own AWS identity instead of a key in a file.
Through Bedrock, the model’s maker never sees your prompts. On some top models AWS keeps them up to 30 days for review, so we choose those workload by workload with you.
WhyCompany data stops going to a model maker’s servers on every call.
Every AI agent has its own identity in AgentCore Identity. Acting for one of your people, it gets that person’s rights and no more. Answering a customer, it reaches only that customer’s records. No credential ever enters the model’s context: each tool call gets a short-lived token scoped to that call, and each session runs apart from every other.
What it may touch is a Cedar policy at AgentCore Gateway, deny by default and checked on every tool call. Your security team keeps it outside the AI agent’s code, so tightening a rule needs no new release. AWS Service Control Policies allow only the models you approved and refuse any call that skips the gateway. Network egress is deny by default as well, and Amazon Bedrock Guardrails screen what goes in and out.
WhyNo AI agent acts under a key nobody owns, and one that is tricked still can’t reach a tool it was never meant to use.
We check that each model you use is offered in your Region. Some are served only through a cross-Region inference profile, which keeps requests inside a geography, such as the US or the EU, rather than one Region. Where your data must stay in one Region, we pin it there and choose from the models offered in that Region, and the workshop says which AI agents that affects.
Bedrock’s quotas are per account, per model and per Region. We size them against your busiest hour and ask AWS for more before the switch, not after the first throttled call. The gateway has a fallback for the day a model or a Region has a bad hour.
WhyThe move keeps your residency rules, and it holds up in the busiest hour of the month.
Your own prompts run on the Bedrock model beside today’s provider, scored with Bedrock Evaluations on real cases your people signed off. Each AI agent switches only when it matches or beats today’s answers: in shadow first, then on a share of real traffic, with the old route kept for rollback until its scores hold. After the switch, the same tests run whenever a model changes.
Where the model you use isn’t on Bedrock yet, or reaches it later than its maker’s own API, we say which AI agents that affects. They stay outside on purpose, listed, until it does.
WhyNobody finds out from a customer that the answers got worse.
Once the last AI agent has switched, we revoke the shared keys and close the route to the model makers’ endpoints, so nothing can quietly go back. Spend lands by the token on the AWS bill you already have, beside the rest of your cloud, with no seat licenses.
Whether that spend counts toward your AWS commitment depends on your agreement, and the workshop checks it with AWS before anyone signs. Cost by team and by outcome is its own piece of work, and we'll say if you need it yet.
Why This is what stops a second AI bill landing on a card nobody owns, with the old keys already revoked so none of them still work.
From a client: technical leadership at a media archive and licensing company near Chicago, on the AI agents we built with them.
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.
We start with a one-day workshop, How to run your AI inside your own AWS account, with your engineering, security and finance leads and the engineers who’d do the work. Bring the usage exports from your model providers, the last few AI invoices, and a list of where your keys live, if you have one.
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, repointing the calls takes about two weeks. Each one runs in shadow beside the old route until its answers hold. There’s a go-or-stop decision in week 4, on a scorecard you sign. If it’s stop, you keep what we found: the map of every call and key, the tests and the scorecard.
For example, six AI agents: live calls still going the old wayAt week 10
go to a model maker’s
servers, from week 10
Example
Three clients, two on Amazon Bedrock and one whose platform we moved onto AWS, where all its business data stays inside AWS.
A government case management platform
Software for agencies and non-profits
Months to weeks
Onboarding a new client took six to seven months, configured by hand, and the timeline lost it deals. We built form and checklist onboarding with generative AI on Amazon Bedrock, with a person editing and approving each one. Generative AI now does 80 to 90% of the form onboarding effort. Onboarding went from about six months to weeks. AWS largely funded the work.
An executive search firm
C-suite and board search
Zero disruption
The vendor behind its platform shut down. We moved every workload onto AWS-native systems the firm owns and runs, and an AI agent now returns a ranked, reasoned shortlist. All business data stays inside AWS. AWS fully funded the discovery that opened the work.
A U.S. cinema chain
One of the largest cinema networks in the U.S.
30% faster
Its contact center runs AI agents for showtimes, refunds and loyalty on Amazon Bedrock, AgentCore and AgentCore Gateway, beside Amazon Connect. It takes more than 40,000 calls a day. Call handling is 30% faster across the contact center.
1,300+engineers, across the UK, US, UAE, India and Canada
500+customer launches on AWS
70%of the proofs of concept we scope reach production
AWS has checked our work. We’re an AWS Premier Tier Services Partner, validated by AWS for agentic AI consulting, with the AWS Migration and Modernization Competency.
The five we hear most. Got another? Bring it to the workshop.
Then keep them. Where a model is on Bedrock, nothing changes about the model, only where it runs, which identity calls it and which log shows it. Where one isn’t, or reaches Bedrock later than its maker’s own API, the workshop names the AI agents that affects, and they can stay where they are, on purpose.
That’s right by default. Sent directly, the big providers’ business terms say they don’t train on your data, and they keep it 30 to 55 days to check for abuse. Through Bedrock, the model’s maker never sees it, and on some top models AWS keeps it up to 30 days for review.
Training was never the whole question, though. Security also asks which key sent it, whose rights it used and which log shows it, and inside your account each of those has an answer.
They can, and the new address is the easy part. What takes the real work is the identity each AI agent gets, the rules your security team sets, the tests that check it, and the metering that tracks cost, because that's what gets the move past your security review. We build those with your engineers, inside your business, so your team runs it afterward.
The time is one day with your engineering, security and finance leads, then our engineers do the work beside yours. Moving the calls takes about two weeks. The roadmap gives the rest its weeks, so nothing lands on your team at once.
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. Bedrock bills by the token on the AWS bill you already have, and the workshop checks with AWS whether that spend would count toward your commitment.
Your code keeps speaking the SDK it speaks today, so moving a call again is the same small change. Everything we build, the identities, the rules, the tests and the dashboards, lives in your AWS account and your repositories, and we hold no license over it to keep it running.
We start with a one-day workshop, with the engineers who'd actually do the work. By the end you know where every AI call goes today, what it costs, and which AI agents move easily, which need bounded work around them, and which are a genuine rebuild. We write it up as a findings report and a target architecture, with a phased, costed roadmap you can act on. AWS funding typically covers the workshop.