AI Features and Providers

About this page

This page explains, in plain terms, how StitchOps uses artificial intelligence (AI), which AI providers may process Customer Data, and what Customer controls. It summarizes Section 3.6 and Section 14 of the StitchOps Master Subscription Agreement at https://www.stitchops.io/legal/msa and Section 4.3 of the Data Processing Addendum at https://www.stitchops.io/legal/dpa. If this page and those agreements differ, the agreements control.

1. AI is optional

Customers choose whether to use AI, based on their own needs and requirements. Workflows can be built and run without any AI, using the visual editor, the API, or manual configuration. StitchOps is designed so that customers with strict AI policies can run it with AI turned off.

2. What counts as an AI Feature

An AI Feature is any part of StitchOps that uses an AI, large language model, or other machine learning model:

(a) AI assisted building. The Stitch assistant and AI discovery sessions, which help design workflows from a plain language description.

(b) AI action steps. Workflow steps that call a model while they run, such as text completion, image analysis, AI assisted sign in and form filling, and reading PDF documents. Each step uses a model and AI provider connection that Customer selects.

(c) Machine learning fallbacks. Machine learning vision locators and optical character recognition, used only when a workflow's standard locators or text recognition cannot find what the workflow needs on screen.

(d) AI assistant connectors. Connections that let Customer's own AI assistant, such as Claude, ChatGPT, or Copilot, search, run, and build StitchOps workflows.

3. Execution is deterministic

The StitchOps Digital Worker executes workflows deterministically: the same inputs produce the same steps. A path found with AI assisted building is saved as a workflow that replays without a model in the loop. When AI Features are off, a Digital Worker does not call any AI model while executing a workflow, and vision locators and optical character recognition run without any machine learning model.

4. Turning AI off

(a) Customer can turn off AI assisted building, AI action steps, and AI assistant connectors for its whole tenant or for individual workflows.

(b) Machine learning fallbacks are on by default. StitchOps turns them off for the whole tenant or for named workflows on written request to support@stitchops.io, confirms receipt within one Business Day, and completes and confirms the change within three Business Days. A request made in an Order Form or before the first production run takes effect from that first run.

(c) Customer may record any of these choices in an Order Form, but does not need one to make them.

5. Whose model is used

(a) Customer's own model, by default. Every AI Feature, including AI assisted building, uses a model that runs on the Digital Worker or through an AI provider Customer connects with its own account. That provider processes data under Customer's own agreement with it and is not a StitchOps subprocessor. Customer chooses the provider and the region.

(b) StitchOps Provided Inference, on request. If Customer does not have its own provider, it can ask StitchOps to supply one, elected in an Order Form with the fees stated there. Only the providers Customer elects process Customer Data to run AI Features. StitchOps currently offers:

(i) Anthropic, PBC, in the United States; and

(ii) Google LLC (Google Cloud), in the United States.

StitchOps may offer other providers on request, after giving notice under its subprocessor process.

(c) Monitoring. When Customer uses AI Features, StitchOps may use LangChain, Inc. (LangSmith), in the United States, to monitor and troubleshoot them. Its records may include Customer Data sent to a model.

The full list, with purposes, locations, and data, is at https://www.stitchops.io/legal/subprocessors.

6. Customer Data and training

(a) No training. StitchOps does not use Customer Data to train AI models, and requires each provider it engages for StitchOps Provided Inference not to use Customer Data to train its models.

(b) What a model receives. A model receives the inputs of the AI Feature that calls it, such as a prompt, text from the workflow, or an image of the screen. Personal Data that a workflow handles otherwise stays with the Digital Worker on Customer's own infrastructure.

(c) Credentials. Credential values are never sent to an AI model. Digital Workers resolve them at run time from Customer's own credential vault, or from StitchOps storage only if Customer asks StitchOps to store them.

7. AI assistant connectors

When Customer connects its own AI assistant, that assistant's provider processes data under its own terms with Customer. The connection uses OAuth, so no keys are pasted into prompts. Read and write tools are granted separately, and destructive actions stop for confirmation. An assistant cannot see a credential value, run shell commands, skip a confirmation on a destructive step, or reach another tenant.

8. Human oversight

AI output can be inaccurate, whichever model produces it. Customer reviews AI generated workflows and output, and confirms actions an AI assistant proposes, before they run in production or are relied on. The Acceptable Use Policy at https://www.stitchops.io/legal/aup describes uses of AI Features that are not allowed, including automated decisions with significant effects on people without human review.

9. Changes

StitchOps gives at least 30 days notice before adding a provider that processes Customer Data. Subscribe to notices at https://www.stitchops.io/legal. Changes to this page follow Section 15.10 of the Master Subscription Agreement.