Skip to main content

Amazon Bedrock

Run Coeffection's AI in your own AWS account. You create one IAM role in AWS. Coeffection uses it to call Amazon Bedrock for your workspace. This guide covers setup, what the role can do, and what to do when something goes wrong.

Who this is for#

Amazon Bedrock is part of HIPAA mode. HIPAA mode is turned on by Coeffection on request. Contact us to discuss your requirements. Once it is on, every AI feature in the workspace runs in your own cloud account: Amazon Bedrock in your AWS account, Google Vertex AI in your Google Cloud project, or both. Anthropic, OpenAI and the Gemini API are turned off. To use Google Vertex AI as well, see the Google Vertex AI setup guide.

If the AI Assistant page has no Amazon Bedrock card, HIPAA mode is not on for your workspace. Contact Coeffection support.

โ„น
What this promises. AI runs only in your AWS or Google Cloud account. That is all. Your records and chat history stay in Coeffection's database, and some items are still open. Read HIPAA mode: what runs where before you decide what to store in the workspace.

Before you start#

  • HIPAA mode is on for your workspace.
  • An AWS account you control. AI requests run in that account, so your own agreement with AWS (the AWS BAA, which you accept in AWS Artifact) is the one that applies to them. AWS decides which services that agreement covers. Check AWS's HIPAA Eligible Services Reference for Amazon Bedrock and for the models you use. Coeffection can't do this for you.
  • An AWS user who can create CloudFormation stacks and IAM roles, and who can turn on model access in Amazon Bedrock.
  • A Coeffection role with the permission to manage Amazon Bedrock (integrations_bedrock:manage). The Owner and Admin roles have it. You can give it to another role on the Roles page. Without it you can't change the connection or see the External ID.

Set it up#

Go to Admin Settings โ†’ Integrations โ†’ AI Assistant and open the Settings tab. While HIPAA mode is on, a banner reads โ€œHIPAA mode: AI runs only in your AWS or Google Cloud account.โ€ A setup checklist follows it. The Amazon Bedrock card below the checklist holds the four steps: Region, External ID, Role ARN and Test. The Google Vertex AI card sits below it.

Step 1 of 4: Choose an AWS region

Choose the AWS region where Amazon Bedrock runs for your workspace. Every option is in the United States.

AWS regions you can choose
RegionCode
US East (N. Virginia)us-east-1
US West (Oregon)us-west-2
US East (Ohio)us-east-2

Select Continue. Coeffection makes your External ID. If you come back to this step and select Continue again, you keep the same External ID.

Not every model is offered in every region. Check that the model you want is available in the region you choose.

Step 2 of 4: Create the access role in AWS

This step shows your External ID in a read-only box with a Copy button. Below it is the Open AWS CloudFormation button. It opens AWS in a new tab with your External ID filled in.

  1. Sign in to the AWS account you want to use, with a user that can create IAM roles.
  2. Select Open AWS CloudFormation. The Create stack page opens. The stack is named CoeffectionBedrockAccess. Two fields are filled in: ExternalId, whose value AWS hides, and BrokerPrincipalArn. Check that BrokerPrincipalArn is arn:aws:iam::616415435588:user/coeffection-bedrock-broker. Leave it as it is unless Coeffection support tells you otherwise.
  3. Tick the box that says CloudFormation might create IAM resources with custom names. The role has a fixed name, so AWS asks. A link can't tick it for you.
  4. Select Create stack. Wait until the stack shows CREATE_COMPLETE.
  5. Go back to Coeffection and select Continue.

If your browser blocks the Copy button, the External ID is selected so you can copy it with Ctrl+C (Cmd+C on a Mac).

โš 
If the Open AWS CloudFormation button is greyed out and the note says โ€œSetup link coming soonโ€, the quick-create link is not published yet. Contact Coeffection support. Support can send you the template file. In AWS, open CloudFormation, choose Create stack, upload the template file, and paste your External ID into the ExternalId field.

Step 3 of 4: Paste the role ARN

When the stack has finished creating, copy the ARN of the CoeffectionBedrockAccess role from AWS and paste it into the Role ARN field.

  • In AWS, open the stack, choose the Outputs tab, and copy the value of RoleArn.
  • It looks like arn:aws:iam::123456789012:role/CoeffectionBedrockAccess. It must be in the standard aws partition, and the role name has no path.
  • Select Save role ARN. Coeffection keeps only the last four digits of your AWS account number on screen.
  • If a role is already saved, the step says so and the button reads Continue. Paste a different ARN to replace the saved one.

Step 4 of 4: Test the connection

Select Test connection. Coeffection assumes the role and sends one small test request to Amazon Bedrock. It takes a few seconds.

  • If the test passes, it says โ€œAmazon Bedrock is connected.โ€ The card then shows Connected, the last four digits of your AWS account, and when it was last tested. Use Re-test whenever you want to check again.
  • If the test fails, the card shows why and offers Re-test, Edit role and Disconnect. Edit role opens the setup at step 3. See Troubleshooting.
  • If a model is missing, the test adds โ€œNot enabled in your AWS account:โ€ and the model names. See Turn on model access in AWS.

What the access role can do#

The stack creates one IAM role named CoeffectionBedrockAccess. It creates nothing else: no user, no access key and no policy for you to manage. Its one output is RoleArn.

Who can use it

  • Only Coeffection's broker user can assume the role: arn:aws:iam::616415435588:user/coeffection-bedrock-broker.
  • It can do that only with your External ID, and only with sts:AssumeRole.
  • The broker user has one permission, which is to assume roles named CoeffectionBedrockAccess. It has no console access.

What it can do

  • Call bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on Anthropic and OpenAI models in the three regions above. These two actions also cover the Converse API.
  • Call the same two actions on US inference profiles (us.*) in your own account, in those regions.
  • Do that only when the request is made in one of the three regions.
  • Call bedrock:ListFoundationModels and bedrock:ListInferenceProfiles, so Coeffection can show which models your account has.

What it cannot do

  • Use any IAM, S3 or STS action. Your data in S3 and other services is out of reach.
  • Create, change or delete anything in Bedrock (no bedrock:Create*, Delete* or Put*). It cannot start Bedrock agents.
  • Use global inference profiles (global.*). They can run outside the United States. Coeffection uses US profiles only.
  • Run other model families, such as image models. The role covers Anthropic and OpenAI models only.
  • Search the web. A web search permission is added only after Coeffection verifies web search on Bedrock. It comes as a new template version that you apply to your stack.

Limits Coeffection adds

Each time Coeffection assumes the role, it also passes a session policy. The policy allows only bedrock:InvokeModel* and bedrock:List*. Even if you widen the role later, Coeffection's sessions can't do more than that. A session lasts at most 60 minutes.

The External ID

The External ID starts with cx-. Coeffection makes it. You can't choose it. It is not a password. It is in the role's trust policy, so people in your AWS account who can read the role can see it. Its job is to stop another Coeffection workspace from using your role.

To end Coeffection's access at any time, delete the CoeffectionBedrockAccess stack in AWS. That deletes the role.

Turn on model access in AWS#

A model has to be enabled in your AWS account, in the region you chose, before Coeffection can use it. The access role can't enable models for you.

On Amazon Bedrock, the default model for Grace, the daily brief, the field service brief and Clay contact discovery is Anthropic Claude Haiku 5.5. Coeffection reaches it through a US inference profile.

  1. In the AWS console, switch to the region you chose in step 1.
  2. Open Amazon Bedrock and find model access. AWS may call this the model catalog.
  3. Enable Anthropic Claude Haiku 5.5. If AWS asks you to accept terms or describe your use, do that first.
  4. Go back to Coeffection and select Test connection or Re-test.

You only need Claude for now. Lead enrichment and web research are off on Amazon Bedrock until web search on Bedrock is verified, so you don't need to enable GPT models yet.

โ„น
AWS changes its console from time to time. If a menu name differs, search the AWS documentation for โ€œAmazon Bedrock model accessโ€.

Choose models for each task#

Below the connection cards, the AI Assistant page lists each AI task with the model it uses. Each task shows a Default or Chosen badge.

  • From Amazon Bedrock, the list shows only text models that are enabled in your account and region. If Google Vertex AI is connected too, its Gemini models are listed under their own group. See the Google Vertex AI setup guide.
  • Pick another model for a task, or select Reset to defaults.
  • For now, Grace and the two briefs use only Claude models from Amazon Bedrock.
  • If the list is empty, it says โ€œNo text models enabled in this region. Enable model access in the AWS console.โ€ See Turn on model access in AWS.
  • If a model you picked is retired, the task runs its default model until you pick a new one.
  • A task that is off says โ€œOff in HIPAA mode until AWS web search is verified.โ€

Troubleshooting#

When a test fails, the card shows the reason. AI features show the same words when a request fails. Find the message below.

Messages you may see, what they mean and what to do
MessageWhat it meansWhat to do
โ€œWe can't reach your AWS. Check the Bedrock connection.โ€Coeffection could not use your role. The role may be missing, the External ID may not match, or AWS may have refused the request. Coeffection does not say which. That stops the message from being used to probe other AWS accounts.Check that the stack shows CREATE_COMPLETE and the role exists. Check that the role ARN in Coeffection is the RoleArn output of the stack. If you rotated the External ID, update the stack. Check that no AWS Organizations policy blocks sts:AssumeRole from Coeffection's account. Then select Re-test.
โ€œThis AI model isn't enabled in your AWS account. Enable it in the AWS console, then re-test in Admin > Integrations > AI.โ€AWS refused the request because the model is not enabled in your account for this region. The message may name the model.Enable the model in the region you chose, then select Re-test. A message like this never marks the model as retired.
โ€œThis AI model was retired. Ask your admin to pick a new one.โ€AWS has retired the model.Pick another model for the task on the AI Assistant page.
โ€œAI is busy, try again. New AWS accounts start with low Bedrock quotas.โ€AWS slowed the request down, or was briefly unavailable, before the answer began. Coeffection already tried up to three times.Wait a minute and try again. New AWS accounts start with low default Bedrock quotas. If it keeps happening, ask AWS for a higher quota in Service Quotas.
โ€œThe answer was cut off, try again.โ€The answer started and then stopped, for example because AWS slowed the request in the middle of it.Ask again.
โ€œAI is temporarily unavailable.โ€The problem is on Coeffection's side, not yours. Coeffection is alerted.Try again in a few minutes. If it continues, contact Coeffection support.
โ€œConnect Amazon Bedrock to use AIโ€HIPAA mode is on, but there is no working Amazon Bedrock connection.Finish the four setup steps, or connect Google Vertex AI instead. AI features stay off until one cloud is connected.
โ€œConnect Amazon Bedrock or Google Vertex AI to use AIโ€HIPAA mode is on, but neither Amazon Bedrock nor Google Vertex AI is connected.Finish the four setup steps here, or connect Google Vertex AI with its setup guide. AI features stay off until one cloud works.
โ€œWeb research is off in HIPAA modeโ€This is by design. Live web search stays off until web search on Bedrock is verified. Lead research can run without live web search if Google Vertex AI is connected.Nothing to fix.
โ€œThis workspace is in HIPAA mode, so AI runs only through Amazon Bedrock or Google Vertex AI, in your AWS or Google Cloud account.โ€Something tried to use a direct provider. Anthropic, OpenAI and the Gemini API are turned off while HIPAA mode is on.Nothing to fix. Use Amazon Bedrock or Google Vertex AI.

While you set up, you may also see these.

Messages during setup, what they mean and what to do
MessageWhat it meansWhat to do
โ€œEnter the ARN of the CoeffectionBedrockAccess role in your AWS account.โ€The text you pasted is not the ARN of that role.Copy the RoleArn output of the stack. It ends in :role/CoeffectionBedrockAccess, with no path.
โ€œStart the Amazon Bedrock setup first.โ€The setup was cleared, for example by a Disconnect in another tab.Go back to step 1 and select Continue.
โ€œToo many connection tests. Try again in a few minutes.โ€Test connection is limited for each workspace.Wait a few minutes, then test again.

Rotate the External ID#

Rotate the External ID if it was shared by mistake, or if your security policy asks for it. Your role still trusts the old External ID, so Amazon Bedrock stays unavailable until you update your stack and the test passes again. Coeffection keeps your role ARN and region.

  1. On the connection card, select Rotate external ID, then Rotate now. The setup opens at step 2 with the new External ID.
  2. Select Copy to copy the new External ID.
  3. In AWS, open CloudFormation and select the CoeffectionBedrockAccess stack. Choose Update stack and keep the existing template. Paste the new External ID into the ExternalId field and finish the update.
  4. When the stack shows UPDATE_COMPLETE, go back to Coeffection. Select Continue on step 2 and on step 3, then select Test connection on step 4.
โš 
Do not use the Open AWS CloudFormation button again for a rotation. It tries to create a new stack, and a stack with that name already exists. Update the stack you have. AWS hides the External ID in the stack's parameters, so the test is how you check that it matches.

Disconnect#

On the connection card, select Disconnect, then Yes, disconnect. The dialog says what happens to AI next. With no other cloud connected it says โ€œGrace and AI features stop until reconnected.โ€ With Google Vertex AI connected it says โ€œAI tasks that use Amazon Bedrock move to Google Vertex AI where it has a model for them. Tasks it has no model for stay off until you reconnect.โ€

  • Disconnect erases the saved role ARN, the External ID and the connection details in Coeffection. It does not change your AWS account.
  • To end Coeffection's access in AWS too, delete the CoeffectionBedrockAccess stack. That deletes the role.
  • While HIPAA mode is on and Amazon Bedrock is disconnected, a task that Google Vertex AI can run uses Google Vertex AI, if it is connected. If no cloud is connected, AI features stay off. Coeffection does not switch to Anthropic, OpenAI or the Gemini API.
  • To connect again, start from step 1. You get a new External ID.

Cost and usage#

You pay AWS directly for Amazon Bedrock. Coeffection does not bill you for it. Coeffection still counts tokens, so your usage caps keep working, and the AI Assistant page shows an estimated cost. Your AWS bill is the number that counts.