> For the complete documentation index, see [llms.txt](https://doc.wearepatchworks.com/product-documentation/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://doc.wearepatchworks.com/product-documentation/ai/patchworks-ai-studio/models.md).

# Models

AI Studio isn't built on a single AI model. This page explains how that works, and what it means for you.

## Frontier models on AWS Bedrock

For tasks that need deep reasoning, natural conversation, and the ability to write scripts and documentation, AI Studio uses frontier models - currently Anthropic's Claude family - served through **AWS Bedrock**.

We chose Bedrock as our primary route to frontier models because:

* **It runs where Patchworks runs.** The Patchworks platform is built on AWS, so requests to frontier models are handled inside our AWS environment, alongside the rest of the platform, with the same security controls and per-tenant isolation.
* **It's enterprise-grade.** Bedrock is a managed AWS service with the security, compliance and availability guarantees that our customers already expect from Patchworks infrastructure.
* **Your data isn't used for training.** Requests processed through Bedrock are not used to train the underlying models.
* **We can adopt new models quickly.** As new frontier model releases become available, we evaluate them against our integration benchmarks and adopt them when they perform better - without any change to the agents you use.

{% hint style="info" %}
The specific model versions behind AI Studio change over time as new releases are evaluated and adopted. You don't need to do anything when this happens - the Assistant and Platform Agents work exactly as before.
{% endhint %}

### Always on: provider fallback

AI providers, like any cloud service, can experience outages. AI Studio is designed so that a problem with one provider doesn't stop you working.

<div data-with-frame="true"><figure><img src="https://2440044887-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FLYNcUBVQwSkOMG6KjZfz%2Fuploads%2Fpdkstkfm9VDR2U4kTlGe%2Fai-studio-models-resilience.svg?alt=media&amp;token=8b944777-10c5-40de-b9da-77357a28102d" alt=""><figcaption></figcaption></figure></div>

Every request goes to AWS Bedrock first. If Bedrock can't be reached, AI Studio automatically falls back to the **Anthropic API** directly - the same Claude models, via a different route. If that's also unavailable, it falls back again to the **OpenAI API**, an entirely independent provider. The result is that AI Studio remains operational even if two providers are down at the same time.

A few things worth knowing:

* Failover is automatic. There's nothing to configure and nothing you need to do - your conversation simply continues.
* AI Studio returns to the primary provider as soon as it's healthy again.
* During a fallback period you may notice small differences in how responses are phrased, particularly if the OpenAI fallback is in use. The agents' capabilities, and their access to your integration context, are unchanged.
* The same data-protection commitments apply to every provider we use: requests are processed only to answer them and are not used to train any provider's models.

### Patchworks-trained models using AWS SageMaker

{% hint style="warning" %}
Coming soon&#x20;
{% endhint %}

Using **Amazon SageMaker**, Patchworks fine-tunes models on integration patterns and best practice from across the platform: how connectors, process flows, mappings and transformations are structured, and what good looks like. The result is a set of specialised models that are fast, consistent and cost-efficient at the integration-specific tasks that make up much of an agent's work - such as suggesting field mappings, matching schemas between systems, and recommending flow structure.

<div data-with-frame="true"><figure><img src="https://2440044887-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FLYNcUBVQwSkOMG6KjZfz%2Fuploads%2F2Iwf2zsDtTGTLfJudrIz%2Fai-studio-models-lifecycle.svg?alt=media&amp;token=80346aee-3d6e-42be-9e28-938658dc6c8f" alt=""><figcaption></figcaption></figure></div>

These models are trained and tuned in cycles. Each cycle produces a candidate model, which is benchmarked against the current Patchworks model *and* the latest frontier releases on real integration tasks. A candidate is only promoted if it performs better - so the model you're using is always the one that has proven itself, and we're never locked into our own models when a frontier model does a better job.

{% hint style="info" %}
Patchworks models are trained on integration structures and best-practice patterns - not on your flows or business data. The records that flow through your process flows are never used as training material.
{% endhint %}

#### A hybrid approach

{% hint style="warning" %}
Coming soon&#x20;
{% endhint %}

Frontier models and Patchworks models run side by side. When an agent works on a task, AI Studio routes each step to the model best suited to it - specialised models for well-defined integration tasks, frontier models where broad reasoning and language understanding matter most.

<figure><img src="https://2440044887-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FLYNcUBVQwSkOMG6KjZfz%2Fuploads%2FbQlxQ8unE0UUzKJC0zw5%2Fai-studio-models-hybrid.svg?alt=media&amp;token=92df5378-0516-4e30-bcce-dc6b7333914d" alt=""><figcaption></figcaption></figure>

This is what the hybrid approach lets us do:

| Benefit                                 | What it means                                                                                                                                                 |
| --------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Faster responses**                    | Routine integration tasks are answered by small, focused models, so agents spend less time waiting on the parts of a job that don't need deep reasoning.      |
| **Better integration-specific results** | Models trained on how integrations are actually built produce more consistent mappings, structures and suggestions than a general model working from scratch. |
| **The best of the frontier**            | Complex, open-ended requests still get the full reasoning power of the latest frontier models - and we upgrade as better models are released.                 |
| **Cost efficiency**                     | Using the right-sized model for each task keeps AI Studio efficient, which is what allows it to be available on every plan.                                   |
| **Resilience**                          | No single model or provider is a point of failure - between provider fallback and our own models, AI Studio stays operational.                                |
| **Continuous improvement**              | Patchworks models improve with every training cycle, and frontier models improve with every release. You get both, automatically.                             |

### Your data

The same principles that apply across the Patchworks platform apply to AI Studio:

* Agents only work with the connectors, data structures and flows that belong to your company profile - the per-tenant isolation described in our [Patchworks infrastructure](/product-documentation/getting-started/technical-overview/patchworks-infrastructure.md) overview carries through to AI Studio.
* Requests sent to any AI provider are processed only to generate a response, and are not used to train that provider's models.
* Patchworks models are trained on integration patterns and structure, not on the business data that flows through your integrations.

For more on how Patchworks handles security and compliance generally, see Product security and [Data security (SOC 2)](/product-documentation/getting-started/technical-overview/data-security-soc-2.md).

### Need to know

* Model selection and provider fallback are handled automatically - there's nothing to configure.
* Token usage is tracked per company regardless of which model or provider handles a request. See [Tokens](/product-documentation/ai/patchworks-ai-studio/tokens.md) for details.