Trust
What we do with your data, who else touches it, what our AI is and isn’t allowed to do, and how to get it deleted. It covers Bobos Dispatcher (dispatcher.bobos.ai), the console agencies run their clients’ AI from, and the Bobos agency platform (v2.bobos.ai), where our managed-service clients see their work; where the two differ, it says which. If your security team has a questionnaire, most of it is answered below.
§1
The short version
- We never use your data to train AI models. You choose the AI model provider and connect it with your own account, so its own terms apply to what it receives.
- We use 8 subprocessors. They’re named below, with what each one can see.
- Connected-platform access tokens and AI-provider keys are encrypted at rest with AES-256-GCM, and never written to our logs.
- Disconnecting a platform deletes its stored token at once. Deletion requests are completed within 30 days.
- AI changes things in your systems only where you’ve allowed it. In Dispatcher an assistant can only read until a workspace admin lets its role change data, and the admin can make it ask a person first. On the agency platform, changes to your properties are proposed, and a person confirms each one.
- One client’s data is kept apart from another’s in code. A query against a client’s tables must carry that client’s filter, enforced by our build and our tests, and in Dispatcher by a runtime guard as well.
§2
Subprocessors
Last reviewed:
| Subprocessor | Purpose | Data it may process | Region |
|---|---|---|---|
| DigitalOcean (via Cloudways) | Hosting for both products: application servers, databases, Redis, Dispatcher’s AI gateway | Everything the products store: accounts, workspaces, chat history, reports, encrypted credentials | US |
| Cloudflare | DNS for our domains; the bot check on some forms | DNS queries; IP address and browser signals when a bot check runs | Global |
| Stripe | Subscriptions, invoices and card payments | Billing name, email and address; payment details (we never see card numbers) | US |
| Resend | Transactional email (sign-in links, invitations, notifications, reports) | Recipient address and the content of the emails we send | US |
| Sign-in with Google | Name and email at sign-in | US | |
| Sentry | Error monitoring | Error reports with request context; on the agency platform’s web pages, session replays with all text masked (5% of sessions and every session with an error) | US |
| DevJock | Runs the agency platform’s AI agents, and Dispatcher agents a workspace connects | Prompts and results of the agents it runs, which can include client reports, email and meeting-transcript text | US |
| ApiFlash | Screenshots of public web pages for site reviews | The address of a public page; no account data | US (AWS us-east-2) |
Each subprocessor is bound by confidentiality and security obligations and receives the data its job needs. Customer, connected-platform and Gmail data are never sold or transferred for anyone’s independent use.
Not on this list, because you choose them and they act under your own account: the AI model provider you connect with your own key, and the systems you connect — your ad accounts, Google Workspace, Microsoft 365, QuickBooks, CRMs and other APIs.
When this list changes materially, it changes here and, where it affects the public disclosure, in our privacy policy.
§3
What our AI is allowed to do
In Dispatcher, each workspace decides. A role’s assistant can only read until a workspace admin lets it change data in connected systems, and the admin chooses whether it acts directly or asks a person first. Scheduled workflows always stop and ask before a change. An assistant reaches only the tools its role was given, and a person reaches only what their role allows, in chat and over MCP alike.
In Dispatcher, every call has a ceiling. Each model call has a per-call dollar ceiling and a 120-second timeout, a tool-using turn has a per-turn ceiling, and a workspace can set a monthly cap. A role may name backup models, which answer only when its first model fails.
On the agency platform, our agents propose and a person confirms. 14 tools are available to our agents, and every tool that would change something on a property of yours returns a proposal for a person to confirm. Tools that write to our database record drafts, documents and notes for review. One process can send without a click per message: the cold-outreach autopilot, which emails prospects with AI-drafted messages and is switched off.
On the agency platform, in-house agents have hard limits. Each agent that runs in our own code carries a per-call dollar limit and a timeout set in code; an overage stops everything downstream rather than switching to a cheaper model. The analysis fleets that run on DevJock don’t yet carry a per-call cap of ours.
Nothing switches on by accident. On the agency platform every agent fleet is off until its identifier is provisioned, every scheduled process is registered with a risk class, and the high-risk ones — those that reach customers or change outside systems — are off until someone enables them and need typed confirmation to run by hand. Deploying our code does not activate a capability. One process is deliberately always on: the job that redacts personal data after 30 days.
Findings must trace to sources. In the weekly campaign and SEO reviews, code checks that every prioritized finding traces to a specialist’s finding and drops any that doesn’t. Our other analysis is held to that by instruction, not yet by code.
§4
Separation between customers
Dispatcher: 70 tables carry an explicit tenant scope, and four layers keep workspaces apart: the annotation on each table, a generated registry, a build rule that fails any query missing its tenant filter, and a runtime guard that refuses a protected query without that filter before it reaches the database. 17 cross-tenant tests run against a real database. A query across workspaces needs a written reason, which is logged before it runs.
Agency platform: 146 tables are classified in our tenancy registry, and our build fails if a query against a client-owned table is missing its filter or reaches it through raw SQL. A test suite runs 20 cross-tenant probes, and every one must return zero rows before anything ships. A query across clients needs a written reason, and an audit record carrying it is written before the query executes; if that write fails, the query still runs and the reason goes to our logs instead.
The honest limit: on the agency platform this is enforced at build time and by tests, not by runtime row-level security; Dispatcher adds its runtime guard, which refuses a query rather than rewriting it. Neither uses database row-level security. We’d rather you read that here than discover it in a questionnaire.
§5
Retention and deletion
Audit records: personal fields are redacted automatically after 30 days by a nightly job driven by a rule registry. On the agency platform, a record whose action has no rule yet is redacted in full and flagged.
Erasure requests: run as a defined job with a retry queue, not as a manual database session. It removes the person’s chat messages and the agent activity attributed to them, and in Dispatcher their sign-in factors too.
Disconnection: disconnecting a platform deletes its stored token at once. Reports and reviews built from it are kept until the client relationship ends or you ask us to delete them; we complete deletion requests within 30 days.
Diagnostics: error reports go to Sentry. The code that reports errors keeps tokens and secrets out of them; there is no automatic scrubbing pass yet. On the agency platform’s web pages, Sentry records 5% of sessions, and every session with an error, as a replay with all text masked.
/data-deletion.html — the request mechanism/privacy.html — the full policy/privacy-choices
§6
Access and identity
Human accounts support time-based one-time passwords, hardware security keys, recovery codes and trusted-device registration. Sessions can be invalidated centrally. On the agency platform, the tokens from signing in with Google are kept by our sign-in code outside the encrypted vault that holds connected-platform tokens.
In Dispatcher, an assistant connected over MCP is bound to the one workspace the person approved, and its permissions are re-checked on every call. A partner’s connection reaches only that partner’s own workspaces.
On the agency platform, a client’s AI connection uses a key scoped to that client and cannot reach beyond it. Our internal agents use a server-held credential per access tier, so what an agent may do is enforced where the agent can’t change it, not stated in its instructions. Our own team’s keys can reach every client’s documents.
Every production secret has a written rotation procedure (15 in all), and commits and dependencies are scanned in our build pipeline. Rotation is done by hand when a procedure or an incident calls for it, not yet on a schedule.
§7
Regulated industries
We maintain a registry of regulated industry categories — healthcare, financial services, legal, insurance, supplements, real estate and others — each with a severity classification and the disclosures it requires. Today it is reference material for our team; nothing gates output on it automatically yet.
§8
What we don’t have yet
We are not SOC 2 or ISO 42001 certified. We are a small firm and both are on a roadmap rather than in a filing cabinet. What we have instead is in the open: the controls above, the enforcement in our build, the test suites, and 78 written operational runbooks. If a certificate is a hard procurement requirement, we’ll tell you at the first call rather than the fifth.
We have not had an external penetration test. It’s a planned gate, not a completed one.
Runtime tenant filtering on the agency platform is not implemented. Dispatcher has a runtime guard; the agency platform relies on build-time rules and tests (see “Separation between customers” above).
Automatic scrubbing of error reports and scheduled secret rotation are not in place yet. Both are done by discipline and by hand today.
Questions this page didn’t answer?
Reply to whoever you’re talking to at Bobos, or email info@bobos.ai. If your security team needs a questionnaire completed, send it over; we’d rather fill it in than have the deal wait.
