# Ballet for enterprise

> Canonical: https://ballet.dev/enterprise

Describe the work in plain English. Ballet writes it as code your engineers review, and the workspace decides who is allowed to change it.

## Put AI in production, not in another pilot

- SOC 2 Type II
- ISO 27001:2022
- Role-based access
- Approval gates
- Encrypted secrets
- Git-backed versions
- Replayable runs
- Bring your own models
- Usage and cost visibility
- Human in the loop

## Why pilots stall on the way to production

| Fine in a pilot | What production asks for |
| --- | --- |
| One good run, on stage | The same work done on the run nobody is watching |
| A prompt somebody owns | Code your engineers can diff and approve |
| Whoever built it, runs it | Roles that decide who can author and who can only run |
| Screenshots of the happy path | Every run logged, and replayable step by step |
| An API key in a notebook | Credentials the workspace holds, encrypted |
| Whichever model was the default | A gateway you point at your own provider |

## The same work, on the run nobody watches

| | Ballet | AI agents (Claude Sonnet 5 and GPT-5.6) |
| --- | --- | --- |
| Tasks that ran the same way all eight times | 100% | 78% |
| Time to finish one task | 0.08s | ~3.2s |
| Model spend at 10,000 tasks a month | $0 | $67–$112 |

Benchmark: https://docs.ballet.dev/articles/deterministic-operations-measured-A0mAp58RxY

## The controls your review board asks about

### Who can do what

Three roles, plus a setting for how much the AI does before it stops and asks you.

- Owners hold billing, invites and API tokens
- Builders author playbooks, connectors and secrets
- Members run what they are granted and author nothing
- Ask mode makes the AI propose every action for approval
- Approval gates live in the workflow, so no route around them

### What changed, and who approved it

Playbooks are code, so they go through the review your engineers already run.

- Connect a collection to a Git repository
- Changes land as pull requests your team reviews
- Version history in the repo you already audit
- Drift is surfaced when Ballet and the remote disagree
- Any run replays step by step with the data it saw

### Where your data goes

The Composer writes the workflow. It then runs as code, so live data moves through what your engineers signed off rather than through a model call.

- Point the AI Gateway at your own provider account
- Or use managed Ballet Models, with usage tracked
- Secrets are encrypted and never shown again in plain text
- Members cannot reach the Secrets page at all
- Covered by Brainfish's SOC 2 Type II programme

## Roles and permissions

| Capability | Owner | Builder | Member |
| --- | --- | --- | --- |
| Run playbooks | Yes | Yes | Yes |
| Create and edit playbooks | Yes | Yes | No |
| Manage connectors and secrets | Yes | Yes | No |
| Manage agents, skills and tools | Yes | Yes | No |
| AI Gateway settings | Yes | Yes | No |
| Create API tokens | Yes | Own | No |
| Invite users | Yes | No | No |
| Billing and plans | Yes | No | No |

## Systems it already reaches

Salesforce, ServiceNow, Zendesk, HubSpot, Gong, Freshdesk, Slack, Stripe, plus any API, database or MCP server.

## What procurement asks

### Can we use our own model provider?

Yes. The AI Gateway routes every AI request from the Console, agents and playbook steps, and you can point it at your own OpenAI or Anthropic account instead of Ballet's managed models. Owners and Builders configure it per workspace.

### How do we stop someone shipping a workflow nobody reviewed?

Authoring is a role. Members can run the playbooks they are granted and cannot create or edit any. Connect the collection to Git and every change reaches production as a pull request your engineers approve.

### What happens to our credentials?

They are stored as Secrets in the workspace, encrypted, and never displayed in plain text again. Steps reference them by name, so keys are not embedded in code, in the repo, or in shared run output. Members cannot access the Secrets page.

### Does the same task always run the same way?

On the compiled path, yes — that is what a playbook is. On our published operations benchmark a compiled playbook did the work identically on 54 of 54 tasks across 8 runs each, where frontier models managed 42 of 54 on the same tasks.

### Who supports a rollout?

Design partners get a specialist who builds the first playbooks with your team. Security reviews start at the Trust Center, where the SOC 2 Type II report, policies and subprocessor list are published: https://trust.brainfi.sh.

[Get a demo](https://ballet.dev/signup)