Trust centre
AI and your data.
The short version: AI runs in accounts in your name, your data is not used to train models, models see only what a task needs, and a person approves anything that matters.
Last reviewed 29 October 2026. Questions: hello@smallbaydigital.com
Six commitments in every build
Your account, your keys
Model calls run through business API accounts in your name, so the usage, the bill and the data agreement are yours. We do not route your data through an SBD account in production.
Not used to train models
We use business APIs whose published terms exclude training on customer data by default, and we never use your data to train or fine-tune a model ourselves.
Only the fields a task needs
A model receives the fields required for that step, not whole records or tables. Card numbers, government IDs and fields you mark as sensitive are removed before a call is made.
A person approves what matters
Anything that moves money, contacts a customer for the first time, changes access or deletes data waits for a named person to approve it.
Actions are logged
Each model call and agent action is logged with its input summary, output, cost and who or what approved it, so you can audit any decision later.
You can switch it off
Every AI step has a manual fallback and an off switch you control, so the process keeps running if you pause the model.
Where the model runs
We choose where a model runs from what the data is and where it is allowed to go, and write the choice into the statement of work.
- 01
Provider business API
Most builds. Strong models, business terms that exclude training by default, retention set by the provider’s terms.
Zero or reduced data retention is available from some providers for eligible accounts; we apply for it where your use case needs it.
- 02
Your cloud account
When data must stay inside your existing cloud agreement or region, models can be called through the AI services of the cloud you already use.
Region and processing location follow your cloud account’s settings.
- 03
Private, self-hosted model
When data may not leave your infrastructure at all. Open models run on servers you control; nothing is sent to a model provider.
Smaller models suit classification, extraction and drafting; we test quality on your data before committing.
Controls on agents
- Prompt injection
- Content from emails, documents and web pages is treated as data, never as instructions. Agents are not given authority to act on commands found inside the content they read.
- Scoped tools
- An agent can only call the tools and records its job needs, with the same least-privilege credentials as any integration.
- Evaluation before go-live
- Each AI step is tested on a sample of your real cases, with the accuracy and failure cases written down before it runs unattended.
- Confidence thresholds
- Below an agreed confidence, the step goes to a person instead of acting. The threshold is reviewed with you after launch.
- Cost limits
- Model spend per run and per month is logged and capped, with an alert before a cap is reached.
Plan the controls for your own agent with the free AI agent risk and controls planner.
Provider terms
Retention periods and eligibility for zero-retention arrangements are set by each provider and change over time, so we link to the terms rather than quote them: Anthropic commercial terms, OpenAI enterprise privacy and Gemini API terms. The providers SBD itself uses are listed on the sub-processors page.