---
title: Building private AI in-house vs. a done-for-you service
description: Should you build private AI in house or buy it? See what Azure OpenAI and Amazon Bedrock handle, and the people and upkeep you'd supply yourself on each path.
canonical: https://privatesuperintelligence.si/blog/build-vs-buy-private-ai/
last-updated: 2026-10-15
---

# Building private AI in-house vs. a done-for-you service

You can build private AI in house on Azure OpenAI (now Foundry Models sold by Azure) or Amazon Bedrock, and both give you strong commitments: your prompts are not used to train the models, and the model providers cannot see them. The cloud provider secures the model and its infrastructure. Everything around it is yours to design, configure and run: identity, logging, retention settings, connections to your email and documents, the application your staff use, and the upkeep as models and terms change. Buying a done-for-you service moves that work to a team that does it for you, and the choice comes down to whether you have the people to own it.

Every statement about Microsoft and AWS below comes from their own documentation, read on 2026-10-15.

## What Azure OpenAI and Amazon Bedrock give you

Both clouds make clear commitments on the model layer, and they are a sound base to build on.

**Microsoft (Foundry Models sold by Azure, including Azure OpenAI).** Microsoft states that your prompts, completions, embeddings and training data "are NOT available to other customers," "are NOT available to OpenAI or other providers of Models sold by Azure," and are not used to train generative AI foundation models without your permission. Microsoft hosts the models in its Azure environment, and they "do NOT interact with any services operated by providers," such as the OpenAI API. Microsoft also states: "The models are stateless: no prompts or completions are stored in the model."

**Amazon Bedrock.** AWS runs each model in a Model Deployment Account that the Bedrock service team owns and operates. "Model providers don't have any access to those accounts," so they "don't have access to Amazon Bedrock logs or to customer prompts and completions." Bedrock also lets you set data retention yourself: in `none` mode, "no request or response data is written to durable storage by AWS or shared with the model provider."

## What you take on when you build

The cloud documents also say where their responsibility ends. AWS applies its shared responsibility model to Bedrock: "You are responsible for maintaining control over your content that is hosted on this infrastructure. You are also responsible for the security configuration and management tasks for the AWS services that you use."

Here is the work the two providers' docs hand to you.

| Area | Amazon Bedrock (from AWS docs) | Azure OpenAI (from Microsoft docs) |
| --- | --- | --- |
| Identity and access | Set up users with IAM Identity Center or IAM, least privilege, MFA on each account | Not covered on the data privacy page; you set it up in your own Azure tenant |
| Activity logging | AWS recommends CloudTrail for API and user activity | Not covered on the data privacy page; you set it up in your own Azure tenant |
| Retention | Set `data_retention_mode` per Region, at account or project level; enforce with IAM or Service Control Policies | Apply for modified abuse monitoring, then verify that `ContentLogging` reads false |
| Human review | Some models require `aws_review`, with prompts and completions kept within AWS for up to 30 days | Without modified abuse monitoring, flagged samples may go to authorized Microsoft reviewers |
| Stored data | Responses API `store` behavior depends on mode | Files, vector stores, Responses API and Assistants threads stay in your tenant until you delete them |
| Data location | Cross-region inference stores retained data in the destination Region | Global and DataZone deployments may process prompts outside the resource's own geography |
| Encryption and network | AWS encryption solutions, Amazon VPC and AWS PrivateLink | Stored data encrypted with AES-256 by default, with the option of a customer-managed key |

A few details from the docs show how much configuration is involved:

- **Retention is per Region.** AWS says the Bedrock retention setting "does not propagate to other Regions," and a Region you have not configured falls back to each model's default.
- **The newest models may need different settings.** AWS lists Claude Fable 5 and Claude Fable 5.1 as models that require `aws_review`. If your account is set to `none`, those models appear as unavailable.
- **The setting lives in an API.** The Bedrock retention guide notes that at launch there is no console screen for it, so you configure it through the API or SDK.
- **Free-text fields leak.** AWS recommends that you "never put confidential or sensitive information, such as your customers' email addresses, into tags or free-form text fields," because those may appear in billing or diagnostic logs.
- **Previews differ.** Microsoft notes that Azure preview features "may employ different privacy practices."

All of this covers the model endpoint. The [zero data retention guide](/blog/zero-data-retention-ai/) walks through the parts that sit outside it: application logs, vector stores, backups and connected apps.

## The roles an in-house build needs

The model is one component. A private AI your staff will use daily also needs an application, connections to your data, and people to keep it running. Expect to fill these roles, whether with employees or contractors:

- **Cloud and security engineering:** accounts, identity, network isolation, retention policies, logging, key management, and the SCPs or policies that enforce them.
- **Application engineering:** the chat or assistant interface, connectors to email, calendar and document stores, retrieval over your files, and citations back to sources.
- **Permissions design:** who can see which mailbox, which entity's documents, and which actions the AI may take. Permissions in a family office follow people, entities and the principal's wishes, not a generic org chart.
- **Operations:** monitoring, incident response, model upgrades, and reviewing each provider's terms when they change. Microsoft's data privacy page lists five revisions between June 2023 and October 2025.
- **Evaluation:** checking that answers are correct and cited before anyone acts on them.

Citi Institute's May 2026 report found that 57% of family offices cite lack of internal expertise as the biggest barrier to AI adoption. That number is the build-vs-buy question in one line. Citi also notes that single family offices "tend to operate with lean teams."

## Build or buy: how to decide

| Question | Points to build | Points to buy |
| --- | --- | --- |
| Do you have cloud security engineers on staff today? | Yes, with time to spare | No, or they are fully booked |
| Who owns the system when the person who built it leaves? | A team, with documentation | Unclear |
| Do you need the AI across email and documents, with permissions? | You have engineers to build connectors | You want it working in weeks |
| Do you want to choose the model and where data lives? | Either path can give you this | Either path can give you this |
| Who reviews provider terms when they change? | Your counsel and engineers | The service, with you signing off |

Building makes sense for a firm with an engineering team that already runs production cloud systems and wants to own every layer. Buying makes sense when the expertise isn't in house and you want the result rather than the project. Ask any vendor the same questions you would ask your own team: which provider, which region, which retention mode, who can see what, and how you verify it. The [private LLM vs private AI service guide](/blog/private-llm-vs-private-ai-service/) covers the difference between hosting a model and running a service.

## A done-for-you private AI

Private SuperIntelligence from Mitosis Labs is the most capable AI, trusted with everything, delivered as a service. Each deployment is built for one client, from the hardware to the choice of model provider to where the data is stored, and Mitosis Labs works with leading vendors and providers to fit each one. It works across your own email, documents and data, answers with citations, and acts within permissions you set, and delivery takes weeks. Read more on the [family offices page](/family-offices/).

## Frequently asked questions

### Is Azure OpenAI private?

Microsoft states that prompts and completions are not available to other customers or to OpenAI, and are not used to train foundation models without your permission. Abuse monitoring may still send flagged samples to authorized Microsoft reviewers unless you are approved for modified abuse monitoring.

### Does Amazon Bedrock share my prompts with model providers?

No. AWS says model providers have no access to the Model Deployment Accounts, so they cannot see Bedrock logs, prompts or completions.

### Can I get zero data retention on Bedrock?

Yes, by setting `data_retention_mode` to `none` for each Region, which you can enforce with IAM or Service Control Policies. Models that require retention then become unavailable unless AWS approves ZDR for your account and that model.

### What does it take to build a private ChatGPT instance?

A cloud model endpoint plus identity, logging, retention settings, an application, connectors to your data, and people to run it. The cloud provider handles the model and infrastructure; the rest is your responsibility under its shared responsibility model.

### Is building private AI cheaper than buying?

It depends on the engineers you already employ and the time they can give it. Count the roles above, not only the cloud bill.


## Related reading

- [Private AI for family offices](/blog/private-ai-for-family-offices/)
- [Consolidated reporting across custodians, without the Excel export](/blog/family-office-consolidated-reporting/)
- [Tracking K-1s across 30 entities](/blog/k-1-tracking-across-entities/)
- [Capital call notices: catching every one, across every trust and LLC](/blog/capital-call-tracking/)
- [The family office technology stack](/blog/family-office-technology-stack/)
- [Family office operations, and where AI changes the staff math](/blog/family-office-operations-and-ai/)
- [What is private superintelligence?](/blog/what-is-private-superintelligence/)

## Sources

- [Data, privacy, and security for Foundry Models sold by Azure (Microsoft Learn, updated 2026-06-05, accessed 2026-10-15)](https://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/openai/data-privacy)
- [Data protection (Amazon Bedrock User Guide, AWS, accessed 2026-10-15)](https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html)
- [Data retention (Amazon Bedrock User Guide, AWS, accessed 2026-10-15)](https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html)
- [Shared responsibility model (AWS, accessed 2026-10-15)](https://aws.amazon.com/compliance/shared-responsibility-model/)
- [AI in the Family Office (Citi Institute, May 2026)](https://www.citigroup.com/rcs/citigpa/storage/public/ai_in_the_family_office.pdf)